# Grok Bot users made it a dispatcher for Claude Code and Codex before xAI's routing

*In seven weeks of posts, 498 people praised running Grok Bot as a team of bots, each with its own job, and Grok Bot is 1st of 5 on multi-agent teams. Separately, 202 people made Grok Bot a dispatcher for agents from other makers, most often Claude Code and Codex, and 134 of them did so before Musk's 7 October post brought other makers' models into Grok Bot. One in five say they do it to save Grok Bot's allowance or use subscriptions they already pay for.*

What users credit, Grok Bot. By Feedback Bench, 17 October 2026. Feedback Bench 1.1. Posts 2026-08-31 to 2026-10-18; bench built 2026-10-19. Web: https://feedbackbench.com/#/articles/grok-bot-users-made-it-a-dispatcher-for-claude-code-and-codex-before-xais-routing

> yesterday while driving my @tesla on fsd, i voice-to-texted grok @bot an automation for cold outreach to find motivated seller leads for my real estate team, that i want to run on my @nousresearch hermes vps.

A user on X, in a post that tagged Nous Research, 5 September ([post](https://twitter.com/1504949293291085824/status/2096219497166197159))

Grok Bot drew the workflow out of Claude Code and gave the job to Hermes Agent. "[grokbot, connected to my claude code, extracted the workflow from there, then spun up a new profile on my hermes vps to handle it, and sent that new profile to build the automation and wire up the cold outreach workflow.](https://twitter.com/1504949293291085824/status/2096219497166197159)"

A month later, on 7 October, Elon Musk posted that Grok Bot would use other makers' models in its own back end ([post](https://x.com/elonmusk/status/2107724314451878104)):

> Important note regarding Grok @Bot: Going forward, @SpaceX will use the best back end model for any given task, including Claude Opus 5.5, MidJourney, Suno and other leading APIs. Whatever is most likely to give you the best outcome.

What users had built by then is related but different. They signed in coding agents they already pay other makers for, on Grok Bot's own computer, and had Grok Bot hand those agents the work. One in five say they did it to save Grok Bot's allowance or use subscriptions they already pay for.

## The strength

Grok Bot users run it as a staff, each bot with its own job. In the seven weeks we read, 498 people praised running it this way from their own use. On [Multiple agents, teams and delegation between them](https://feedbackbench.com/criteria/auto.multi_agent.md), Grok Bot is 1st of 5, and its interval is clear of Hermes Agent's in second place. Store reviews are about a tenth of its rated user-weeks here. Without them, its praise share is 84%, against 79% for Hermes Agent.

**Exhibit 1: Grok Bot is the only personal agent clearly above the norm on multi-agent teams.** Customer love on Multiple agents, teams and delegation between them, with 95% intervals, for the five agents with at least 30 rated user-weeks

| Agent | Customer love | 95% interval | Rated user-weeks |
|---|---|---|---|
| Grok Bot | 0.619 | 0.580 to 0.653 | 1170 |
| Hermes Agent | 0.486 | 0.448 to 0.528 | 574 |
| OpenClaw | 0.471 | 0.428 to 0.513 | 282 |
| Muse | 0.454 | 0.419 to 0.489 | 99 |
| ChatGPT Dots | 0.415 | 0.377 to 0.454 | 148 |

*Collection differs: Grok Bot is collected on every channel, including both app stores. Store reviews are about a tenth of Grok Bot's rated user-weeks here; without them its praise share is 84%, against 79% for Hermes Agent, which has no store reviews. ChatGPT Dots launched on 29 September, so its base is shorter. Hark, Instinct, Gemini Spark and Poke have too few rated user-weeks to place. n = 2,273 rated user-weeks. Source: Feedback Bench 1.1. Posts 31 Aug to 18 Oct 2026.*

## The dispatcher users built

xAI's launch post describes a chief of staff over specialist bots, all of them xAI's own, and the documentation never names another maker's agent.

We wrote the rule before counting. A person counts if a post from their own use says their Grok Bot directs, dispatches work to or controls an agent from another maker. Any sentiment counts. Wishes, how-to advice, news relays, peer channels where Grok Bot directs nothing, and setups where another agent directs Grok Bot do not. Cursor's cloud agents do not count either, because Cursor sells Grok Bot. Under that rule, 202 people count, and for 134 of them a post before Musk's announcement already says so.

**Exhibit 2: 134 of 202 people described Grok Bot directing other makers' agents before Musk's 7 October post.** People, by whether their first post naming Grok Bot with another maker's agent, or their first request to see, pick or pin the model, came before 7 October or from 7 October

|  | People |
|---|---|
| Directs other makers' agents: first post 31 Aug to 6 Oct | 134 people |
| Directs other makers' agents: first post 7 to 18 Oct | 68 people |
| Asked to see, pick or pin the model: first asked 31 Aug to 6 Oct | 30 people |
| Asked to see, pick or pin the model: first asked 7 to 18 Oct | 27 people |

*Both groups were read in full under rules written before counting. For each of the 134, a post before 7 October says Grok Bot directs the agent. Requests: 30 people first asked in the five weeks before the post and 27 from 7 October to the end of the window on 18 October. Opt-out-only asks are not in these rows. n = 202 and 57 people. Source: posts read by AI agents under written rules. Posts 31 Aug to 18 Oct 2026.*

The agents they direct are mostly coding agents. Claude and Claude Code come first, then Codex, then Hermes Agent. About two in three of these people name Claude or Codex, and about two in five name agents from more than one maker.

