# OpenClaw users complain about setup at more than twice the rate of Hermes Agent users

*OpenClaw's getting-started page says setup takes about five minutes. Of the people who judged OpenClaw, 11.6% complained about setting it up, more than twice the 5.3% for Hermes Agent, the other personal agent people install themselves.*

Capability case, OpenClaw. By Feedback Bench, 15 November 2026. Feedback Bench 1.1. Posts 31 August 2026 to 8 November 2026. Web: https://feedbackbench.com/#/articles/openclaw-users-complain-about-setup-at-more-than-twice-the-rate-of-hermes-agent-users

In October an OpenClaw user on X described the one time they set it up:

> I have set up an openclaw server exactly once and plan on never touching it until I can do the same thing as a one touch solution. The entire process from install to account linking to Discord linking took over two hours of tinkering and broke on multiple occasions.

[A user on X, 2026-10-07](https://x.com/HWM_Alex/status/2107750624687448497)

OpenClaw's getting-started page describes the same job, from install through onboarding to a first chat, and says it takes about five minutes. In the ten weeks of posts we read, the distance between that page and what people went through shows up as a complaint OpenClaw users make far more often than users of other personal agents.

Of the 7,393 people who judged OpenClaw from their own use, 858 complained about the work it took to set it up. That is 11.6%. Among people who judged any other personal agent, it is 1.3%.

Hermes Agent is next, at 5.3%, and we come back to it below. Every other agent with more than 300 judges is at 1.1% or under.

That is 8.7 times the field, but the field is a loose comparison. More than a third of its people wrote store reviews, mostly of consumer apps, while most of OpenClaw's judges posted on X.

Comparing like with like narrows the gap without closing it. On X alone, OpenClaw is 6.9 times the field. On Reddit it is 3.5 times.

**Exhibit 1: Setup complaints by personal agent.** Share of people who judged each agent and complained about setup, for agents with more than 300 judges. OpenClaw is 11.6% (858 of 7,393), more than twice Hermes Agent at 5.3% (460 of 8,715). Every other agent shown is at 1.1% or under. The field row is everyone who judged another personal agent and did not judge OpenClaw, 1.3% (769 of 57,607).

|  | Share of judging people |
|---|---|
| OpenClaw | 11.6% (858 of 7,393) |
| Hermes Agent | 5.3% (460 of 8,715) |
| Instinct | 1.1% (4 of 366) |
| ChatGPT Dots | 0.9% (22 of 2,527) |
| Grok Bot | 0.7% (137 of 19,769) |
| Cue | 0.6% (2 of 359) |
| Muse | 0.5% (129 of 24,659) |
| Gemini Spark | 0.5% (5 of 964) |
| Hark | 0.5% (11 of 2,175) |
| The field | 1.3% (769 of 57,607) |

*31 Aug to 8 Nov 2026. Counts are people, not posts. OpenClaw judges are removed from every other agent and from the field.. Agents with 300 or fewer judges are left out: Claude Duet (4 of 200), Fo (0 of 187), Poke (0 of 153) and Wabi (0 of 143).. Source: Feedback Bench labels, build of 9 Nov 2026.*
Setup-complaint rate by agent. OpenClaw and Hermes Agent first, then every other personal agent, Grok Bot named. OpenClaw 858/7,393 (11.61%), Hermes Agent 460/8,715 (5.28%), field 769/57,607 (1.33%).

## Installing it yourself explains part of it

The bigger question is whether this is just what happens when you install and run an agent yourself. Hermes Agent is the fair test, because it is the other personal agent people install themselves and it has enough judges to compare.

Hermes users do complain about setup more than the field, at 5.3%. So installing is part of the story. OpenClaw's rate is still 2.2 times Hermes's.

The gap holds inside each channel, though it is smaller there. On X, OpenClaw is 2.5 times Hermes. On Reddit it is 1.9 times.

Most of that evidence comes from X. With X left out altogether, the gap to Hermes narrows to 1.5 times.

**Exhibit 2: The gap to Hermes Agent holds on X and Reddit, and narrows without X.** On X, OpenClaw is 2.5 times Hermes Agent and 6.9 times the field; on Reddit it is 1.9 times Hermes and 3.5 times the field. Without X, the gap to Hermes narrows to 1.5 times.

|  | Share of judging people |
|---|---|
| OpenClaw, X | 11.8% (762 of 6,441) |
| Hermes Agent, X | 4.8% (299 of 6,258) |
| Field, X | 1.7% (458 of 26,696) |
| OpenClaw, Reddit | 9.4% (41 of 438) |
| Hermes Agent, Reddit | 5.0% (84 of 1,683) |
| Field, Reddit | 2.7% (104 of 3,875) |
| OpenClaw, without X | 10.1% (96 of 952) |
| Hermes Agent, without X | 6.6% (161 of 2,457) |
| Field, without X | 1.0% (311 of 30,911) |

*31 Aug to 8 Nov 2026. Counts are people, not posts. Hermes Agent and the field have OpenClaw judges removed. Without X is Reddit, store reviews and other combined.. Store reviews are 37.3% of field judges (21,482 of 57,607). OpenClaw has 2 of 56 there and Hermes has none.. Source: Feedback Bench labels, build of 9 Nov 2026.*
The gap to Hermes Agent by channel: OpenClaw, Hermes Agent and the field on X, on Reddit, and with X removed. OpenClaw is 2.5 times Hermes on X, 1.9 times on Reddit and 1.5 times without X.

Some people who tried both say so directly. One user who had "[tried both openclaw and hermes](https://x.com/Mcheitiheitison/status/2114086951212990946)" wrote that Hermes was much easier to set up.

