AI Decision System
Replayable, offline-verifiable records for AI-assisted go/no-go decisions: experiment promotions, typed model answers and human-approved actions.
翻译准备中
尚未检测到更新。GitHub 发布版本后会显示在这里。
个人 GitHub 账号下公开仓库的所有者可以登录并验证所有权,自行管理介绍,无需管理员批准。
登录并生成验证文件,在 30 分钟内将其提交到此仓库的默认分支。其他分支或分叉仓库中的文件无效。然后在下方检查文件,即可无需管理员批准直接管理项目。请保留文件以维持管理权限。
如需登录,将前往登录页面。登录后请回到此处,再次点击所选的验证按钮。
在项目文件夹中运行 /nomorevibe verify,即可验证所有权并自行管理介绍。
一分钟完成验证Self-hosted execution and governance for reproducible AI-powered software delivery over private code.
翻译准备中
取自仓库 README
正在加载完整 README。 以下为已保存的摘录。
Self-hosted execution and governance for reproducible AI-powered software
delivery over private code.
Coding agents can write code. MAIster makes the process around them repeatable:
versioned delivery flows, isolated workspaces, scoped capabilities, human
decisions, evidence gates, cost controls, and promotion to a pull request or
local target branch.查看完整 README ↗Self-hosted execution and governance for reproducible AI-powered software delivery over private code.
Website · Public documentation · Getting started · Architecture · Releasing
Status: beta until v1.0 — the beta ends with full multi-host support. Versioning and release mechanics are described in RELEASING.md.
Coding agents can write code. MAIster makes the process around them repeatable: versioned delivery flows, isolated workspaces, scoped capabilities, human decisions, evidence gates, cost controls, and promotion to a pull request or local target branch.
Task → version-pinned Flow → isolated worktree → ACP coding agent
→ human input + evidence gates → review → PR or local merge
MAIster is designed for a technical owner or small engineering team operating multiple private repositories and coding agents on its own infrastructure.
| Area | Purpose |
|---|---|
web/ | Next.js control plane, Postgres state, boards, review, and promotion |
supervisor/ | Fastify daemon that owns ACP sessions and agent processes |
mcp/ | MCP facade for governed agent access to MAIster operations |
site/ | EN/RU product website published at imaister.dev |
site-docs/ | EN/RU public product documentation |
docs/ | Engineering contracts, architecture, APIs, and operating details |
.ai-factory/ | Plans, adversarial reviews, ADR context, and the design record |
The web tier and supervisor run as separate Node processes on the same host and communicate over HTTP and server-sent events. PostgreSQL is the durable control plane store; run workspaces use Git worktrees. External ACP adapters remain the agent runtimes.
Prerequisites: Node 24, pnpm, Git, Docker with Compose, and a coding agent signed in on this host (Claude Code, Codex, Gemini CLI, OpenCode, or MiMo). Only Postgres runs in Docker; the web tier and the supervisor run on the host because they spawn agent CLIs and work on local git repositories. Linux and macOS are supported; on Windows run everything inside WSL2 with Docker Desktop's WSL integration.
One command clones the repository, installs dependencies, writes the env files
with a generated AUTH_SECRET, starts Postgres, applies the migrations, and
builds the MCP facade:
curl -fsSL https://imaister.dev/quickstart.sh | bash
The script is scripts/quickstart.sh. Read it first if
you prefer, then run ./scripts/quickstart.sh from a checkout; it never
overwrites an existing env file and is safe to repeat. Then start both host
processes:
cd maister
pnpm dev # supervisor on :7777, web on :3000
Open http://localhost:3000/login and sign in as [email protected] with
the password maister-admin; the first login asks for a new one. Before the
first Run, open Settings → ACP runners: on a fresh install the first visit
registers a native runner for each adapter found on the host (for example
claude-code); wait for Ready. Then
add a repository under Projects → Add project (a manifest that names a
default_runner needs that runner Ready first).
