# Amp (Amp)

Feedback Bench, coding agents, built 2026-10-01, window 2026-08-31 to 2026-09-27. Web page: https://feedbackbench.com/#/agent/amp

| Measure | Value |
|---|---|
| Rank | =11 of 17 (rank range 11–12) |
| Feedback Score | 35.3 (95% interval 34.7–35.8) |
| Popularity | 0.222 (share of voice 0.75%) |
| Customer love | 0.561 (95% interval 0.544–0.577) |
| Top quadrant | no |
| Authors | 737 |
| Posts counted | 2786 |
| Posts that judge the agent | 1292 |
| Criteria better / worse than peers | 5 / 0 of 63 |

## The brief

Written by Claude Opus 5.5 from 85 labelled posts and the numbers on this page. Interpretation, not measurement: every quote is verbatim and links to its post.

**A polished orb factory that users love, until the servers blink.**

TL;DR:

- Parallel orbs and cloud threads are the draw; users run work from phones and wake to results.
- Outages are the loudest complaint, and each one pushes loyal users toward other tools.
- The top request is Claude subscription support, which an Amp reply says Anthropic's terms forbid.

### What hurts

- **Outages block work and drain loyalty** ([Outages, server errors and capacity or rate errors](https://feedbackbench.com/criteria/rel.service_errors.md), [Client crashes, freezes and failed tool execution](https://feedbackbench.com/criteria/rel.client_failures.md)). Service errors dominate the reliability complaints, and because the work runs in cloud threads, a backend outage stops the user cold.
  Posts describe threads that fail to reconnect and users left waiting with nothing to do. The pain is sharper for fans. One self-described advocate says each outage makes them want to use Amp less.
  
  An Amp reply in the window concedes recent instability from growth and infrastructure limits. It points to multi-model routing as a hedge against a single provider going down.
  Evidence:
  - Complaint, Amp, @AmpCode, 2026-09-08: “i really enjoy @ampcode and have recommended it to plenty of people. every time i get blocked because of an outage makes me want to use it less and less.” [source](https://twitter.com/423086332/status/2097324055942541682)
  - Complaint, Amp, @AmpCode, 2026-09-09: “@ampcode, how long before we will be up again - threads seem to reconnect unsucessfully.” [source](https://twitter.com/38651218/status/2097807908071489654)
  - Complaint, Amp, @AmpCode, 2026-09-01: “@ampcode severe withdrawal symptoms! #hugops and hope this gets resolved quickly!” [source](https://twitter.com/1491081/status/2094821977538122223)
  - Complaint, Amp, @AmpCode, 2026-09-06: “do it! let me know any feedback you have or anything you see missing. the biggest complaint right now is some instability we've had over the last few weeks (<strict_link>) as we've been growing and pushing the limits of our infrastructure, but we're working really hard to get better there. (and with multi-model, you can use other models if one is down, which saves you single-model-provider downtime.)” [source](https://twitter.com/784008/status/2096678351411060901)

- **Claude subscriptions stay locked out** ([Using an existing subscription across tools](https://feedbackbench.com/criteria/billing.subscription_portability.md)). Users want to bring a Claude plan into Amp, but an Amp reply says Anthropic's terms forbid it, so Claude users pay twice or leave.
  Bringing an existing subscription is the most-requested change by a wide margin. Posts call missing Claude support a deal breaker for orbs. One user thought Claude Code ran underneath and watched Amp credits drain instead.
  
  Teams already invested in the Claude ecosystem say the extra cost is hard to justify without it. OpenAI and other subscriptions route fine, which makes the gap more visible.
  Evidence:
  - Complaint, Amp, @AmpCode, 2026-09-25: “one last thing would make the anthropic plan a winner for me and that's proper support for other harnesses (@ampcode in my case). not being able to orb properly with claude is a deal breaker. <strict_link>” [source](https://twitter.com/33135576/status/2103289194231853557)
  - Complaint, Amp, @AmpCode, 2026-09-26: “@sqs @purefunctor @ampcode i mistakenly thought we can now use claude subscriptions through amp where amp uses claude code underneath only for anthropic models. then my amp credits keep going down…. haha” [source](https://twitter.com/1960166959388971008/status/2103902887965491272)
  - Complaint, Amp, @AmpCode, 2026-09-25: “@raygenrupe @ampcode anthropic disallows is in their terms of service. we would love to support claude subs, we just can't” [source](https://twitter.com/1369860113423609866/status/2103482010102563266)
  - Praise, Amp, @AmpCode, 2026-09-03: “oh nice! i'll try out the non-commit setup. and the event-driven orbs looks like exactly what i was looking for. i'll try it out. would be great to replace my cursor automations - those are fairly limited because it only exposes a subset of webhooks on the various platforms. def rooting for you guys so i'll keep trying out amp and testing new features. ideally would love to introduce it to some of my engineers at work. the spaces feature looks neat but obv couldn't play around with it myself. we are deeply ingrained in the claude ecosystem right now, which is far from perfect. but yeah the challenge is justifying the additional cost - esp since we can't use anthropic subscriptions on the platform. anyways keep up the great work! it's orbin time!” [source](https://twitter.com/1404603670382084098/status/2095657817260847610)

- **Interface controls confuse and misfire** ([How the interface shows work, and what the user can configure](https://feedbackbench.com/criteria/ui.display_settings.md), [Reasoning effort setting and its defaults](https://feedbackbench.com/criteria/models.effort_control.md)). Display and settings draw more complaints than praise because small controls misbehave, from greyed-out model levels to keybinds leaking characters into the prompt.
  Users report they cannot change the model level mid-conversation, and the gauge only flashes a tooltip. TUI mouse selection highlights the wrong text inside terminal multiplexers, and copy mode grabs scrollbar characters.
  
  A settings keybind leaves its trigger letter in the prompt box. The outline feature reads as a good idea that is hard to use. Requests ask for presets, a session dashboard and sidebar grouping.
  Evidence:
  - Complaint, Amp, @AmpCode, 2026-09-14: “@sqs @ampcode another thing that's confusing to me is i can't change the model level (i.e. "medium") mid-conversation. it's just greyed out. and clicking on the little gauge just pops up a tooltip saying "medium" that quickly disappears which is confusing” [source](https://twitter.com/721234540/status/2099487663384334421)
  - Complaint, Amp, @AmpCode, 2026-09-23: “some of @ampcode tui's rough edges: - mouse selection inside of herdr is broken: it highlights text from other part of the screen - herdr's copy mode usually also doesn't work as wanted: part of the selected characters are the scroll bar - which forces you to remove them later” [source](https://twitter.com/2062509956574650368/status/2102830957283954986)
  - Complaint, Amp, @AmpCode, 2026-09-23: “@ampcode other small details: - when opening amp settings with a configured keybind (like `&lt;leader&gt; s`), and then closing the editor, then an `s` is filled in the prompt input (the character of the keybind) - the "outline" feature i think is a good idea but quite confusing to use” [source](https://twitter.com/2062509956574650368/status/2102851244205465948)
  - Complaint, Amp, @AmpCode, 2026-09-25: “@toolmantim @thorstenball @ampcode cmd+option+b appears to be opening a weird arc global "new window" i've never used!” [source](https://twitter.com/5444392/status/2103329240972415252)

- **Plugins and MCP auth fail silently** ([MCP servers, plugins, skills and hooks](https://feedbackbench.com/criteria/setup.extensions_mcp.md), [Onboarding, discoverability and documentation](https://feedbackbench.com/criteria/setup.onboarding_docs.md)). Extensions break at the handoff, where an MCP sign-in succeeds on the provider side but Amp never detects it, and personal plugin modes vanish from the web app.
  Users describe reauthenticating, removing and re-adding an MCP connection with no change. A personal plugin mode shows up in the CLI but not the web or app. Event-driven orbs work only through plugins for now, which surprises people who look for a UI toggle.
  
