Coordinator
Plans the work
Turns your request into deliverables and finish conditions, and asks only the teammates it needs.
Ask once. Another AI checks
the version it made.
For people who will not ship an AI result they cannot verify. It runs on your computer and hands you the files plus the record of which version was checked, and how.
Open source · MITRuns on your computerNever sends anything out on its own
The app
The work lives on one page.
Ask once, then stay on one page. The team and the results sit side by side.
These are real screens of the app. Steps 2 and 3 show a run recorded on 14 September 2026 with a local model (Qwen 3.5 9B).
The team
Don't leave it all to one AI.
Whoever wrote it is the last to notice its mistakes. So making and checking are separate jobs.
Researcher
Gathers evidence
Reads your files and public pages, and writes notes with sources.
Maker
Turns it into files
Code, documents, tables. When a finding comes back, it fixes that version and publishes again.
Reviewer
Sometimes says no
Checks each finish condition. Whatever could not be checked stays marked unverified.
What you get
It doesn't end with a chat reply.
What stays with you is a file you can use.
Real files
Code, Markdown, CSV, HTML. Download them together as a ZIP.
Visible revisions
Compare the first draft with the fix, and choose the version you will use.
Step in any time
Send a new direction from the chat while the work is under way.
The whole trail
Who made what, what was checked, what is still unverified. A record you can follow later.
A real record
The reviewer really sends work back.
From the run of 13 September 2026: findings the reviewer raised on a memo comparing three public pages.
research.md r1 → research-final-r2.md
Before · r1 (as recorded)(The comparison table stated a "backend / frontend" layout; section 2-2 had no quote supporting it.)
After · r2 (as recorded)Unverified (F-1): the "backend / frontend layout" in the r1 table has no supporting quote in the fetched text. The statement is removed from the table and not replaced by a guess.
Before · r1 (as recorded)Reference (the official repository linked at the top of the page, https://github.com/agentskills/agentskills): "Apache-2.0 license" …
After · r2 (as recorded)Evidence the link exists (F-2, re-fetched 21:27:35Z): the page header shows a link labelled with the repository name twice. Quote: "… ⌘ I / agentskills/agentskills / agentskills/agentskills"
Limitation: only extracted text was fetched, without href attributes, so it is not confirmed that the link targets https://github.com/agentskills/agentskills.
Limitation: only extracted text was fetched, without href attributes, so it is not confirmed that the link targets https://github.com/agentskills/agentskills.
Before · r1 (as recorded)Note: the page also mentions a ready-made coding agent for the terminal, so non-library users are in scope.
After · r2 (as recorded)Quote (re-fetched 2026-09-12T21:27:36Z): "Deep Agents Code — a pre-built coding agent in your terminal, similar to Claude Code or Cursor, powered by any LLM."
Note (F-3): the source says "a pre-built coding agent in your terminal"; the word "CLI" does not appear. This memo follows the source wording.
Note (F-3): the source says "a pre-built coding agent in your terminal"; the word "CLI" does not appear. This memo follows the source wording.
Result partial (not complete)Events 171Model calls 91Cost $6.07 at list priceTime 23m 51s
The final review stopped at the budget limit, so this run stays "partial". Matching quotes against the original pages is recorded as "unverified", because the review environment could not fetch them. Open the full record
Honest numbers
We keep the records that went badly.
The same small local model, as a team and alone (18 September 2026, 10 tasks, one run each).
Coding tasksgraded by hidden tests · 5 tasks
Research-and-summarise tasksauto-graded · 5 tasks
It is slower.
Across the 10 tasks the team took 164 minutes. One AI took 37.
It was more accurate where tests can check the work.
All five coding tasks were correct. One AI missed two.
Research is still a weak spot.
With a small model, all five were graded wrong: timeouts, and checks that missed gaps.
A small test. The differences may be chance, and this is not a general claim of better quality. Conditions and full results
Your data
It runs on your computer.
No sign-up. No cloud account.
- The work record and files are stored on your computer.
- With a local model (Ollama), your input never leaves.
- If you choose a cloud model, your input goes to that provider.You decide which model to use in Settings.
- Nothing is posted or sent externally.It stops at drafts. Anything that widens permissions asks for your approval.
Under the hood
You can trust it because it keeps the record.
For developers: what actually happens during a run.
An append-only event log
Every message, tool call, check and review is an event in SQLite. The chat, timeline and report are views of that log.
Checks bound to the bytes
A check or review applies to one revision's SHA-256. A new revision starts out unchecked.
Limits in code, not prompts
Budget reservation, tool and write scopes, approvals and the run's time limit are enforced by the runtime.
Commands run isolated
Docker with no network, or the macOS seatbelt. With neither, commands are refused rather than run unsandboxed.
Bring your own model
Claude Code (no API key), the Anthropic API, OpenAI-compatible APIs or Ollama. A real probe call confirms the connection first.
Take the record with you
A REST and SSE API. Export any run as JSONL, or as a ZIP with a SHA-256 manifest.
Get started
One command to your first request.
No Python setup. If you have uv, it just starts.
uvx --from "git+https://github.com/FORIFOR/Multibot#subdirectory=backend" agentteam quickstartWhat you need macOS, Linux or Windows, and uv. If you do not have uv: curl -LsSf https://astral.sh/uv/install.sh | sh (Windows: official steps). No Python setup.
- Start itRun the command above. The app opens in your browser.
- Connect a modelPick Claude Code, an API or Ollama, and run the connection check.
- AskWrite what you need in your own words.
The app is free. If you connect a cloud model, that provider bills you for usage.
Claude Codeno key neededOpenAI-compatible APIsOllamalocal
Questions
How is this different from ChatGPT or Claude?
It is not one AI answering once. Making and checking are separate jobs. You receive files, not a reply, along with a record of what was checked and what is still unverified.
Will the result always be right?
No. The reviewer can miss things. That is why partial and unverified results are shown as they are, and why we publish the runs that went badly. Check important work yourself.
What does it cost?
The app is free and open source (MIT). A cloud model bills you for usage. You can set a budget limit per request, and work stops when it is reached.
Where does my data go?
The work record and files stay on your computer. With a local model your input never leaves. Only if you choose a cloud model does your input go to that provider.
What kind of work suits it?
Work that tests or conditions can check: small tools, data conversion, documents with a fixed format. Research with a small local model is not yet reliable.
Using it at work
Ask whether it fits your work.
We narrow it to one task and its pass conditions, and check together. Paid pilots on customer data, production SLAs, SSO and organisation isolation are not offered as available.