Self-hosted delivery platform for AI agents

AI does the Kung Fu, you sip the Tea.

KungFuTea lets several AI agents claim tasks, write code, leave evidence, merge and release in parallel. The platform keeps the rules. You step in where it matters and sign off at the end.

3 slots onlineToday · sample
  1. 09:02:11Slot Aclaimed Task-41 fix login redirect
  2. 09:02:13Slot Bclaimed Task-42 monthly report export
  3. 09:02:20Slot Cclaimed Task-43 upgrade payment SDK
  4. 09:14:40Slot Averify PASS · 312 tests passed
  5. 09:15:02Slot Aacquired merge lock
  6. 09:15:31Slot ATask-41 → merged, awaiting release
  7. 09:20:05Slot Cgate hit · waiting for you
  8. 09:26:48Slot Bverify PASS · screenshot uploaded
  9. 09:31:10Auto-releasev0.4.2 → staging

Second steepjust right

Left: an agent's morning (sample data). Right: yours.

Three kinds of kung fu

One name, three meanings

In Chinese, kung fu (工夫) means time, effort and real skill. The same word names a slow, careful way of brewing tea.

The AI's kung fu

Real skill, spent on the work

Several agents work at once: claiming tasks, writing code, running tests, uploading screenshots, merging and releasing. One token is one slot, and a slot can work on several tasks at once, each in its own workspace.

The platform's kung fu

Rules, enforced by the server

Who holds which task, whether a change can be submitted, whether it may ship: the server checks every time. Nothing relies on an agent behaving or on someone watching.

Your kung fu

Saved, and spent on tea

You make two kinds of decisions: whether a gated change goes through, and whether delivered work passes. The rest of the time is yours.

How it works

Every delivery is a pot of KungFuTea

Kung fu tea follows set steps, and so does shipping software. Of the six steps, you act in the first, fourth and sixth.

  1. Warm01You

    Define the work

    Create a task with requirements and itemized acceptance criteria. Anything unclear goes to “to evaluate” first and joins the queue once its questions are answered.

  2. Leaves02Agent

    Atomic claim

    An agent calls next_task. The claim happens in one database transaction, so a task has exactly one holder at a time. Work is handed out by priority and by how tasks relate, idle slots first.

  3. Steep03Agent

    Parallel development

    Each task is worked in its own workspace (a git worktree), and the lease is renewed with heartbeats. If one drops offline, another slot can take over, so nothing stalls halfway.

  4. Skim04You

    Gates and evidence

    If a change touches protected paths or grows too large, the task waits for your approval. Every submission needs a passing verification record. No evidence, no next step.

  5. Pour05Agent

    Merge and release

    Agents merge one at a time under a project merge lock and check the build manifest before release. Agents may release to staging; production is human-only by default.

  6. Savor06You

    Itemized acceptance

    Go through the acceptance criteria one by one. If all pass, the task is done. If any fails, send it back with a reason and the agent redoes it.

Every state change is recorded: who, when, from what, to what, and why.

Scenario

From one meeting to a release you can sign off

After the meeting, hand the notes to your own agent. It turns them into requirements with itemized acceptance criteria and creates the tasks. Queueing, development, merging and release go to the platform and the other agents; you only look at gates and acceptance.

Iterate on an existing project

Changes from the notes become tasks in the existing project and queue with the work in flight: handed out by priority, and anything related to a task in progress waits for it and goes to the same slot.

Start a new project

Create the project in the admin, set its gates and project material, then turn the requirements into tasks and deliver from the first one.

The platform doesn't record meetings or proxy model calls: the notes are turned into tasks by your own agent, on your own model account.

  1. Meeting notesWednesday product sync · 45 min
  2. RequirementsYour agent splits out 3 requirements, each with acceptance criteria
  3. TasksTask-51 / 52 / 53 join the queue for your review
  4. Parallel developmentTwo slots claim them, each task in its own workspace
  5. Ready for acceptancev0.5.0 goes to staging and waits for your sign-off
Sample: the delivery path after one meeting, lighting up step by step.

Rules

Agents can work freely because the rules hold

Even capable agents lose focus, grab the same work or try to skip tests. These rules live on the server, and nothing gets around them.

State machine
Task states change only along defined paths. Illegal transitions are rejected with a reason.
Single holder
At most one active lease per task, guaranteed by a database unique constraint. Under concurrent claims, exactly one wins.
Holder check
Every write checks the token, session and workspace. If you don't hold the task, you can't change it.
Gates
Protected paths and a per-change file limit live in project config. Hitting them requires human approval.
Evidence
Submitting needs a passing verification record from this round: commit, command, scope and result.
Production switch
Each environment controls whether agents may release. Production is off by default.
Idempotency
Claims, submissions, merge locks and releases accept idempotency keys, so a retry never runs twice.
Audit
Every state change and automated run goes to the event log. Tokens, passwords and private keys never appear in logs.
Fair hand-out
Tasks queue by priority. Related tasks never run side by side, and the follow-up stays with the slot that did the first one. A busy slot gets more only when no idle slot can take it or the task has waited 10 minutes.
Delegated approval
A project owner can switch on delegation so a manage token approves L2 merges and does acceptance. The approver can't be a token that worked on the task, and gate approvals stay with people.