**Exhibit 3: Claude and Codex are the agents Grok Bot directs most often.** People who say their Grok Bot directs, dispatches work to or controls an agent from another maker, by the agent named; one person can appear in several rows

|  | People |
|---|---|
| Claude and Claude Code (Anthropic) | 106 people |
| Codex (OpenAI) | 67 people |
| Hermes Agent | 41 people |
| Other makers' agents, such as OpenCode, Droid and Copilot | 20 people |
| ChatGPT and other OpenAI models | 19 people |
| OpenClaw | 16 people |
| Gemini and Antigravity (Google) | 14 people |
| Muse | 8 people |
| T3 Code | 6 people |
| ChatGPT Dots | 4 people |
| Devin (Cognition) | 3 people |

*Coded by reading each person's posts, not by keyword. 133 of the 202 name Claude or Codex; 80 name agents from two or more makers. Cursor's cloud agents are left out, because Cursor sells Grok Bot. n = 202 people. Source: posts read by AI agents under written rules. Posts 31 Aug to 18 Oct 2026.*

One user had Grok Bot sign in to three coding agents and send each the same job. "[this time i had grok bot act as the main dispatcher, installing and logging in codex, grok build, and antigravity on its cloud computer.](https://twitter.com/1588747531248230400/status/2109212957410259011)" Another let it start a long job on a Mac. "[my grok bot opened a new terminal and instructed claude code, which then coordinated with its own skills and subagents.](https://www.youtube.com/watch?v=c9Z7ixVr4CU&lc=UgwVDokwEPwuQt94-Vd4AaABAg.AbbVZgtc6ZiAbbxt98EOsw)"

A separate cluster sits in Hermes Agent's own community. These users keep self-hosted Hermes agents and put Grok Bot in front for its phone app and its voice. One wrote in a Nous Research thread, "[my grok bot "bilby" plugs straight in and hands off jobs from voice mode.](https://twitter.com/2040767007390384128/status/2108482612721524888)" Another, in a post about OpenClaw, runs several Hermes agents and OpenClaw behind Grok Bot. "[grok bot has its place, i have grok bok as my high level orchestrator connecting to hermes ( multiple ) and openclaw via cloudflare tunnels using mcp and telegram desktop as a backup to the mcp path.](https://twitter.com/146463747/status/2108244624342970467)" The same post adds, "[grok bot voice is the best interface of any bot i have used bar none and its not close.](https://twitter.com/146463747/status/2108244624342970467)"

At home, one parent gave a child ChatGPT as a tutor, and Grok Bot now reads the results each day and updates the tutor's instructions. "[chatgpt is the tutor.](https://twitter.com/1015038773959090178/status/2101230809382535655)" "[grok bot is the manager of the tutor running in the cloud.](https://twitter.com/1015038773959090178/status/2101230809382535655)"

These users credit Anthropic and OpenAI with the heavy coding and Nous Research with a stable self-hosted agent. One user answered a skeptic this way. "[it is only through @bot as a coordinator that i can really get something out of codex, claude, and so on.](https://twitter.com/804676521529110528/status/2105567568463307024)" The driver above stays loyal to Hermes Agent. "[hermes is now way too good and stable at this point to switch off of it.](https://twitter.com/1504949293291085824/status/2094585207420604706)"

## Why they dispatch

xAI's pages do not say why anyone would send Grok Bot's work elsewhere. Under a second written rule, 39 of the 202 say they send work to another maker's agent to save Grok Bot's allowance or to use a subscription they already pay for. "[i use grokbot to launch sessions on cursor/claude/chatgpt to conserve tokens.](https://www.reddit.com/r/cursor/comments/1wznrfb/grokbot_vs_cursor_token_limits/peesmvh/)" Another keeps it short. "[i have grokbot delegate certain tasks to openai models.](https://twitter.com/28800334/status/2096708769388667037)" It adds, "[helps with the usage limit problems as well.](https://twitter.com/28800334/status/2096708769388667037)" The rest give no reason, and we did not infer one.

## How it works in the product

Each Bot "works on a persistent cloud computer with a browser, filesystem, and terminal", in the words of xAI's overview. xAI's page on the computer adds that "Command-line credentials are shared" across every Bot on the account. A user who signs in to Claude Code or Codex once has given every Bot a coding agent to call.

Grok Bot already lets users choose the model in one place. For the Cursor agents it starts, Cursor's support team wrote in August that they use the account's default model unless the user tells Grok Bot which model to use, and Grok Bot passes that along.