  Some users fall back to a local runner to get browser auth. Getting a new model working through a plugin also stumps both users and the agent.
  Evidence:
  - Complaint, Amp, @AmpCode, 2026-09-05: “@sqs @ampcode tried it all, deactivated, reauthenticated, removed, added again. basically it says authenticated in netlify website after i click in sign in, but amp does not detect it.” [source](https://twitter.com/2931128860/status/2096362214970048959)
  - Complaint, Amp, @AmpCode, 2026-09-04: “@ampcode i was building a personal plugin, with an specific mode but i cannot see it in the web/app but in the cli it does show. are personal/user plugin modes excluded while workspace modes are included? puck told me this after auditing againts official plugin api "conclusion: your plugin meets every applicable documented requirement. the strongest diagnosis is an amp platform bug in server-side hosted-mode discovery: personal/user plugin modes are excluded while workspace modes are included."” [source](https://twitter.com/13267532/status/2095907323759718403)
  - Complaint, Amp, @AmpCode, 2026-09-24: “@ampcode @ampcode @thorstenball please support browser auth for mcp. 🙏 using a local runner as a workaround for now. <strict_link>” [source](https://twitter.com/295919262/status/2103084754153197846)
  - Complaint, Amp, @AmpCode, 2026-08-31: “oh my bad, thanks! i assumed there was small only for some reason i like the ship button 👀 1 small thing (probably a skill issue) is that i was having a decent amount of trouble figuring out how to get grok 4.6 working through my x premium+ subscription i had an orb helping me set it up by reading the docs and it ran added the grok 4.6 mode into my amp personal plugins but it doesn't work - the agent's struggling to figure out how to get it set up as well” [source](https://twitter.com/1762175022527877120/status/2094318577352597835)

- **Orb setup and cold starts stall** ([Cloud and remote sandbox execution](https://feedbackbench.com/criteria/surfaces.cloud_sessions.md), [Mobile, remote-control and voice access](https://feedbackbench.com/criteria/surfaces.remote_mobile.md), [How much use a plan's price buys](https://feedbackbench.com/criteria/limits.plan_value.md)). Cloud orbs that fail to initialise or shut down between check-ins force users to wait or debug environments before any real work starts.
  One user reports an orb setup retrying about ten times as interrupted, with no way to file a report. Another got stuck on Node version mismatches and environment variables.
  
  Mobile users say returning to a portal takes a minute or two to spin back up and ask for orbs to stay warm longer. Others want a one-click move to a bigger orb when small ones choke.
  Evidence:
  - Complaint, Amp, @AmpCode, 2026-09-11: “@ampcode if you're on and off orbin on your phone, it gets annoying pretty fast. so you send a task and then decides to check the portal it takes like 1-2 minutes to have stuff up running back to where you were. please let me keep then running for like 1 hour or detect pattern” [source](https://twitter.com/1857935142670450688/status/2098368506207088895)
  - Complaint, Amp, @AmpCode, 2026-09-21: “@ampcode my orb setup keeps failing and retrying. it’s retried about 10 times, each time saying it was interrupted even though i didn’t do anything. i wanted to give you a report i did but i cannot submit a report because the order didn't start @puckofamp could you take care of that? :>” [source](https://twitter.com/1993812110162448386/status/2102026996557472199)
  - Complaint, Amp, @AmpCode, 2026-09-04: “at this point, the problem is not amp code at all - i'm at a stage where i'm trying to make it a factory with github issues or trello as the conveyor belt. got stuck at orb initialisation, node versioning mismatch and environment vars. i stupidly tried to configure 3 projects at once so fixing it takes 3 times longer! i just need a little bit of daylight to really sink my teeth into it.” [source](https://twitter.com/1898512810272763904/status/2095692363050185110)
  - Complaint, Amp, @AmpCode, 2026-09-25: “@codermatt @sqs @ampcode a one click “put these changes on a bigger orb” would do the trick. i’ve bumped my default up to medium because every project i have kills the small!” [source](https://twitter.com/9111552/status/2103370105505943560)

### What works

- **Orbs that spawn and coordinate orbs** ([Subagents, parallel agents and orchestrators](https://feedbackbench.com/criteria/work.multi_agent_orchestration.md), [Long unattended runs and goal/loop mode](https://feedbackbench.com/criteria/work.long_running_autonomy.md)). Multi-agent orchestration is a standout because orbs can spin up other orbs and message between threads with little setup, and users say it just works.
  Praise outnumbers complaints heavily here. Users describe orbs watching GitHub issues, the built-in agent treated like a teammate for simple tasks, and cross-runner thread spawning that feels intuitive.
  
  The gaps are about control, not function. Users ask which settings child orbs inherit and want automations owned by the workspace rather than one person's tokens.
  Evidence:
  - Praise, Amp, @AmpCode, 2026-09-08: “@thorstenball i also enjoy orbs spinning up other orbs and having them communicate to each other - it's rather seamless and "just works" kudos to the whole @ampcode team, i'm rather shocked at how much just works and with the ux polish, it's all wonderful and useful, form meeting function” [source](https://twitter.com/2000470921329709056/status/2097210482234884569)
  - Praise, Amp, @AmpCode, 2026-09-23: “@parcadei @ampcode my thoughts exactly 👍 built a plugin to use it with poteto mode from @poteto . talking with other threads or spawning orbs and threads in other runners is just do intuitive. but sometimes still got fomo from tools like @capydotai , not knowing how their a2a workflow is.” [source](https://twitter.com/2416367251/status/2102750911441539320)
  - Praise, Amp, @AmpCode, 2026-09-26: “i love that we now have @harrystuck77 on the bandwagon. if you haven't played with @ampcode yet, then @ampcode + @ollama @ $20usd is perfect for you. how i use ampcode for my small projects: - cloud hosted codebase on gh. - orbs monitor gh issues for when my team creates them - puck treated on the same level as my infra guys for simple tasks. - projects hosted on vercel. - ci/cd to publish main.” [source](https://twitter.com/1898512810272763904/status/2103973089122116055)
  - Complaint, Amp, @AmpCode, 2026-09-05: “quick thought, @ampcode: it’d be great to have a “default” settings profile for new orbs spawned by orchestrator/admin orbs — unless this already exists and i just haven’t found it yet. right now, it’s a bit unclear which settings are inherited when an orb spawns other orbs to parallelize its work.” [source](https://twitter.com/2931128860/status/2096223195233718466)

- **Ships ambitious builds end to end** ([Can do the user's kind of task](https://feedbackbench.com/criteria/work.capability.md), [Diagnosing and fixing reported bugs](https://feedbackbench.com/criteria/work.bug_diagnosis.md)). Users credit Amp with finishing large, multi-part jobs, from a stalled CI/CD rewrite to an app it built and then tested itself.
  Posts describe a browser piano built and verified in an hour, and a bug-report pipeline where each report spawns an orb that drafts a fix as a pull request. One user says a months-delayed rewrite finally moved.
  
  Dissent targets specific models and task types. Users report mixed results on blog writing, no web search for research, and one model reverted for wasting tokens.
  Evidence:
  - Praise, Amp, @AmpCode, 2026-09-10: “@thorstenball i've been procrastinating a big ambitious rewrite of our internal ci/cd for months. fable 5.1/astra on @ampcode got me back to it this week. the work done so far is *tremendous*, almost glad to have waited so long. now i'm thinking "what's next". we must dream wayyy bigger” [source](https://twitter.com/7501222/status/2098018378375942313)
  - Praise, Amp, @AmpCode, 2026-09-05: “i asked @ampcode in high mode(gpt astra) to build a playable browser piano with felt controls, open samples and recording, then prove it worked by playing beethoven's für elise end to end. amp built the app, tested it and captured the performance you hear here. all in an hour. <strict_link>” [source](https://twitter.com/1822979861113417728/status/2096322588700230021)
  - Praise, Amp, @AmpCode, 2026-09-26: “i just did something that i wasn't aware it was possible using @ampcode . i had an idea on how to improve bug fixes. since everything now runs on an orb (their vm system), i thought: "it would be so cool if amp was my "support team""" so, i asked it to build. since we have a very good tracing system, we can check pretty much everything the user do, so we added an "send bug report" (the user allows us to see the last 10 minutes of his actions, there's a checkbox for that) and when he sends, it automatically creates an github issue, that by itself spawns an orb that starts checking what happened right away, if it needs an fix, it creates it and submit as a pr (and link it to the github issue) so we can validate. that made us a lot faster to debug and fix user issues. what a time to be alive.” [source](https://twitter.com/2931128860/status/2103847911725404390)
  - Complaint, Amp, @AmpCode, 2026-09-13: “@csinco @ampcode it’s very good. but no web search to do web research.” [source](https://twitter.com/1879194803700670464/status/2099081018301128914)

- **Work from the phone, wake to results** ([Cloud and remote sandbox execution](https://feedbackbench.com/criteria/surfaces.cloud_sessions.md), [Mobile, remote-control and voice access](https://feedbackbench.com/criteria/surfaces.remote_mobile.md)). Cloud execution plus a mobile app lets users start tasks away from the desk and check progress later, which they rate as smoother than rivals' remote stories.
  Users call Amp their most-used phone app and describe kicking off ideas while doing other things. Notifications from overnight cloud runs are a recurring point of pride. Voice now works in CarPlay, per one post.
  