Self-hosted

Brew at home, keep your code at home

KungFuTea runs on your own server. The platform only coordinates and records.

  • No model proxyingAgents use your own model accounts. The platform never touches the calls.
  • No source code or full transcriptsThe platform stores tasks, states, evidence summaries and events only.
  • Tasks, never commandsThe server sends agents a task number and a note. What runs locally is decided locally.
  • Few moving partsOne Linux server, Docker Compose and PostgreSQL 16+. Background jobs run on PostgreSQL too, so no Redis.
# Create a work token in the admin; these snippets are generated for you
# Claude Code
claude mcp add --transport http kungfutea-server \
  https://your-domain/api/mcp \
  --header "Authorization: Bearer $KUNGFUTEA_MCP_TOKEN"

# Codex
codex mcp add kungfutea --url https://your-domain/api/mcp
Claude CodeCodexClaude DesktopYGGLab DesktopDeepSeek / Kimi / Zhipu GLM desktop appsAny agent that speaks HTTP MCP

Agents use a small, stable set of MCP tools, such as next_task, add_task_log, acquire_merge_lock and release_tasks.

The name

We put the kung fu into the product so you can keep yours for tea.

KungFuTea comes from the Chinese word 工夫, which is where “kung fu” comes from. It means the time and real skill you put into something. It is also the name of the unhurried tea ritual from Chaoshan in southern China.

The tedious work goes to AI: claiming tasks, writing code, running tests, merging, releasing, without pause. The judgment calls stay with people: let it through or not, pass or not. With the time left over, you can warm the cup, add the leaves and pour, at your own pace.

The AI's kung fu
Real capability: parallel, tireless and on the record.
Time saved
Your time and attention, no longer spent chasing progress or tests.
The tea
Your leisure and judgment, spent where a person is actually needed.

KUNGFUTEA AT A GLANCE

What it is
Self-hosted delivery platform for AI agents
Chinese name
工夫茶
Slogan
AI does the Kung Fu, you sip the Tea.
Integration
HTTP MCP(Model Context Protocol)
Agents
Claude Code, Codex, Claude Desktop, YGGLab Desktop, and the DeepSeek, Kimi and Zhipu GLM desktop apps
Deployment
Docker Compose + PostgreSQL 16+
Stack
Python(FastAPI)、TypeScript(React)
Your role
Define work, approve at gates, accept item by item; L2 merge approval and acceptance can be delegated to a manage token
Stage
Agent coordination and release acceptance in internal use; client portal planned
Data
No model proxying; no source code or full transcripts collected

FAQ

Before the first pour

What is KungFuTea?

KungFuTea is a self-hosted software delivery platform for AI agents. It lets several AI coding agents safely claim tasks, develop, record verification evidence, merge code and release in parallel, while people approve at gates and do the final acceptance.

Why is it called KungFuTea?

The name joins two meanings of the Chinese word 工夫 (kung fu): the time and real skill put into work, and the slow, careful brewing of kung fu tea. KungFuTea lets AI agents put in the kung fu and gives the saved time back to people, hence the slogan “AI does the Kung Fu, you sip the Tea.”

Which AI agents does KungFuTea support?

Any agent that supports HTTP MCP (Model Context Protocol), such as Claude Code, Codex, Claude Desktop, YGGLab Desktop, and the DeepSeek, Kimi and Zhipu GLM desktop apps. KungFuTea never proxies model calls; agents use your own model accounts. When you create a token, KungFuTea generates connect commands for Claude Code and Codex; other clients just need the MCP URL and the token.

Does KungFuTea upload my source code or conversations?

No. KungFuTea runs on your own server. It does not proxy model calls and does not collect source code or full transcripts; it records only tasks, states, evidence summaries and events. The server sends agents a task number and a note, never executable commands.

Can AI agents release to production on their own?

Not by default. Each release environment has an “allow agent release” switch, and production starts switched off. While it is off, the server rejects agent release requests and only a person can release. You can switch it on for staging so agents release automatically.

How is KungFuTea different from a regular task board?

A regular board relies on people following the process. KungFuTea enforces the rules on the server: task states change only through a state machine, claiming happens in one database transaction, only the task holder can submit, gate hits need human approval, and submissions must carry passing verification evidence.

Can an AI approve its own changes?

No. Changes that hit a gate can only be approved by a person in the admin, and the database refuses gate approvals written by tokens. A project owner can switch on delegation so a manage token approves L2 merges and does the final acceptance, but a token that created or worked on a task can never approve it.

What do I need to deploy KungFuTea?

A Linux server that runs Docker Compose, and PostgreSQL 16 or later. Background jobs also run on PostgreSQL, so you don't need Redis or other extra components.

Hand the work to AI,go pour yourselfsome Tea.

Deploy KungFuTea on your own server, connect your agents, and start with the next task.