IndustrySoftware engineering / AI-assisted delivery

KungFuTea — a Self-Hosted Delivery Platform for AI Coding Agents

Several AI agents claim tasks, write code, leave evidence, merge and release in parallel; the server enforces the rules, and people step in only to scope work, approve gates and accept results.

In short

KungFuTea is a self-hosted delivery platform for AI coding agents, built by YGG. Claude Code, Codex and other agents connect over HTTP MCP to claim tasks in parallel, write code, file verification evidence, merge under a project lock and release. State transitions, single ownership, gates and evidence are enforced by the server, and people step in only to scope tasks, approve gates and accept results item by item. A perpetual licence costs ¥76,000, on-premise deployment included.

What the agent takes off people's plates

Core capabilities

01

Atomic claims, separate workspaces

Agents call next_task to claim work. The claim happens inside one database transaction, so a task has exactly one holder at any moment. Each task runs in its own git worktree workspace and renews its lease with a heartbeat; if a slot drops out, another slot takes the task over instead of leaving it stuck halfway.

DB transaction · single lease · takeover

02

Rules enforced by the server

Task states can only move along defined paths; an illegal transition is rejected with a reason, and every write checks the token, session and workspace. The rules live on the server, not in a prompt — an agent that forgets them or tries to skip them still cannot get through.

State machine · holder check · idempotency keys

03

Gates stop and wait for a person

Protected paths and a per-change file limit live in the project configuration; hitting either stops the task until a person approves it, and gate approvals can only come from people. Production releases are human-only by default, while test environments can be opened to automatic releases by agents.

Protected paths · file limit · production switch

04

No evidence, no merge

Before submitting, a task needs a passing verification record from the current round — commit, command, scope and result, all four filled in — plus a local screenshot for page tasks. Merges happen one at a time under the project merge lock, and releases are checked against a list first.

Verification record · merge lock · release checklist

05

One MCP toolset for many agents

Agents connect over HTTP MCP: Claude Code, Codex, Claude Desktop, the YGGLab desktop app and any other agent that speaks HTTP MCP use the same tools to claim tasks, write records and take the merge lock. One token maps to one slot, and one slot can work on several tasks at once in separate workspaces.

HTTP MCP · one token per slot

06

Runs on your own server

A single Linux server with Docker Compose and PostgreSQL is enough; no Redis is needed. The platform does not proxy model calls and does not collect source code or full conversations — it hands out tasks, never commands. Every state change is written to an event log, and tokens, passwords and private keys never appear in the logs.

Self-hosted · no model proxying

Workflow

How it works

Steps a person signs offDone by AI agents

Side by side

KungFuTea vs a single AI coding assistant editing the repo vs a task board with manual assignment

AspectKungFuTeaSingle AI coding assistantTask board + manual assignment
Several agents in parallelAtomic claims and separate workspaces; one holder per taskOne thread per session; multiple windows need a person to coordinatePeople assign work and spot conflicts by hand
Where rules are enforcedOn the server: state machine, holder checks, gatesIn prompts and habits, which a model may ignoreIn process documents, relying on discipline
Evidence before mergingA passing verification record from this round is required (commit, command, scope, result)Depends on whether tests were run in the conversationDepends on whether a reviewer checked
Where people step inThree points: scoping, gate approvals, item-by-item acceptanceEvery step needs watchingAssignment, follow-up and acceptance all by people
DeploymentPlatform self-hosted; agents still call their own model vendorsNo platform layer; the assistant calls its model vendor directlySaaS or self-built board, independent of AI

Quantified before / after

Measured impact

Engineering lead
BeforeSeveral AI coding assistants run in separate windows, and finding out who is changing which file and how far along it is means asking each one
AfterTasks, holders, heartbeats and verification records sit in one place, and anything touching a protected path stops for your approval
Requester / acceptor
BeforeMismatches with the requirement surface only on delivery, and the reasons for rework are scattered across chat history
AfterAcceptance criteria are written item by item at scoping time, each is passed or sent back at acceptance, and the reasons stay in the task record
Operations / security
BeforeWorry that AI will touch production directly, and no clear record of who changed what and when if something breaks
AfterProduction releases are human-only by default, and every state change records who, when, from where to where and why