  A competitor's TUI-first tool is described as lacking a remote or cloud story by comparison. The remaining complaint is that reviewing diffs on a phone still feels cramped.
  Evidence:
  - Praise, Amp, @AmpCode, 2026-09-02: “.@ampcode is now my most used app on my phone. it’s fun to be able to kick off ideas and check on ongoing work while doing other stuff. it’s helped me spend less time sitting at the computer while still getting as much if not more done. <strict_link>” [source](https://twitter.com/33135576/status/2095052108492099678)
  - Praise, Amp, @AmpCode, 2026-09-02: “this is what a successful night of cloud agents looks like. love waking up to @ampcode notifications <strict_link>” [source](https://twitter.com/1535774830225944576/status/2095242181426679863)
  - Praise, Amp, @AmpCode, 2026-09-14: “i'm already doing all my work from the cloud (orbin' on @ampcode). working from mobile more than ever now. the idea of the iphone duo is getting more appealing 🤔” [source](https://twitter.com/15281463/status/2099611203668455881)
  - Praise, Amp, @AmpCode, 2026-09-13: “@askpascalandy @ampcode pi is great but is catered more towards the tui and doesn’t have remote/cloud. codex app computer is still the best so i still use it for that, but i still find their remote/cloud story not as seamless as amp.” [source](https://twitter.com/394369752/status/2099181205346504850)

- **Support fixes bugs within the hour** ([Support, refunds and issue handling](https://feedbackbench.com/criteria/account.support.md)). Account and support is the strongest-rated area because the built-in bug reporter feeds a team that ships fixes fast and sometimes issues credits unprompted.
  One user filed a CLI crash through the reporter and saw a fix released about 30 minutes later. Others praise proactive credits after a broken thread as a trust rebuild.
  
  The complaints are about follow-through. Users want to see their past reports, and one says a web-submitted report got no reply.
  Evidence:
  - Praise, Amp, @AmpCode, 2026-09-02: “@codermatt @ampcode support amazing 💯👍 quick fixes like this create so much goodwill” [source](https://twitter.com/911541767333466114/status/2095011824236138852)
  - Praise, Amp, @AmpCode, 2026-09-21: “i know i glaze @ampcode often but the attention to detail and the speed with which issues are being addressed and fixed is unheard of in the pre-ai era. i had a problem with the cli crashing, it suggested i file a bug report using the built-in reporter. about 30 minutes later @thorstenball already fixed the issue and a new cli is being released. whether a human was involved here or the entire thing came through completely automated is really not of any interest - i, as an end-user, got a satisfactory conclusion to an issue i reported. whether you agree with thorsten's predictions on the future of software - it's undeniable that llm-driven development is an exponential multiplier and there's no going back <strict_link>” [source](https://twitter.com/1491081/status/2101982752799183067)
  - Praise, Amp, @AmpCode, 2026-09-24: “@iannuttall @ampcode proactive credits after a broken thread is a trust rebuild, not a coupon. most agent tools hide the failure. amp named it, priced the apology, and kept you in the loop. overlooked: reliability support is part of the harness. the model is not the product.” [source](https://twitter.com/2093342701227806720/status/2103160708943012157)
  - Complaint, Amp, @AmpCode, 2026-09-26: “would be cool if we could see our previous bug reports here @ampcode sometimes i can't remember if i filed something already, or i wanna add more detail, or even just have the bug id. <strict_link>” [source](https://twitter.com/33135576/status/2103636986116681898)

### Under the surface

- **Outage handling buys back goodwill** ([Outages, server errors and capacity or rate errors](https://feedbackbench.com/criteria/rel.service_errors.md), [Support, refunds and issue handling](https://feedbackbench.com/criteria/account.support.md)). The same users who hate downtime praise how Amp responds, because fast recovery and usage resets turn incidents into loyalty moments.
  Posts thank the team for quick restorations and for monthly usage resets after an upstream cloud failure. This cushions the reliability complaints for now.
  
  The cushion has limits. Goodwill from handling coexists with users saying repeated blocks make them use Amp less. If incidents keep recurring, the praise for response may stop offsetting the frustration.
  Evidence:
  - Praise, Amp, @AmpCode, 2026-09-01: “this is why i totally support @ampcode out and loud. they tried everything they could, the problem was around gcp infra, not theirs. they will still issue monthly usage resets to those affected. this is awesome and i really appreciate the work you all have been doing. <strict_link>” [source](https://twitter.com/2931128860/status/2094876670586974591)
  - Praise, Amp, @AmpCode, 2026-09-03: “really like how @ampcode handled today's outage. <strict_link>” [source](https://twitter.com/1350354798717984769/status/2095400240144843120)
  - Praise, Amp, @AmpCode, 2026-09-18: “@rockorager @ampcode so quick, i left my desk for less than five minutes and it was back up 👍” [source](https://twitter.com/33135576/status/2101085551021654105)
  - Praise, Amp, @AmpCode, 2026-09-18: “@rockorager @ampcode @sqs @thorstenball looks to be all good now, thanks for resolving quickly! muh orbs!” [source](https://twitter.com/1118175955/status/2101086063066583079)

- **BYOK widens reach and edge cases** ([Connecting own API keys, local models and custom endpoints](https://feedbackbench.com/criteria/setup.provider_byok_local.md), [Free tier and free model availability and limits](https://feedbackbench.com/criteria/billing.free_tier.md)). Bring-your-own-key brought back a free path and open-model support, but each new provider surfaces compatibility bugs that users hit first.
  Users recommend BYOK to students again and call Amp a strong harness for open models. One says the built-in agent configured a custom backend for them.
  
  The failures come from provider quirks. Auto-titling breaks on a reasoning-enabled custom endpoint, and custom models do not appear in dial tuning. An Amp reply asks for bug reports and patience across many provider edge cases.
  Evidence:
  - Praise, Amp, @AmpCode, 2026-09-12: “@ampcode has a free mode again through byok. if you already pay for @chatgpt, i strongly recommend giving amp a try with your own key. really glad i can properly recommend it to other students again without the cost being the first hurdle.” [source](https://twitter.com/1999233079202975744/status/2098891991551779314)
  - Complaint, Amp, @AmpCode, 2026-09-20: “hi, @ampcode : when a thread runs on a custom openai-compatible connection with reasoning/thinking enabled (e.g. kimi k3 via custom url, high effort), auto title generation fails with "tool_choice 'specified' is incompatible with thinking enabled". titling always uses the thread's model with a forced tool call — it should drop reasoning or degrade gracefully for aux requests.” [source](https://twitter.com/701611192704327681/status/2101610388714648009)
  - Praise, Amp, @AmpCode, 2026-09-14: “tested @ampcode byok over the weekend and in a rather funny twist amp is now the best harness and best for open source models. preparing a few bug reports but i added @synthetic_new and a meta api key for very cost effective additions to the roster. main issue i have is that my custom stuff doesn’t show up in the dial tuning options.” [source](https://twitter.com/1637683046395592705/status/2099530527170584907)
  - Complaint, Amp, @AmpCode, 2026-09-12: “@codermatt @ampcode yeah! now submit bug reports and please be patient as we encounter and fix like <zip_code> edge cases in all these providers and their models and their bugs. but we will get there and make it smooth.” [source](https://twitter.com/784008/status/2098874437961658556)

### Fine print

- Nearly all posts come from X conversations around the @AmpCode account; Reddit contributes only a handful.
- Several cited posts are replies from Amp staff, which frame vendor positions rather than user experience.
- Many criteria have too few posts to rank, so smaller pain points may be under- or over-stated.