Hermes users also praise its setup more often than OpenClaw users praise OpenClaw's, 3.2% against 2.8%. That argues against the idea that OpenClaw users just talk about setup more.

## It is there every week

The gap did not come from one bad week. In every week from 31 August to the week of 2 November, OpenClaw's complaint rate sat between 5.7 and 10.9 times the field's.

**Exhibit 3: OpenClaw's setup-complaint rate was 5.7 to 10.9 times the field's every week.** Weekly share of OpenClaw judges who complained about setup, and how many times the field's rate that is. Every week from 31 August to the week of 2 November sits between 5.7 and 10.9 times.

|  | Share of judging people |
|---|---|
| 31 Aug | 11.3% (213 of 1,888), 6.0 times |
| 7 Sep | 9.3% (93 of 1,002), 5.7 times |
| 14 Sep | 8.1% (75 of 927), 7.4 times |
| 21 Sep | 8.6% (65 of 757), 8.6 times |
| 28 Sep | 8.6% (106 of 1,233), 7.8 times |
| 5 Oct | 8.5% (93 of 1,099), 7.7 times |
| 12 Oct | 8.6% (66 of 770), 7.4 times |
| 19 Oct | 8.8% (69 of 782), 10.9 times |
| 26 Oct | 8.3% (62 of 744), 6.7 times |
| 2 Nov | 9.2% (66 of 715), 7.7 times |

*31 Aug to 8 Nov 2026. Counts are people, not posts. A person is counted in a week when one of their complaint posts, and one of their judging posts, falls in that week.. The field is everyone who judged another personal agent and did not judge OpenClaw.. Source: Feedback Bench labels, build of 9 Nov 2026.*
The same complaint rate by week, OpenClaw against the rest of the field. Every week from 31 August to 2 November sits between 5.7 and 10.9 times.

The posts name specific steps. One user had the gateway running but could not get it to accept the gateway token: "[Have tried every way around that i have researched and still same error.....any help would be appreciated. Bout 2 seconds from giving up on running local.](https://x.com/Jenn92030/status/2112217017851806171)" Another downloaded the new desktop app and landed in what they called "[gateway update hell](https://x.com/whydowebuy/status/2107916304263094576)," with the app asking for an update that Windows said was already done.

The getting-started page measures setup in minutes. One user who tried it now measures every agent another way: "[My first benchmark now is how long it takes before the agent saves me more time than I spend configuring it](https://x.com/HRadka/status/2113040035209695713)."

## How we counted

The build is the one dated 9 November 2026. Posts run from 31 August through 8 November 2026. Counts are people, not posts and not user-weeks. One author string is one person. Every judging post that person wrote about an agent is folded into a single yes or no.

A person counts as judging an agent when a post is marked as judging, the post is about that agent, the firsthand check is true, and the author is not dropped by the author screen (automated for every agent; staff, promotion, or declined for that agent) or by a staff handle on the agent. An agent that launched inside the window contributes posts from the Monday of its launch week. OpenClaw, Hermes Agent, and Muse are in from 31 August. Dots from 5 October, Hark from 12 October, Fo and Cue from 28 September, Wabi from 19 October, Duet from 26 October.

A complaint is a `access.setup` label with polarity complaint. Praise uses the same criterion with polarity praise. Mixed labels count in neither. The criterion asks how much configuration, instruction, and teaching the user had to do before the agent was useful. The two rates use the same rule.

The field denominator is every person who judged some other personal agent, with anyone who also judged OpenClaw removed, so the two sides share no one. With the about filter that denominator is 57,607. Without it, it is 57,612. The five-person gap is judging rows that are not about the agent. The Hermes comparison removes those same people from Hermes's side. A same-channel rate keeps only the authors whose id is on that channel. The corrected bar for the eight candidates is 0.05/8 = 0.00625. The scan-wide bar is 0.05/256 = 0.000195. The Hermes comparison on the full sample clears both.

Leaving out the busiest author, or any one source channel, leaves the field comparison in the same direction and still under both bars. The week-by-week rates do the same. The label file has no thread id and no subreddit. Reddit accounts for 41 of the 858 complainers, and dropping every Reddit author leaves the result, so no single subreddit can be carrying it.

X holds 762 of the 858 complainers. With X removed, 96 of 952 OpenClaw judges complained against 311 of 30,911 in the field, still 10 times, p = 2.8×10⁻¹³³. Against Hermes alone with X removed the ratio is 1.54 (96 of 952 against 161 of 2,457), chi-square 12.28, p = 4.59×10⁻⁴. That p is under 0.00625 and over the scan-wide bar of 0.000195, so the Hermes comparison does not clear the stricter bar once X is removed. That is the caveat on rule (d) for the Hermes comparison. The busiest author wrote 9 of the complaint posts, across three weeks. Sixteen OpenClaw judges both praised and complained about setup; they count once in each rate.

The quotes are from posts marked as clear setup complaints in the committed quote check (`data/labels/quote-checks-personal-2026-08-31.jsonl.gz`), read on X on 14 November 2026.

The getting-started page was read on 14 November 2026. It describes a five-minute path and a Quick start. It does not state this rate, or a comparison with other agents.

## Sources

- Feedback Bench personal labels, `data/labels/labels-personal-2026-08-31.jsonl.gz`, firsthand file, and author screen, build of 9 November 2026, window 31 August to 8 November 2026. Recount: `recount.py` in this folder.
- Criterion `access.setup`, "Effort needed to set up and train the agent", in `pipeline/criteria_personal_v1.json`.
- OpenClaw, Getting started, read 14 November 2026: [docs.openclaw.ai/docs/getting-started](https://docs.openclaw.ai/docs/getting-started). The page opens with install, onboarding, and a chat "in about 5 minutes."