How the agent and human supervision collaborate

System architecture (conceptual)

human

Project owner (scoping · gate approvals · acceptance)

source

KungFuTea server (state machine · leases · gates · merge lock · event log)

agent

AI agent slots (Claude Code / Codex etc., via HTTP MCP)

output

Repository and environments (workspaces · merges · releases)

Flow

  • Project owner (scoping · gate approvals · acceptance)→KungFuTea server (state machine · leases · gates · merge lock · event log)Tasks and acceptance criteria; gate approvals; accept or send back
  • KungFuTea server (state machine · leases · gates · merge lock · event log)→AI agent slots (Claude Code / Codex etc., via HTTP MCP)Hands out tasks, never commands
  • AI agent slots (Claude Code / Codex etc., via HTTP MCP)→KungFuTea server (state machine · leases · gates · merge lock · event log)Heartbeats, verification records, merge-lock requests
  • AI agent slots (Claude Code / Codex etc., via HTTP MCP)→Repository and environments (workspaces · merges · releases)Build in workspaces, merge and release under the lock

Verifiable facts

Numbers and sources

In the six-step delivery flow, people act only at scoping, gate approval and item-by-item acceptance; agents do the rest

Self-reportedKungFuTea product description (kungfutea.ygg.com.cn)2026-11-21

Deployment needs only one Linux server, Docker Compose and PostgreSQL 16 or later, with no Redis dependency

Self-reportedKungFuTea product description, self-hosting section2026-11-21

The platform does not proxy model calls or collect source code and full conversations; it hands out tasks, not commands

Self-reportedKungFuTea product description, design constraints2026-11-21

Maintenance tasks for this website (YGG’s official site) are already dispatched, claimed, verified and merged through KungFuTea

Self-reportedThis site’s KungFuTea task records (the “check automatic publishing” task, 2026-11-21)2026-11-21

Delivery and pricing

Pricing

¥76,000Perpetual licence · on-premise deployment included

Included

  • · Perpetual licence for one KungFuTea installation
  • · On-premise deployment: installed, configured and brought live on the customer’s own Linux server
  • · Q&A support: 20 hours of remote technical support in the first year, covering deployment, configuration and use of the KungFuTea platform itself; additional hours billed separately
  • · Level planning: we help assign permission levels to each agent slot for your team (develop only, may merge, may release)
  • · Free upgrades for one year: all version upgrades within 12 months of delivery are free

Provided by the client

  • · AI agents and model costs (Claude Code, Codex etc. subscriptions or API usage, billed to the customer’s own accounts)
  • · Server and domain (one Linux server, provided by the customer)

Add-ons

  • · Annual service fee from year two: ¥15,200 per year (20% of the price), covering continued upgrades and 20 hours of remote technical support per year
  • · Outside the support scope (quoted separately on request): the customer’s own code and business logic, issues with the AI agents or models themselves, server operations, custom development
WeChat consultation QR code

Scan for the proposal and a demo account

Scan with WeChat to reach the solution lead and request the complete documentation and deployment checklist.

Demo accounts are not handed out publicly: scan the code to reach us on WeChat, tell us what you need, and we will set one up for you.

Where it fits

When this approach does not apply

  • Current stage: agent coordination and release acceptance are in use for YGG’s internal development; a customer-facing portal is still being planned, so customers cannot yet follow progress in a self-service portal.
  • No model is included: customers bring their own agents (Claude Code, Codex, etc.), with subscriptions or API usage billed to their own accounts. Agents send code context to their respective model vendors under those vendors’ privacy terms — self-hosting keeps tasks, records and audit data local, which is not the same as code never leaving the machine.
  • It fits software delivery with a code repository and acceptance criteria that can be written down; tasks with unclear requirements stay in “to evaluate”, and the platform does not do requirements analysis for you.
  • A Linux server of your own is required; project settings such as gates, branch naming, merge and release methods need to be configured first, since agents follow them.
Written by: YGG Technology Platform Engineering TeamPublished: 2026-11-21Last reviewed: 2026-11-21