## Top requests

What users ask to add or change, most asked first. 247 author-weeks ask for something. Requests do not change the Feedback Score. Rule: A separate pass by Claude Sonnet 5 reads every counted post and extracts what the author asks the agent or its vendor to add or change, with the criteria it maps to and a short normalised wording; it does not touch the labels or the Feedback Score. Claude Opus 5.5 groups the wordings within each criterion (the first criterion the request maps to) into themes; code counts them. A theme counts distinct author-weeks that ask for it, per agent; across agents, one author-week per agent. Themes asked in fewer than 2 author-weeks, and requests that share no theme, are not shown. Examples: up to 3 posts per theme from different authors, without slurs, preferring posts of 60 to 450 characters, most recent first.

| Rank | Request | Criterion | Author-weeks | Posts |
|---|---|---|---|---|
| 1 | Bring existing subscription into this agent | [Using an existing subscription across tools](https://feedbackbench.com/criteria/billing.subscription_portability.md) | 32 | 34 |
| 2 | macOS and Windows cloud environments | [Cloud and remote sandbox execution](https://feedbackbench.com/criteria/surfaces.cloud_sessions.md) | 4 | 5 |
| 3 | Access to subscription integration beta | [Using an existing subscription across tools](https://feedbackbench.com/criteria/billing.subscription_portability.md) | 4 | 4 |
| 4 | Add Opus 5.5 model | [Which models are offered on a plan and when](https://feedbackbench.com/criteria/models.catalog_access.md) | 4 | 4 |
| 5 | Custom base URL OpenAI-compatible endpoints | [Connecting own API keys, local models and custom endpoints](https://feedbackbench.com/criteria/setup.provider_byok_local.md) | 4 | 4 |
| 6 | Session status dashboard | [How the interface shows work, and what the user can configure](https://feedbackbench.com/criteria/ui.display_settings.md) | 4 | 4 |
| 7 | Allow subscription use in third-party harnesses | [Using an existing subscription across tools](https://feedbackbench.com/criteria/billing.subscription_portability.md) | 3 | 4 |
| 8 | Option to hide sidebar | [How the interface shows work, and what the user can configure](https://feedbackbench.com/criteria/ui.display_settings.md) | 3 | 4 |
| 9 | Remove model selector slot limit | [Which models are offered on a plan and when](https://feedbackbench.com/criteria/models.catalog_access.md) | 3 | 4 |
| 10 | Savable model and mode presets | [How the interface shows work, and what the user can configure](https://feedbackbench.com/criteria/ui.display_settings.md) | 3 | 4 |
| 11 | Sidebar project grouping, sorting and filtering | [How the interface shows work, and what the user can configure](https://feedbackbench.com/criteria/ui.display_settings.md) | 3 | 4 |
| 12 | Access to experimental and unreleased features | [Onboarding, discoverability and documentation](https://feedbackbench.com/criteria/setup.onboarding_docs.md) | 3 | 3 |

### 1. Bring existing subscription into this agent

- Amp, 2026-09-27, @AmpCode (X): “@anthropicai set opus 5.5 free, i wanna use my sub in @ampcode .” [source](https://twitter.com/1541582568029831169/status/2104046097211720023)
- Amp, 2026-09-26, @AmpCode (X): “@bedesqui @t3dotcodes @ampcode i tried to use claude code this way in amp and is a horrible experience. i don’t want claude code i want the amp experience with anthropic models… but api prices are unsustained for normal people” [source](https://twitter.com/13267532/status/2103890053625786519)
- Amp, 2026-09-26, @AmpCode (X): “@ampcode if you guys make claude sub in the app seamless, i will 1000% shill the shit out of you at my work” [source](https://twitter.com/129354616/status/2103884835416678538)

### 2. macOS and Windows cloud environments

- Amp, 2026-09-25, @AmpCode (X): “@sqs @sixhobbits @ampcode @sqs they’re good. i really liked devin macos orb. amp macos orb wen?” [source](https://twitter.com/1553373273639256065/status/2103437963472539744)
- Amp, 2026-09-20, @AmpCode (X): “@sqs @ampcode but what i mean orbs that run in a macos/ios environment, not the macos app or runners. right now orbs are debian vms, smth like devin mac,do we have this” [source](https://twitter.com/1592160489965948933/status/2101648500354437132)
- Amp, 2026-09-15, @AmpCode (X): “@jeffwang hope @ampcode gets something like this. macos orbs? 😳” [source](https://twitter.com/33135576/status/2099936171468136737)

### 3. Access to subscription integration beta

- Amp, 2026-09-23, @AmpCode (X): “@sqs @purefunctor @ampcode @sqs can i also get this private beta setting? i'm in the same shoes as justin, claude sdk / subscription is the only thing holding me back from full amp/orb usage (i already use the claude cli workaround but it's a bit clunky)” [source](https://twitter.com/2000470921329709056/status/2102730076936942058)
- Amp, 2026-09-23, @AmpCode (X): “@sqs @purefunctor @ampcode would love to test this out! ill give as much feedback as possible 😊” [source](https://twitter.com/129354616/status/2102727209295503396)
- Amp, 2026-09-23, @AmpCode (X): “@sqs @purefunctor @ampcode ooh, how might one get access to try this out?!” [source](https://twitter.com/72656715/status/2102727195655577829)

### 4. Add Opus 5.5 model

- Amp, 2026-09-22, @AmpCode (X): “@ampcode it's been sometime - looking forward to your post about opus 5.5 and update to the default dial :p” [source](https://twitter.com/85549810/status/2102480563596591364)
- Amp, 2026-09-22, @AmpCode (X): “i'm seeing visions of 6 sol agent + opus 5.5 oracle and i need it yesterday in @ampcode” [source](https://twitter.com/1281249457703538688/status/2102474111570366753)
- Amp, 2026-09-22, @AmpCode (X): “@ajkemps any idea when claude-opus-5.5 and gpt-6-sol will be available in @ampcode” [source](https://twitter.com/1940807256377118724/status/2102463085697192138)

### 5. Custom base URL OpenAI-compatible endpoints

- Amp, 2026-09-07, @AmpCode (X): “when will the megawatt subcription account be enabled for custom base url byok?, i’ve been waiting for it so long @sqs @ampcode 🥲” [source](https://twitter.com/1800425844491620352/status/2096782766776221869)
- Amp, 2026-09-01, @AmpCode (X): “@sqs @socksmyrocks @ampcode would it add support for any baseurl or only whitelisted?” [source](https://twitter.com/817707228912357377/status/2094878138320695803)
- Amp, 2026-09-01, @AmpCode (X): “i know @ampcode has byok (which i love, btw) but is it possible to add openai compatible endpoints? for example openrouter? that would be a game changer!” [source](https://twitter.com/1567548474794835969/status/2094866120213983725)

### 6. Session status dashboard

- Amp, 2026-09-24, @AmpCode (X): “would be great to have a status indicator in the thread picker for @ampcode 's tui... the web ui has a simple animation. any indicator would be enough tbh..” [source](https://twitter.com/2062509956574650368/status/2103149291468419198)
- Amp, 2026-09-24, @AmpCode (X): “@sqs @ampcode the built-in thread overview in claude code projects automatically grouping threads into idle, working, waiting for you and resolved collapsible groups is really nice. i've tried doing something similar in amp with portals but i'm missing a live thread view like that in amp” [source](https://twitter.com/132882990/status/2103060185367351574)
- Amp, 2026-09-10, @AmpCode (X): “@ampcode is amazing but its so hard to see status on threads... why not simplify this? <strict_link>” [source](https://twitter.com/17822452/status/2098112579180621888)

### 7. Allow subscription use in third-party harnesses

- Amp, 2026-09-25, @AmpCode (X): “@jkudish @theo @ampcode the continued anthropic gatekeeping is annoying. i need another $200/m openai sub but can't get one... and i won't get a $200/m anthropic sub because i can't use it in all the places i want to...” [source](https://twitter.com/16740797/status/2103558201971274216)
- Amp, 2026-09-25, @AmpCode (X): “@theo i miss being able to use my harness of choice @ampcode with the sub” [source](https://twitter.com/33135576/status/2103406298943615393)
- Amp, 2026-09-24, @AmpCode (X): “@addyosmani @buildwithrajath let us use the sub on other harnesses. @ampcode is the future, and i'll use whatever subscription it supports.” [source](https://twitter.com/1796637816589529088/status/2103227532090933711)

### 8. Option to hide sidebar

- Amp, 2026-09-24, @AmpCode (X): “@evanphx @thorstenball @ampcode i think there should be a ui element, there is in the web app. probably a bug.” [source](https://twitter.com/1164939261675622400/status/2103217148227358805)
- Amp, 2026-09-24, @AmpCode (X): “@thorstenball @ampcode there is no icon for me. how do i hide it?” [source](https://twitter.com/5444392/status/2103214625210908882)
- Amp, 2026-09-03, @AmpCode (X): “@camden_cheek @ampcode @sqs appreciate the great work! is it possible to make a toggle for the sidebar icons, so users can choose?” [source](https://twitter.com/1471360971876622343/status/2095332046675886139)

