Handbook
What this is, and how it works
You describe what you want built. A manager plans it, hires a team, and the team writes the code — on your accounts, on your servers, in your version control. No coding knowledge assumed.
The idea in one paragraph
Imagine hiring a small software agency. A manager reads your brief, works out the pieces, and hands each person a task. They go away, write the code, test it, and submit it for review. The manager checks the work, sends back what is not good enough, and approves what is. devteam is that agency, staffed by AI. Not one chatbot playing every part — genuinely separate programs running at once, each with its own instructions and its own copy of the code, coordinating through tasks, handover notes, reviews and pull requests.
Why everything is yours
It runs on your machine or your cloud account, spends your AI credits, and pushes to your code repository. There is no hosted service and no account to sign up for. Each user's agents run on that user's own credentials — the software refuses to spend one person's key on another person's work.
Who is on the team
- You
- humanWrite the brief, answer when asked, decide what ships.
- The manager
- AIPlans the work, assigns it, reviews it, decides. Deliberately cannot write code.
- The teammates
- AINamed people with a job, a personality and a memory of what they have built here.
- The scheduler
- ordinary softwareDecides what can start now. No AI, so it costs nothing to run.
- The round table
- AIOptional. Models from different companies argue about your idea before anyone builds.
Why the manager cannot write code
The manager can only create tasks, read reports, ask questions, send work back and approve things. It physically cannot open a file or run a command.
An AI that can both assign work and do the work will, under pressure, quietly do the work itself and report that the team delivered it. That is not hypothetical — it happened during development, and the manager reported completed work nobody had built. Removing the ability removes the temptation.
What happens between your brief and a pull request
- The manager plans. Your brief becomes concrete tasks, recording which must wait for others. Unrelated work runs side by side.
- The team is hired. Roles become named people with personalities.
- Tasks are handed out. Whatever is unblocked starts. When it cannot start everything, it starts whatever is holding up the most other work.
- A teammate builds. Their own private copy of the code, plus handover notes from whoever they build on top of.
- The platform runs the tests — not the AI. The raw result is recorded before the AI writes its summary.
- The work is reviewed. Test result first, write-up second. Teammates can review independently; the manager sees the disagreement, not a verdict.
- Approved or sent back. Approved work becomes a pull request. Two failures escalate to a stronger model.
Tests are run by the plain, non-AI part of the system, and the exit code is recorded before any AI describes what it did. Without that you have only the AI's word that it works, and AIs will describe broken code in glowing terms. With it, the manager judges evidence rather than a claim.
What the machine is made of
Six pieces, and only two of them cost money to run.
| Piece | What it does | Costs money? |
|---|---|---|
| Dashboard | This page. Plain files read off disk — no build step. | No |
| Conductor | One web server: serves the dashboard, answers the API, holds the records, runs the manager, starts teammates. | No |
| Records | A single database file — no separate database server to maintain. | No |
| Scheduler | Decides what starts when and notices stalled work. | No |
| Manager | The AI that plans and supervises. | Yes |
| Teammates | Separate programs that write the code. | Yes |
AI is used for judgement; ordinary software for bookkeeping. Which tasks can start and who is free is arithmetic — asking an AI would cost money on every check and occasionally be wrong in a way that is hard to spot.
Three doors, three different keys
| Door | Who gets in | Why separate |
|---|---|---|
/api/… | You, signed in | Normal use. Login attempts are limited, and the limit survives a restart. |
/internal/… | Teammates, with a shared token | Three requests only: report work, post activity, look up a colleague. A teammate is not a user. |
/api/health | Anyone | A monitor that had to sign in could not tell you that signing in was broken. |
Where things run
| Shape | Where | What is isolated |
|---|---|---|
| Local | A program on your laptop | Nothing — your real records and keys. |
| Sandbox | A second program on your laptop | Own records; every credential blanked, not removed. |
| Staging | Its own space on the cluster | Own records, storage and address. Real keys, and it merges ordinary work — but never into the repository this platform is built from. |
| Production | Its own space on the cluster | Own storage, real credentials. The live system. |
One public entrance serves every environment; which one you reach is decided by the web address, not by a separate machine. That is why staging costs almost nothing to keep running. Traffic only ever goes outward — nothing on the internet can start a conversation with a teammate.
As deployed today a teammate is a separate process, not a separate sealed box. They cannot overwrite each other's work, because each gets its own directory — but they share one container, its filesystem and its memory. The stronger arrangement is built and is a configuration change, not new work. The boundary that is real today is around the whole system, not between teammates.
How much you are involved
| Mode | What happens |
|---|---|
| Supervised the default |
Works on its own, but stops and waits for you on the three decisions that quietly cause damage: accepting work nobody delivered, merging past failing tests, and giving up after repeated failure. Other questions wait up to an hour before it proceeds on its own judgement. |
| Autonomous | Never blocks on you. It still recognises those same three decisions and still records each one loudly for audit — it just does not wait. |
At the end of each sprint you are asked what should change before the next one is planned — the cheapest possible moment to redirect, because nothing has been built yet. It does not block: an overnight run that stalled an hour per sprint waiting for you would destroy the thing you asked for. Your answer is read whenever it arrives.
What it costs
Nothing for the software. A small managed cluster runs the whole thing for roughly $40 a month — the largest single item is the load balancer, not the computing power. AI is billed by your provider directly: in a measured run, a complete small feature written, tested and submitted cost about ten cents.
What is solid, and what is early
The full loop is proven: brief, plan, build, test, review, pull request, run live and repeatedly with real accounts. It has shipped a fix to itself. Over seven hundred automated checks pass, including inside the deployed system.
- Teammates on OpenAI or Gemini are unproven. Built and tested against simulated responses, never a real call. Planning on those providers is proven; building is not.
- Amazon and Google's enterprise AI services are configurable, not supported. Both need authentication that is not a password in a settings box.
- One cloud provider is implemented. The seam for others exists but is untested by use.
- The manager does not interview you before it starts. It asks between sprints, not before the first.
Words used here
- API
- The list of requests a program will answer — the difference between a system you can only click and one you can drive.
- Branch
- A parallel version of the code where work happens without affecting anyone else.
- Canary
- Starting a new version privately, with no real traffic, to see whether it works.
- Pull request
- A formal, reviewable proposal to add a change to the main codebase.
- Sprint
- One round of planning, building, testing and shipping.
- Staging
- A full second copy used to try changes before they reach real users.
- Version control
- The system that stores code and remembers every change and who made it.