OpenMausBot (milind-soni) + Claude Code CLI → DeepSeek / 智谱 GLMPartly tested🇨🇳 CN

@zhengtang364848 tested OpenMausBot, which is widely promoted on X, with DeepSeek and Zhipu GLM: delegation, memory and schedules worked for about ¥3 a round, but the app has to stay running

OpenMausBot is an open-source, Grok Bot-style local bot-team UI (Apache-2.0): a roster of bots on the left, chat in the middle, and a Chief of Staff that hands work to other bots. Rather than repeat the hype, 秋/十三 (@zhengtang364848) pointed the Claude Code CLI at DeepSeek and Zhipu and tested item by item: three small tasks with one prompt, schedule accuracy and catch-up after downtime, a two-step Chief delegation checked against a known answer, long-term memory after restart—and estimated cost from token counts. Verdict: it's a management layer over the command-line AI already on your machine; quick install, punctual schedules, delegation and memory worked, but it brings no model, must stay running, and every scheduled run is a full conversation. We cloned commit 88d54eb and confirmed two behaviours in source—5-minute minimum interval and catching up only the latest missed run—without installing deps, running it or calling any model.

Visit site OpenMausBot (milind-soni) + Claude Code CLI → DeepSeek / 智谱 GLM ↗
Pricing
Free
Region
Global
Where it's from
Chinese community
Added
Nov 1, 2026
Tested on
Nov 1, 2026
Version we used
milind-soni/OpenMausBot commit 88d54eb (0.1.136); author's test: Claude Code CLI → DeepSeek V4 Pro / deepseek-flash, 智谱 GLM-5.3, web dev mode
Categories
Automation & AgentsProductivityCoding
Original author
秋/十三 @zhengtang364848 ↗
Our test
Partly tested
Cost / time
OpenMausBot is open source and free (except the enterprise/ folder); you pay the model: the author's whole test round was a little over ¥3, about ¥0.29 per scheduled run, per DeepSeek's billing.
Source
X @zhengtang364848(长文 / Article) ↗GitHub milind-soni/OpenMausBot ↗OpenMausBot Docs · approval levels ↗

The task

Run a few AI bots from one chat UI on your own machine—scheduled jobs, a chief that delegates, remembered preferences—using cheap Chinese models instead of an overseas subscription.

Steps

  1. Prereqs: Node 24+, pnpm and at least one logged-in agent CLI (the author used Claude Code CLI). OpenMausBot brings no model; the work is done by the `claude -p ...` subprocesses it spawns.
  2. Point Claude Code at a Chinese model: the author says DeepSeek took "only environment variables" (Claude Code's Anthropic-compatible base URL and key—check DeepSeek / Zhipu docs for exact names).
  3. `git clone https://github.com/milind-soni/OpenMausBot`, `pnpm install` (6.1 s for the author), then start the backend and web dev mode; the author skipped the desktop package.
  4. "Create Chief of Staff" only writes a local bot record and costs nothing; billing starts when you send a task.
  5. Create a scheduled routine: minimum interval is 5 minutes—2 minutes returned HTTP 400; use a cron expression for anything tighter.
  6. Test delegation: give the Chief a task that needs two steps (e.g. Researcher counts errors in logs, Writer summarizes), with an answer you computed yourself.
  7. Test memory: state two preferences (how to address you, a sign-off format), restart, open a new chat and see if they're applied.
  8. Estimate cost as tokens × list price; trust the provider's billing page.

Prompt

First have Researcher count the total ERROR lines across the three log files in logs/, the count per file, and the most frequent error type. When it reports back, verify the numbers yourself, then have Writer write a summary of at most 5 lines from the results. Finally, tell me how many times you delegated work.

What the author says

@zhengtang364848's long-form test: pnpm install 6.1 s; three small tasks with one prompt (write a README, write and run a word-frequency script, summarize a web page)—DeepSeek V4 Pro 3/3 at ~14–20 s each; Zhipu GLM-5.3 1/3 on the first round with 3 more chats ending in unrecognized_model, then 3/3 on a retest at ~33–68 s each (one run per task—not a model comparison). Schedules fired ~4 s after the mark while running; a ~7 min 12 s outage missed 3 slots and only 1 was caught up on restart; each run was ~62k input tokens over two steps, ~¥0.29 per run, ~¥8.8/hour at peak if every 2 minutes. The Chief (deepseek-flash) delegated to Researcher, grepped to verify, then sent Writer; numbers matched the answer key (34 ERRORs, 14/9/11 per file, top error DB connection timeout ×12), 24 model requests in ~70 s—but Writer was dispatched twice and the Chief's account of that didn't match the logs. Memory: after restart a fresh chat opened with "唐老板" and ended with "风险灯:黄" (risk light: yellow). The whole round cost a little over ¥3. Not tested: mobile, Composio, cloud computers, OpenCode engine, group chat, long-task stability.

Our test

Cloned milind-soni/OpenMausBot commit 88d54eb (package 0.1.136, engines node>=24), read the README and docs/approval-levels.md, and checked the scheduler's minimum interval and catch-up logic in source. The box has Node 20; we did not pnpm install, start the server, or call DeepSeek / Zhipu or any model, so the author's timings, success rates and costs are not reproduced.

What we found when we tried it:Source matches the author's observations: interval routines with everyMinutes < 5 are rejected (max 1440); a routine late by ≤12 h (CATCH_UP_MS) catches up once at the latest occurrence, older ones are marked missed. docs/approval-levels.md states "Teammates start at Full access like every bot" and a Chief of Staff's level flows down every hop of coordinate_bots. The README lists Boat (boat.dev) as the project's sponsor for per-bot cloud computers; the author's test didn't use that feature.

How to get in

The test write-up is a Chinese X Article that may need login to read in full. GitHub and docs.openmausbot.com are public. DeepSeek and Zhipu GLM APIs need their own accounts (phone verification may apply); Zhipu also offers an international platform.

Who it suits

People comfortable with the command line who already have Claude Code or a Chinese-model key and want a UI to manage a few bots and small scheduled jobs locally. Not for those who need it to run with the laptop closed.

Pitfalls

  • Our source read: `server/routine-schedules.ts` requires everyMinutes to be an integer 5–1440 and rejects anything under 5—matching the author's 400.
  • Our source read: `server/routines.ts` catches up an interval routine that is late by ≤12 h with a single run at the latest occurrence—matching the single catch-up after downtime; other missed slots are not replayed.
  • It must keep running; it ships no model or runtime—install Node 24+, pnpm and a CLI agent first.
  • Every scheduled run is a full conversation (~62k input tokens in the author's test); tight schedules get expensive fast.
  • The Chief once double-dispatched work, costing an extra round, and its report didn't fully match the logs—verify important results.
  • The docs say bots start at Full access and a Chief of Staff's level flows down the delegation chain; switch to "Ask for approval" before letting it touch real files or connected apps.
  • Zhipu GLM's first round hit unrecognized_model errors the author couldn't explain; smoke-test with a couple of small tasks when switching models.

More in this category