### 9. Remove model selector slot limit

- Amp, 2026-09-23, @AmpCode (X): “@jkudish @ampcode agree! could be an option in "build your own dial" to just add a couple more. sometimes i want "high but with fable" vs my regular high with sol.” [source](https://twitter.com/1567548474794835969/status/2102856772021305745)
- Amp, 2026-09-23, @AmpCode (X): “what i did to use more models and agents is to set them up in orbs with their tokens and then i use <strict_link> to have amp write workflow code. that way i can use a wide range of agents and models in a single orb. that said, one extra dial pick - or customizable as you say, wouldn't be a bad thing imo.” [source](https://twitter.com/1637683046395592705/status/2102845920454787411)
- Amp, 2026-09-23, @AmpCode (X): “@skastr052 @ampcode maybe with build your own dial, you could choose how many spots you have up to a reasonable maximum?” [source](https://twitter.com/33135576/status/2102841110309797938)

### 10. Savable model and mode presets

- Amp, 2026-09-17, @AmpCode (X): “@jellydn @ampcode @nicolaygerold any comment on this - allowing users to create their own presets for fast swtiching?” [source](https://twitter.com/98821843/status/2100566520497569819)
- Amp, 2026-09-17, @AmpCode (X): “@arvislacis @ampcode it’s their preset and i want to create my presets.” [source](https://twitter.com/89360130/status/2100527308033700325)
- Amp, 2026-09-01, @AmpCode (X): “there should be a @ampcode team for every vertical” [source](https://twitter.com/1670026308/status/2094780535163990108)

### 11. Sidebar project grouping, sorting and filtering

- Amp, 2026-09-07, @AmpCode (X): “@sqs @josevalerio @ampcode how about giving a project level toggle if required? basically a dropdown where people can select all projects, or a particular project - to filter in the sidebar. it could also filter for archive, snoozed threads or there could be no filter at all (current view)” [source](https://twitter.com/2067222195064143872/status/2096796163588681966)
- Amp, 2026-09-06, @AmpCode (X): “i beg you @ampcode i don’t want to make projects, or tag shit manually give me the clusters of work again <strict_link>” [source](https://twitter.com/1420340648117510149/status/2096739242193887449)
- Amp, 2026-09-17, @AmpCode (X): “@ampcode some random papercuts: - i'm not a big right sidebar user and prefer to run portals in my browser, but the only way to access the portal url is via the portal in the sidebar or pill (not always visible in the conversation). could we have a button that's always visible to open the portal? even in the thread's context menu would be good enough - can we have a setting to disable the "inactive last 72h" grouping? i'd rather keep my unarchive” [source](https://twitter.com/910328509/status/2100580225629196368)

### 12. Access to experimental and unreleased features

- Amp, 2026-09-23, @AmpCode (X): “@sqs @purefunctor @ampcode would like to try whatever (a) and (b) as recently hermes, raycast and glaze all have been experimenting with it” [source](https://twitter.com/906700218132918273/status/2102741226504258040)
- Amp, 2026-09-23, @AmpCode (X): “@sqs @purefunctor @ampcode if there's room to test it lmk! i am facing the same thing over here” [source](https://twitter.com/2931128860/status/2102737182662266906)
- Amp, 2026-09-23, @AmpCode (X): “@sqs @purefunctor @ampcode can i sneak in this test too?” [source](https://twitter.com/2084694287338393600/status/2102734911148892634)

## Facts

| Fact | Value |
|---|---|
| Version | n/a (rolling releases) |
| Released | Free/BYOK model shift: 2026-09-13 |
| Price | Free with bring-your-own compute/model keys (as of 2026-09-13); earlier in 2026 was pay-as-you-go with a $5 minimum and zero markup on provider pricing for individuals; Smart Mode is the paid tier for zero data sharing |
| Model | Multi-provider, bring-your-own-key |
| Surface | IDE (VS Code), CLI |

## Sources

| Channel | Source | Posts |
|---|---|---|
| X | @AmpCode | 2779 |
| Reddit | Posts that name it | 7 |

## Better than peers on

Can do the user's kind of task, Subagents, parallel agents and orchestrators, Mobile, remote-control and voice access, Cloud and remote sandbox execution, Support, refunds and issue handling

## Worse than peers on

None.

## All 63 criteria

Criterion love: 0.5 is the category norm. n: rated author-weeks.

### Paying and limits: Better than peers (customer love 0.569, n 135)

| Criterion | Reading | Customer love | 95% interval | n | Praise | Complaint |
|---|---|---|---|---|---|---|
| [Using an existing subscription across tools](https://feedbackbench.com/criteria/billing.subscription_portability.md) | Typical | 0.529 | 0.498–0.560 | 60 | 24 | 36 |
| [How much use a plan's price buys](https://feedbackbench.com/criteria/limits.plan_value.md) | Typical | 0.521 | 0.493–0.551 | 41 | 21 | 20 |
| [Single prompt, model or effort level consumes disproportionate quota](https://feedbackbench.com/criteria/limits.burn_rate.md) | Too few posts | 0.501 | 0.477–0.528 | 21 | 4 | 17 |
| [Free tier and free model availability and limits](https://feedbackbench.com/criteria/billing.free_tier.md) | Too few posts | 0.503 | 0.485–0.519 | 12 | 8 | 4 |
| [Usage meter visibility and accuracy](https://feedbackbench.com/criteria/limits.usage_meter.md) | Too few posts | 0.549 | 0.513–0.586 | 9 | 6 | 3 |
| [Quota reset timing and bonus or banked resets](https://feedbackbench.com/criteria/limits.reset_schedule.md) | Too few posts | 0.520 | 0.501–0.543 | 6 | 4 | 2 |
| [Pricing and plan terms stated clearly and consistently](https://feedbackbench.com/criteria/billing.pricing_clarity.md) | Too few posts | 0.493 | 0.487–0.499 | 5 | 0 | 5 |
| [Short rolling usage window blocks or interrupts work](https://feedbackbench.com/criteria/limits.window_interrupts_work.md) | Too few posts | 0.504 | 0.491–0.522 | 4 | 1 | 3 |
| [Price, allowance or plan terms changed](https://feedbackbench.com/criteria/limits.allowance_change.md) | Too few posts | 0.496 | 0.490–0.500 | 3 | 0 | 3 |
| [Prompt cache hits, misses and invalidation](https://feedbackbench.com/criteria/limits.prompt_cache.md) | Too few posts | 0.501 | 0.492–0.508 | 3 | 2 | 1 |
| [Pay-as-you-go overage, fallback billing and spend caps](https://feedbackbench.com/criteria/billing.overage_charges.md) | Too few posts | 0.497 | 0.493–0.500 | 2 | 0 | 2 |

Most recent posts:

- Praise, 2026-09-27, @AmpCode (X): “@ben_rapaport @ampcode yeah, been risking it since april via pi plugin and cliproxyapi without issues.” [source](https://twitter.com/89291422/status/2104021496079560919)
- Praise, 2026-09-27, @AmpCode (X): “@art049 @ampcode ! bring your own gpt sub. mobile and mac app are great” [source](https://twitter.com/971761/status/2104076476350173509)
- Praise, 2026-09-27, @AmpCode (X): “amp usage (basically llms). my orbs doesnt hit too much, often 10% every month. i am prob using it less then i should but i am struggling to find ways to use them more haha (it's already pretty good for my workflow) i wouldn't mind paying like $125 for a reset with less usage (without changing my recurring cycle). pretty much for emergencies. for example, happening right now: i have something i ne…” [source](https://twitter.com/2931128860/status/2104312520639176711)
- Complaint, 2026-09-27, @AmpCode (X): “@anthropicai set opus 5.5 free, i wanna use my sub in @ampcode .” [source](https://twitter.com/1541582568029831169/status/2104046097211720023)
- Complaint, 2026-09-27, @AmpCode (X): “@ampcode create a purchasable reset button &lt;3” [source](https://twitter.com/2931128860/status/2104200521477501280)
- Complaint, 2026-09-27, @AmpCode (X): “@ampcode @sqs @thorstenball the cheaper model thing is real for me. flash model on a refactor did 4 extra tool loops fixing its own edit and the bill went up. do the orbs evals show that or just latency?” [source](https://twitter.com/1835841692852682752/status/2104230660105781726)