git clone https://github.com/maister-dev/maister.git
cd maister
pnpm install --frozen-lockfile
cp .env.example .env
cp web/.env.sample web/.env.local
cp supervisor/.env.sample supervisor/.env
# Put the output in AUTH_SECRET in web/.env.local before starting the web tier.
openssl rand -base64 33
docker compose up -d --wait postgres
pnpm --filter maister-web db:migrate
pnpm --filter maister-web db:migrate:brain
pnpm --filter @maister/mcp build
Then run the two processes in separate terminals:
pnpm --filter @maister/supervisor dev # http://localhost:7777
pnpm --filter maister-web dev # http://localhost:3000
The complete setup, adapter requirements, database workflow, and first-run checks are in Getting Started.
MAIster is under active development and is already dogfooded for governed agent delivery. It is suitable for evaluation and self-hosted experimentation, not a turnkey production platform. Current productization gaps include stronger process/container isolation, OIDC/SSO/MFA and organization administration, automated backup/restore drills, notification routing, and broader end-to-end qualification. See the Product View for current scope.
MAIster drives coding agents through the Agent Client Protocol (ACP) and its vendor-neutral SDK. The supervisor spawns one adapter process per session and enforces capabilities at that boundary.
| Agent | ACP adapter | Status |
|---|---|---|
| Claude Code | claude-agent-acp | Ready by default |
| Codex | codex-acp | Ready by default |
| Gemini CLI | built in (gemini --acp) | Gated by adapter diagnostics and smoke evidence |
| OpenCode | built in (opencode acp) | Gated by adapter diagnostics and smoke evidence |
| MiMo Code | built in (mimo acp) | Gated by adapter diagnostics and smoke evidence |
Agents run with the authentication configured on the execution host, including subscription-backed sessions. API providers such as Anthropic, OpenAI, and Anthropic-compatible gateways are runner configuration, with secrets referenced by environment variable name only.
Ready-made Flow packages live in maister-plugins. Each package is pinned by its own git tag and installs from Settings → Package sources.
| Package | Built on |
|---|---|
aif | AI Factory: spec-driven delivery Flows (dev, bugfix, evolve, roadmap, init, loop, qa) with the vendored skills and agents |
pstack | pstack: evidence-first engineering Flows with typed result profiles and an Evaluation Method |
superpowers | Superpowers: governed Flows with structured design, verification, and review handoffs |
openspec | OpenSpec: spec-driven change Flows with typed change and review handoffs |
spec-kit | Spec Kit: Spec-Driven Development Flows |
bmad-bmm | BMAD Method: planning and build Flows, upstream skills, and platform agents |
bmad-tea | BMAD Test Architect: quality Flows and the Murat agent |
bmad-cis | BMAD Creative Intelligence Suite: a governed discovery Flow and creative agents |
core, core-java, core-react, core-pg, core-skill-authoring, env-e2e | First-party MAIster packages: triage, Project Brain, and evaluation agents; stack skills for Java, React, and PostgreSQL; skill authoring; an ephemeral docker-compose end-to-end environment as readiness evidence |
| Start here | What it answers |
|---|---|
| Public docs | How to understand and operate MAIster |
| Getting Started | How to install, configure, and run locally |
| Personal librarian | Conversation, task routing, privacy and setup |
| Configuration | Environment, maister.yaml, and Flow contracts |
| Architecture | Process boundaries and data flows |
| System analytics | Domain behavior and state machines |
| API contracts | Web, supervisor, external API, and event schemas |
MAIster is built with AI Factory, and
the repository deliberately keeps that engineering record in
.ai-factory/: implementation plans, reviews, and the reasoning
that shaped the current architecture. The same method ships as the aif
package above.
Contributions are welcome. Start with CONTRIBUTING.md, follow the Code of Conduct, and use the private route described in SECURITY.md for vulnerabilities. Please do not open public security issues.
MAIster is available under the MIT License.