## What users still cannot see

Grok Bot has never let users pick its own model. Cursor's help page on the models behind Grok Bot asks "Can I see which model handled a Grok Bot request?" and answers "No." It also says "there is no model picker in Grok Bot settings." Requests run through the whole window. Thirty people asked to see, pick or pin the model in the five weeks before 7 October. From 7 to 11 October, 21 people made such a request, 20 of them for the first time, and seven more asked to opt out of other makers' models. In those five days, 83 people welcomed the change.

Routing showed up before the announcement. Five hours before Musk's post, one user wrote that "[everything i used today on grok bot shows up as the opus 5.5 high fast model under my current usage in cursor dashboard.](https://twitter.com/1379927307989704707/status/2107644203069034566)" Two more said that day that their usage or call records named Opus 5.5, while the help page says usage is recorded "as Grok Bot usage, not by provider or model name".

One new user welcomed the top models and asked in the same post. "[one wish: show me which model is doing the work.](https://twitter.com/2082536149458046976/status/2109288148488618264)" Another saw most of a week's usage go in one morning after the switch ([post](https://twitter.com/1726538544938696704/status/2108205309646999749)). Cursor's help page says routing a task to Opus 5.5 "doesn't change what it costs you per token." One heavy user replied that the model does not matter as long as the bot gets the job right. The routing changes Grok Bot's own back end, and users can still dispatch to their own Claude Code and Codex as before.

## What to build

Let a Bot hand its heavy work by default to an agent the user has signed in on the shared computer, and show in usage which worker and model spent each task's allowance. The dispatchers already do the first part by hand to save their allowance, and the routing now spends part of that allowance on Opus without naming it in usage. xAI's Grok Bot pages describe no default worker, and Cursor's help page says usage is not broken down by model. The shared-credentials design already supports the first part.

## Where it could reach

The team design stops at one account, and households run into that limit. Thirty-one people asked for a way to share bots with a spouse or family, or wired a bridge themselves through email, Slack, webhooks or a message bus. One described the workaround. "[i was but couldn’t get a good solution for me and my wife for shared family organisation even resorting to having it monitor my email every 10 min for messages from my wife directed at my bot but that just sucked as an experience](https://twitter.com/1591951866191302656/status/2096687836527468609)"

xAI's Team Bots do not cover this. The documentation says "Team Bots are available to members of a team" and suits them to "one shared job". It also says, "For an everyday assistant that works from your own accounts and memories, keep a personal Bot." A household version would let each person keep their own Bots and memory, while one chief of staff can message another's and share chosen Bots.

## Limits

The dispatchers are a small part of the team users, and they lean toward developers posting on X. We read what people say and cannot see which model ran. Some users put Hermes Agent or OpenClaw in charge of Grok Bot to save usage, and those did not count. Some users drive their Claude or ChatGPT consumer accounts through Grok Bot's browser; each maker's terms govern that use.

## How we read this

We read public posts about nine personal agents over seven weeks, using Feedback Bench 1.1. AI agents read and classified the posts under written rules, and the bench labels come from large language models. AI agents read every Grok Bot post the bench labelled as praise for multi-agent teams and counted people under a written rule. People counted up to 11 October were re-read by a second reader; the people added in this refresh were read once. For this piece, AI agents searched the posts about all nine agents across the whole window, read every person found under the rules above and coded the agents each person names by reading, and a second AI agent re-read every person they counted. The model requests, welcomes, opt-outs, allowance reasons and household posts were each read in full under rules written before counting. We removed referral posters, sellers, template makers, staff, bots, news relays and borderline accounts before counting, and quote no flagged account. The bench position is the bench's own, computed the same way for every agent.

## Sources

- Feedback Bench, [Multiple agents, teams and delegation between them](https://feedbackbench.com/criteria/auto.multi_agent.md) and [Grok Bot](https://feedbackbench.com/agents/grok-bot.md), built 2026-10-19.
- Elon Musk, [post on X](https://x.com/elonmusk/status/2107724314451878104), 7 October 2026, 06:46 UTC, read 2026-10-17.
- xAI, [Introducing Grok Bot](https://x.ai/news/introducing-grok-bot), 11 August 2026, read 2026-10-17.
- xAI, [Grok Bot documentation, Overview](https://docs.x.ai/grok-bot/overview), last updated 7 October 2026, read 2026-10-17.
- xAI, [Grok Bot documentation, Use the computer and apps](https://docs.x.ai/grok-bot/computer-and-apps), last updated 12 October 2026, read 2026-10-17.
- xAI, [Grok Bot documentation, Team Bots](https://docs.x.ai/grok-bot/team-bots), last updated 15 October 2026, read 2026-10-17.
- Cursor, [Which AI models power Grok Bot?](https://cursor.com/help/grok-bot/models), undated, read 2026-10-17, and the [archived copy](https://web.archive.org/web/20261008020759/https://cursor.com/help/grok-bot/models) of 8 October 2026.
- Cursor forum, [reply from Cursor support](https://forum.cursor.com/t/query-about-grok-bot-cursor-agent-usage-and-model-selection/169160/5), 24 August 2026, read 2026-10-17.