### Setting up and connecting: Typical (customer love 0.528, n 106)

| Criterion | Reading | Customer love | 95% interval | n | Praise | Complaint |
|---|---|---|---|---|---|---|
| [Connecting own API keys, local models and custom endpoints](https://feedbackbench.com/criteria/setup.provider_byok_local.md) | Typical | 0.510 | 0.483–0.535 | 40 | 23 | 17 |
| [MCP servers, plugins, skills and hooks](https://feedbackbench.com/criteria/setup.extensions_mcp.md) | Typical | 0.495 | 0.471–0.520 | 33 | 15 | 18 |
| [Onboarding, discoverability and documentation](https://feedbackbench.com/criteria/setup.onboarding_docs.md) | Too few posts | 0.516 | 0.489–0.545 | 21 | 6 | 15 |
| [Install, launch and sign-in](https://feedbackbench.com/criteria/setup.install_signin.md) | Too few posts | 0.510 | 0.486–0.539 | 20 | 5 | 15 |
| [IDE and editor integration](https://feedbackbench.com/criteria/setup.ide_integration.md) | Too few posts | 0.500 | 0.490–0.512 | 5 | 2 | 3 |

Most recent posts:

- Praise, 2026-09-27, @AmpCode (X): “@benvargas @ampcode local always on was the pain. custom url via orbs is the kind of shortcut that sticks.” [source](https://twitter.com/1502677575776149504/status/2104182819664593107)
- Praise, 2026-09-26, @AmpCode (X): “@bedesqui @catalinmpit @t3dotcodes @ampcode you can wire claude code into amp ui directly btw (this thread is running in cc) <strict_link>” [source](https://twitter.com/3064259332/status/2103785310685647344)
- Praise, 2026-09-26, @AmpCode (X): “@sqs @ampcode @ollama fantastic decision. that was one of the final jigsaw pieces for me. 🫡” [source](https://twitter.com/1590702228234391552/status/2103868668400775310)
- Complaint, 2026-09-27, @AmpCode (X): “@iannuttall @ampcode fresh workers still drop mid-auth state” [source](https://twitter.com/27319585/status/2104269838072091095)
- Complaint, 2026-09-26, @AmpCode (X): “@ampcode, how is the slack bot supposed to work? i know one of the examples is "make a group dm with me and name", but that just doesn't work? it seems to only send dm's as my own user, which is very hard to interact with. i'm on a slack enterprise workspace if that matters.” [source](https://twitter.com/2095743155312467968/status/2103658167469293873)
- Complaint, 2026-09-26, @AmpCode (X): “@homborg @ampcode amp can't be backed by the local codex cli or claude, right? i wish someone would do that. essentially a web orchestration layer on top of my clis.” [source](https://twitter.com/36411940/status/2103851323238297984)

### Choosing models: Better than peers (customer love 0.539, n 45)

| Criterion | Reading | Customer love | 95% interval | n | Praise | Complaint |
|---|---|---|---|---|---|---|
| [Which models are offered on a plan and when](https://feedbackbench.com/criteria/models.catalog_access.md) | Too few posts | 0.515 | 0.491–0.540 | 23 | 10 | 13 |
| [Automatic model routing and fallback](https://feedbackbench.com/criteria/models.routing_auto.md) | Too few posts | 0.531 | 0.510–0.554 | 16 | 11 | 5 |
| [Reasoning effort setting and its defaults](https://feedbackbench.com/criteria/models.effort_control.md) | Too few posts | 0.493 | 0.479–0.507 | 10 | 3 | 7 |
| [Quality got worse or better over time](https://feedbackbench.com/criteria/models.quality_drift.md) | Too few posts | 0.512 | 0.497–0.529 | 6 | 4 | 2 |

Most recent posts:

- Praise, 2026-09-27, @AmpCode (X): “it's incredible how you can just run @ampcode in a medium gpt mode and then, if you need a little more "juju", can pull in any other model to help out.. <strict_link>” [source](https://twitter.com/5408192/status/2104110093218316753)
- Praise, 2026-09-25, @AmpCode (X): “@sellsy @ampcode for me it gives me all of the models and labs (including glm and deepseek) in one ui. the amp team have very good taste, ship very fast and are building things that i need as a dev like shared skills, a very easy way to connect to local runners, etc. it took me a while to get it!” [source](https://twitter.com/9111552/status/2103455738878140534)
- Praise, 2026-09-24, @AmpCode (X): “@leuler2718 @ampcode @synthetic_new i haven't noticed any degraded performance with gpt models in amp, no. i think that was a characteristic of older codex models.” [source](https://twitter.com/1637683046395592705/status/2103058329194856588)
- Complaint, 2026-09-27, @AmpCode (X): “amp usage (basically llms). my orbs doesnt hit too much, often 10% every month. i am prob using it less then i should but i am struggling to find ways to use them more haha (it's already pretty good for my workflow) i wouldn't mind paying like $125 for a reset with less usage (without changing my recurring cycle). pretty much for emergencies. for example, happening right now: i have something i ne…” [source](https://twitter.com/2931128860/status/2104312520639176711)
- Complaint, 2026-09-26, @AmpCode (X): “@benvargas @ampcode solid stack. that routing setup is doing a lot of work though. one server-side flag change and half your agents reroute without asking.” [source](https://twitter.com/1656373234647048192/status/2103698212087353406)
- Complaint, 2026-09-26, @AmpCode (X): “@wendell_adriel i'm in the exact same boat. i haven't used sol at all in the last week after using opus 5.5 for a few tasks. i'm primarily using @ampcode, and i find myself dealing with the experimental external agent feature just for opus.” [source](https://twitter.com/10604/status/2103904355766407327)

### Instructing and context: Typical (customer love 0.525, n 32)

| Criterion | Reading | Customer love | 95% interval | n | Praise | Complaint |
|---|---|---|---|---|---|---|
| [Output degrades as the context window fills](https://feedbackbench.com/criteria/context.long_context_decay.md) | Too few posts | 0.499 | 0.485–0.517 | 7 | 1 | 6 |
| [Direct in-prompt instructions and caps are followed](https://feedbackbench.com/criteria/context.instruction_following.md) | Too few posts | 0.511 | 0.498–0.526 | 6 | 4 | 2 |
| [Context compaction keeps what matters, cheaply and quickly](https://feedbackbench.com/criteria/context.compaction.md) | Too few posts | 0.519 | 0.504–0.534 | 6 | 6 | 0 |
| [Memory and state carried across sessions](https://feedbackbench.com/criteria/context.session_memory.md) | Too few posts | 0.503 | 0.492–0.514 | 6 | 4 | 2 |
| [Persistent project rules files are read and obeyed](https://feedbackbench.com/criteria/context.instruction_files.md) | Too few posts | 0.505 | 0.500–0.513 | 4 | 3 | 1 |
| [Finding the right files in the codebase](https://feedbackbench.com/criteria/context.codebase_retrieval.md) | Too few posts | 0.500 | 0.493–0.506 | 2 | 1 | 1 |
| [Asks the user versus guessing](https://feedbackbench.com/criteria/context.clarifying_questions.md) | Too few posts | 0.498 | 0.493–0.500 | 1 | 0 | 1 |
| [Images, PDFs and file attachments as input](https://feedbackbench.com/criteria/context.attachments.md) | Too few posts | 0.498 | 0.493–0.500 | 1 | 0 | 1 |

Most recent posts:

- Praise, 2026-09-27, @AmpCode (X): “@ileppane @ampcode i use it bare. i just added some global agents like how i want it to respond and some reusable stuff across projects but overall it's bare :)” [source](https://twitter.com/1705384263867379712/status/2104065142564749714)
- Praise, 2026-09-24, @AmpCode (X): “@sqs @ampcode i understand. it is just as easy to spin up a bigger thread to continue the work.” [source](https://twitter.com/1535774830225944576/status/2102937024726769849)
- Praise, 2026-09-24, @AmpCode (X): “@kentcdodds @bot come to think of it, all the agentic tools i use the most, @ampcode @bot and some hermes, all of them abstract compaction away, and i have continuous sessions with all 3, and no dumb zones” [source](https://twitter.com/33135576/status/2103020461017977219)
- Complaint, 2026-09-27, @AmpCode (X): “@solllin @ampcode the harness gap is real. we built a tracer for exactly this: watching drift accumulate across sessions until the agent was effectively operating on a hallucinated codebase.” [source](https://twitter.com/2074234098864816128/status/2104114241443639622)
- Complaint, 2026-09-26, @AmpCode (X): “@thorstenball @ampcode is that live or are you working on it? if it’s live, the model doesn’t seem to know to use if!” [source](https://twitter.com/5444392/status/2103702725733294250)
- Complaint, 2026-09-25, @AmpCode (X): “@ampcode curious why amp doesn't have any question/answer tools the model can use. having instances where the modal outputs a big explanation then in the last sentence: may i do that? i sometimes miss that it's asking at all! a ui question tool would make that al ot more obvious.” [source](https://twitter.com/5444392/status/2103600382303944750)

### Doing the work: Better than peers (customer love 0.597, n 158)

| Criterion | Reading | Customer love | 95% interval | n | Praise | Complaint |
|---|---|---|---|---|---|---|
| [Subagents, parallel agents and orchestrators](https://feedbackbench.com/criteria/work.multi_agent_orchestration.md) | Better than peers | 0.558 | 0.533–0.583 | 57 | 49 | 8 |
| [Can do the user's kind of task](https://feedbackbench.com/criteria/work.capability.md) | Better than peers | 0.550 | 0.522–0.578 | 55 | 48 | 7 |
| [Long unattended runs and goal/loop mode](https://feedbackbench.com/criteria/work.long_running_autonomy.md) | Too few posts | 0.510 | 0.486–0.532 | 26 | 22 | 4 |
| [Git commits, branches and sync](https://feedbackbench.com/criteria/work.git_workflow.md) | Too few posts | 0.497 | 0.481–0.516 | 14 | 4 | 10 |
| [Computer use and browser control](https://feedbackbench.com/criteria/work.computer_browser_use.md) | Too few posts | 0.499 | 0.486–0.512 | 9 | 5 | 4 |
| [Diagnosing and fixing reported bugs](https://feedbackbench.com/criteria/work.bug_diagnosis.md) | Too few posts | 0.505 | 0.492–0.515 | 7 | 6 | 1 |
| [Risky or irreversible actions without confirmation](https://feedbackbench.com/criteria/work.destructive_actions.md) | Too few posts | 0.493 | 0.486–0.499 | 5 | 0 | 5 |
| [Stops mid-task or answers instead of acting](https://feedbackbench.com/criteria/work.premature_stop.md) | Too few posts | 0.492 | 0.485–0.498 | 4 | 0 | 4 |
| [Length and clarity of replies, summaries and comments](https://feedbackbench.com/criteria/work.response_verbosity.md) | Too few posts | 0.508 | 0.498–0.521 | 4 | 3 | 1 |
| [Frontend and visual UI output](https://feedbackbench.com/criteria/work.frontend_ui.md) | Too few posts | 0.493 | 0.481–0.500 | 2 | 0 | 2 |
| [Does unrequested work or over-engineers](https://feedbackbench.com/criteria/work.scope_overreach.md) | Too few posts | 0.497 | 0.493–0.500 | 2 | 0 | 2 |
| [Tool approval prompts and autonomy modes](https://feedbackbench.com/criteria/work.permission_prompts.md) | Too few posts | 0.504 | 0.495–0.517 | 2 | 1 | 1 |
| [Breaks existing code or reintroduces bugs](https://feedbackbench.com/criteria/work.regressions_introduced.md) | Too few posts | 0.498 | 0.495–0.500 | 1 | 0 | 1 |
| [Plan-before-edit mode](https://feedbackbench.com/criteria/work.plan_mode.md) | Too few posts | 0.498 | 0.494–0.500 | 1 | 0 | 1 |
| [Caves to or argues with the user's judgement](https://feedbackbench.com/criteria/work.sycophancy_pushback.md) | Too few posts | 0.498 | 0.494–0.500 | 1 | 0 | 1 |
| [Spins, loops or gets stuck without progress](https://feedbackbench.com/criteria/work.stuck_loops.md) | Too few posts | 0.500 | 0.500–0.500 | 0 | 0 | 0 |
| [Games checks instead of fixing the problem](https://feedbackbench.com/criteria/work.reward_hacking.md) | Too few posts | 0.500 | 0.500–0.500 | 0 | 0 | 0 |
| [Safety filters block legitimate coding tasks](https://feedbackbench.com/criteria/work.safety_refusals.md) | Too few posts | 0.500 | 0.500–0.500 | 0 | 0 | 0 |

Most recent posts:

- Praise, 2026-09-27, @AmpCode (X): “first i didn't know how to use git worktrees then i had @ampcode orbs and didn't need to know how to use git worktrees.” [source](https://twitter.com/89691524/status/2104148649160962391)
- Praise, 2026-09-26, @AmpCode (X): “@simonbs @ampcode its good, right? the other good part is running threads in orbs (ad-hoc vms) and i think it follows from there, that llm provider stuff is mainly handled at a level above each node you can ask the agent to hand off work to any local cli and ask it to use your shared tmux session” [source](https://twitter.com/132882990/status/2103833633802842583)
- Praise, 2026-09-26, @AmpCode (X): “i just did something that i wasn't aware it was possible using @ampcode . i had an idea on how to improve bug fixes. since everything now runs on an orb (their vm system), i thought: "it would be so cool if amp was my "support team""" so, i asked it to build. since we have a very good tracing system, we can check pretty much everything the user do, so we added an "send bug report" (the user allows…” [source](https://twitter.com/2931128860/status/2103847911725404390)
- Complaint, 2026-09-26, @AmpCode (X): “@thorstenball @ampcode i have a skill to address pr feedback and the model usually asks me in text “may i continue?” seems like that’s a perfect place to give the user a ui element” [source](https://twitter.com/5444392/status/2103705622650978634)
- Complaint, 2026-09-26, @AmpCode (X): “@thorstenball @ampcode probably! curious that i never see the modal use the choice tool is all” [source](https://twitter.com/5444392/status/2103706972151587097)
- Complaint, 2026-09-26, @AmpCode (X): “anyone using @ampcode to write blog posts? getting very mixed results so far so curious about your process.” [source](https://twitter.com/881577286927036416/status/2103824362746859692)

### Checking and finishing: Too few posts (customer love 0.492, n 11)

| Criterion | Reading | Customer love | 95% interval | n | Praise | Complaint |
|---|---|---|---|---|---|---|
| [Builds, tests or runs its own changes](https://feedbackbench.com/criteria/verify.self_testing.md) | Too few posts | 0.491 | 0.477–0.502 | 4 | 1 | 3 |
| [Reviewing and approving the agent's changes](https://feedbackbench.com/criteria/verify.change_review_ui.md) | Too few posts | 0.501 | 0.492–0.511 | 4 | 2 | 2 |
| [Claims work is done or fixed when it is not](https://feedbackbench.com/criteria/verify.false_completion.md) | Too few posts | 0.497 | 0.493–0.500 | 2 | 0 | 2 |
| [Agent-performed code review finds real issues](https://feedbackbench.com/criteria/verify.agent_code_review.md) | Too few posts | 0.503 | 0.500–0.509 | 2 | 2 | 0 |

Most recent posts:

- Praise, 2026-09-18, @AmpCode (X): “@adamwathan wrote an @ampcode plugin that checks if an llm’s code output follows the spec i give it. it checks against specific verification criteria and claims. it’s already caught some stuff for me” [source](https://twitter.com/33135576/status/2101006859314561178)
- Praise, 2026-09-11, @AmpCode (X): “`amp sync` is yet another awesome addition from @ampcode - lets you (temporarily) sync changes from a thread to your machine and deletes them when you exit. great for review in your preferred diff tool.” [source](https://twitter.com/1491081/status/2098282433674092676)
- Praise, 2026-09-03, @AmpCode (X): “@jtaby @sethmills21 we use @ampcode, which has solutions to both: * their cloud agents can rpc to a local mac to build and report back (we do this for our ios app) * they have a built-in diff viewer/commenter and multiplayer for other team members other cloud agents might have this solved too?” [source](https://twitter.com/1183203638/status/2095596826544259359)
- Complaint, 2026-09-25, @AmpCode (X): “@badlogicgames i think this is something which is missing from all ai tools like @opencode desktop and @ampcode i want to review/read the code with lsp and code navigation. all of them just shows git diff only” [source](https://twitter.com/1158785224299335680/status/2103409717959705080)
- Complaint, 2026-09-25, @AmpCode (X): “@sqs @ampcode the changes tab on very large repos gets weirdly out of sync showing like 80k+ changes or something. sometimes running git pull or other commands fix it, other times get worse. i think it needs some way to refresh on the ui. could also be comparing wrong commit” [source](https://twitter.com/1857935142670450688/status/2103610446720942394)
- Complaint, 2026-09-25, @AmpCode (X): “@beyang @sqs @ampcode you're not offering that, but i don't mind beta testing. :) this one bugs me a lot in a certain huge ass monorepo” [source](https://twitter.com/1857935142670450688/status/2103615119880224923)

### Interface and sessions: Better than peers (customer love 0.556, n 176)

| Criterion | Reading | Customer love | 95% interval | n | Praise | Complaint |
|---|---|---|---|---|---|---|
| [How the interface shows work, and what the user can configure](https://feedbackbench.com/criteria/ui.display_settings.md) | Typical | 0.512 | 0.481–0.543 | 91 | 37 | 54 |
| [Cloud and remote sandbox execution](https://feedbackbench.com/criteria/surfaces.cloud_sessions.md) | Better than peers | 0.533 | 0.506–0.560 | 51 | 42 | 9 |
| [Mobile, remote-control and voice access](https://feedbackbench.com/criteria/surfaces.remote_mobile.md) | Better than peers | 0.541 | 0.517–0.565 | 36 | 27 | 9 |
| [Saving, switching, resuming and rewinding sessions](https://feedbackbench.com/criteria/ui.session_history.md) | Too few posts | 0.500 | 0.481–0.518 | 16 | 5 | 11 |
| [Stopping and steering a running agent](https://feedbackbench.com/criteria/ui.interrupt_steer.md) | Too few posts | 0.495 | 0.488–0.500 | 2 | 0 | 2 |

Most recent posts:

- Praise, 2026-09-27, @AmpCode (X): “@art049 @ampcode ! bring your own gpt sub. mobile and mac app are great” [source](https://twitter.com/971761/status/2104076476350173509)
- Praise, 2026-09-27, @AmpCode (X): “@ariesthecoder @ampcode @ollama fantastic. thanks for the insight. will do more projects where i don't need the mission views.” [source](https://twitter.com/1590702228234391552/status/2104122342981128653)
- Praise, 2026-09-27, @AmpCode (X): “claude projcets is converting everyone to the @ampcode orb way <strict_link>” [source](https://twitter.com/132882990/status/2104147861302563096)
- Complaint, 2026-09-27, @AmpCode (X): “@sqs @purefunctor @ampcode this was a little too bad for me especially on mobile. i made a plugin that uses the agent sdk and streams html-readable conversation in a portal to monitor claude (screenshots etc) the amp thread acts as a sidekick which can steer the claude subagent. anthrophic tos sucks <strict_link>” [source](https://twitter.com/3852971/status/2104007337744953740)
- Complaint, 2026-09-27, @AmpCode (X): “@ampcode, i miss the press and hold action on "ship" - nothing important but just letting you know in case this was unintentional” [source](https://twitter.com/38651218/status/2104333581762146660)
- Complaint, 2026-09-26, @AmpCode (X): “.@sqs @ampcode can we get per user orb defaults instead of per org?” [source](https://twitter.com/475271251/status/2103692025644077443)

### Reliability and speed: Typical (customer love 0.490, n 74)

| Criterion | Reading | Customer love | 95% interval | n | Praise | Complaint |
|---|---|---|---|---|---|---|
| [Outages, server errors and capacity or rate errors](https://feedbackbench.com/criteria/rel.service_errors.md) | Typical | 0.539 | 0.480–0.597 | 37 | 5 | 32 |
| [Client crashes, freezes and failed tool execution](https://feedbackbench.com/criteria/rel.client_failures.md) | Too few posts | 0.493 | 0.462–0.544 | 24 | 1 | 23 |
| [Latency, throughput and fast mode](https://feedbackbench.com/criteria/rel.response_speed.md) | Too few posts | 0.497 | 0.481–0.512 | 13 | 5 | 8 |
| [Updates break working setups](https://feedbackbench.com/criteria/rel.update_breakage.md) | Too few posts | 0.510 | 0.492–0.533 | 6 | 2 | 4 |

Most recent posts:

- Praise, 2026-09-26, @AmpCode (X): “@ampcode team is shipping. i get like 5 relaunch to update per day. and the experience keeps getting better and better. i know orbs and runners are kinda competing products, but would like to have feature parity between them with browser use and everything else.” [source](https://twitter.com/1993188300719636485/status/2103833810978902137)
- Praise, 2026-09-23, @AmpCode (X): “@sqs @purefunctor @ampcode i’ve never opened amp as fast as i just did” [source](https://twitter.com/1443430538/status/2102751907387498733)
- Praise, 2026-09-23, @AmpCode (X): “@sqs @ampcode wow!! so quick! i'll give it a shot! you're amazing @sqs” [source](https://twitter.com/268615009/status/2102838747721277472)
- Complaint, 2026-09-26, @AmpCode (X): “@benvargas @fastchicken @ampcode i set that up too, when i however have too many concurrent claude requests going on it starts giving me errors, do you have issues with that?” [source](https://twitter.com/2416367251/status/2103773774084550768)
- Complaint, 2026-09-25, @AmpCode (X): “@yjsoon @ampcode not just you. we just updated our status page at <strict_link>. seems chatgpt is having some broader issues.” [source](https://twitter.com/1369860113423609866/status/2103621110655004860)
- Complaint, 2026-09-24, @AmpCode (X): “@sqs @ianlandsman @ampcode i will try tailscale as well, curious how that fits together. but it would be great if portals would be fast and usable. what causes the slowness here? it's one of the main things that's keeping me from moving everything to amp.” [source](https://twitter.com/1330980620617601026/status/2103028237299290550)

### Account and support: Better than peers (customer love 0.641, n 42)

| Criterion | Reading | Customer love | 95% interval | n | Praise | Complaint |
|---|---|---|---|---|---|---|
| [Support, refunds and issue handling](https://feedbackbench.com/criteria/account.support.md) | Better than peers | 0.634 | 0.600–0.664 | 33 | 28 | 5 |
| [Account bans and access restrictions](https://feedbackbench.com/criteria/account.bans_restrictions.md) | Too few posts | 0.492 | 0.486–0.497 | 6 | 0 | 6 |
| [Wrong charges, failed payments and plan provisioning](https://feedbackbench.com/criteria/account.billing_errors.md) | Too few posts | 0.495 | 0.490–0.499 | 4 | 0 | 4 |
| [Data retention, training use and deployment isolation](https://feedbackbench.com/criteria/account.data_privacy.md) | Too few posts | 0.502 | 0.495–0.512 | 2 | 1 | 1 |

Most recent posts:

- Praise, 2026-09-25, @AmpCode (X): “friends at @ampcode, is there a way to opt out of receiving free credits? you guys resolving the bugs means a lot more than credits. <strict_link>” [source](https://twitter.com/1089515309122367489/status/2103360810378752178)
- Praise, 2026-09-24, @AmpCode (X): “i got $30 in credits from @ampcode today because an image &gt; 30mb killed a claude thread and orbs weren't working for about 5 minutes! proactive - they just emailed to let me know. if i ever launch a successful product again, this is how i want to treat customers. &lt;3” [source](https://twitter.com/9111552/status/2103093438988321082)
- Praise, 2026-09-24, @AmpCode (X): “@iannuttall @ampcode the fact they hit you up first instead of waiting for you to notice is wild. most companies just go radio silent when stuff breaks” [source](https://twitter.com/1811356104083083264/status/2103095021352468765)
- Complaint, 2026-09-26, @AmpCode (X): “would be cool if we could see our previous bug reports here @ampcode sometimes i can't remember if i filed something already, or i wanna add more detail, or even just have the bug id. <strict_link>” [source](https://twitter.com/33135576/status/2103636986116681898)
- Complaint, 2026-09-26, @AmpCode (X): “@benvargas @ampcode are folks getting bans with this or is it pretty safe…? 😅🫣” [source](https://twitter.com/7841602/status/2103841399347331477)
- Complaint, 2026-09-25, @AmpCode (X): “@priyashpatil @ampcode @sqs wait, you guys are getting free credits? 😄 i’ve reported plenty of bugs to amp and somehow never got any. <email_address> — time to restore justice 😅” [source](https://twitter.com/1954974932267388928/status/2103459437931548833)