pindoc
Code-pinned team memory for AI coding agents — typed artifacts (Decision/Analysis/Debug/Task), MCP-native workflow, self-host with Docker Compose.
README
Pindoc
<p> <a href="./README.md"><img alt="English README" src="https://img.shields.io/badge/lang-English-2563eb.svg?style=flat-square"></a> <a href="./README-ko.md"><img alt="Korean README" src="https://img.shields.io/badge/lang-%ED%95%9C%EA%B5%AD%EC%96%B4-6b7280.svg?style=flat-square"></a> </p>
Code-pinned team memory for AI-assisted development. Agents write the durable record; humans review, discuss, and steer.
Pindoc is a self-hosted project memory system for teams working with AI coding agents. It turns useful agent discoveries into typed artifacts: decisions, debugging paths, task closeouts, verification notes, and code-linked analyses. Every artifact is scoped to a project area and pinned back to commits, files, URLs, resources, or related Pindoc artifacts.
It is still the wiki you never type into, but the point is not automation for its own sake. Pindoc keeps the parts of agent work that teammates and future agents can reuse.
Why It Exists
AI coding sessions are productive, but team context still falls through the cracks:
- a debugging path dies with the terminal session,
- the same decision is re-explained to every new agent,
- useful analysis stays in one operator's chat instead of becoming team knowledge,
- duplicate documents accumulate across wikis, issue trackers, PRs, and commit messages,
- in real project environments, the person who finds a problem cannot always change the code immediately; structured evidence helps the team discuss and decide.
Pindoc turns agent work worth keeping into searchable, code-pinned team memory. The next teammate or coding agent can ask Pindoc what matters before it edits.
What Makes Pindoc Different
- Collaborative memory layer: artifacts are written for teammates and future agents, not as private chat summaries.
- Agent-only write surface: the Reader UI is for reading and review; durable writes go through agents.
- MCP-native workflow: tools such as
pindoc.context_for_task,pindoc.artifact.propose, andpindoc.task.queueregulate agent behavior instead of acting as a thin CRUD API. - Typed artifacts: Decision, Analysis, Debug, Flow, Task, TC, Glossary, and domain-pack types.
- Code-pinned memory: artifacts can point to commits, files, line ranges, resources, URLs, and related artifacts.
- Record-worthy by design: Pindoc avoids raw chat archives and keeps only decisions, analyses, debug paths, verification, and task context with future value.
- Multi-project daemon: one
/mcpendpoint can serve multiple projects; each tool call carriesproject_slug. - Self-host first: Docker Compose brings up Postgres, pgvector, the Pindoc daemon, and the Reader SPA.
Public Demo
A read-only public demo is a follow-up track and is not part of this OSS
release. Until it ships, the README, docs/, and a
self-hosted clone are the primary proof. Operators who want to evaluate
Pindoc end-to-end run docker compose up -d --build and inspect their own
artifacts.
The follow-up demo plan stays in Public Demo Plan for when a hosted instance is appropriate.
Quick Start
Prerequisites:
- Docker 27+
- 2 CPU cores and 4 GB RAM recommended for local dogfood or small-team use
- 5 GB free disk recommended for Docker images, Postgres data, and the embedding cache; 2 GB is a light fresh-clone minimum
- outbound HTTPS on first run so the bundled EmbeddingGemma model and runtime can be cached
- Go 1.25+ only for host-native development
- Node 20.15+ and pnpm 10+ only for web development outside Docker
The default Docker path includes semantic search through a bundled EmbeddingGemma Q4 ONNX provider, so no embedding sidecar is required. See System Requirements for minimum and optional deployment profiles.
git clone https://github.com/var-gg/pindoc.git
cd pindoc
docker compose up -d --build
Open the Reader:
http://localhost:5830/
On a fresh instance, / first asks for the owner identity (display name and
email), then routes to the first-project wizard. To open the project wizard
directly after identity setup:
http://localhost:5830/projects/new?welcome=1
Connect an MCP Client
The Docker daemon exposes one account-level MCP endpoint:
{
"mcpServers": {
"pindoc": {
"type": "http",
"url": "http://127.0.0.1:5830/mcp"
}
}
}
Project scope is not encoded in the URL. Agents pass project_slug on
project-scoped tool calls. Workspaces generated by pindoc.harness.install
store that slug in PINDOC.md frontmatter.
Common Workflows
Ask an agent to start work with project context:
Use Pindoc context before editing. Find the current project, inspect assigned
Tasks, then implement the next acceptance item.
Typical MCP loop:
pindoc.workspace.detectpindoc.task.queuepindoc.context_for_task- code or doc work
pindoc.artifact.propose- update Task acceptance and closeout state
Configuration
The default Docker path is single-user and loopback-only:
| Variable | Default | Purpose |
|---|---|---|
PINDOC_DAEMON_PORT |
5830 |
Host port used by Docker Compose. |
PINDOC_PROJECT |
pindoc |
Default project for unscoped reads/config. |
PINDOC_PUBLIC_BASE_URL |
http://127.0.0.1:${PINDOC_DAEMON_PORT} |
Public base URL used in generated links and OAuth metadata. |
PINDOC_BIND_ADDR |
127.0.0.1:5830 |
Security intent. Non-loopback values require an IdP or explicit public unauthenticated opt-in. |
PINDOC_AUTH_PROVIDERS |
empty | Identity providers enabled for external requests. Current provider: github. |
PINDOC_ALLOW_PUBLIC_UNAUTHENTICATED |
false |
Explicit opt-in for external exposure without an IdP. Use only behind a trusted network/reverse proxy. |
PINDOC_FORCE_OAUTH_LOCAL |
false |
Development flag that routes loopback /mcp calls through OAuth bearer auth for local QA. |
PINDOC_ALLOWED_ORIGINS |
empty | Comma-separated CORS allowlist. Empty means same-origin only; set explicit origins for cross-origin frontends. |
PINDOC_DEV_MODE |
false |
Development-only flag that permits wildcard CORS for local tooling. Do not enable on public instances. |
Do not expose a writable daemon to the public internet without an identity
provider. For a public read-only demo, keep /mcp and mutating HTTP routes
blocked at the reverse proxy; see SECURITY.md and
docs/22-public-demo.md.
The daemon also sets baseline security headers itself, including nosniff,
clickjacking protection, referrer policy, and hardened asset-blob CSP.
For a writable public or cross-device instance, follow
docs/oauth-setup.md. It covers GitHub OAuth App setup,
the ${PINDOC_PUBLIC_BASE_URL}/auth/github/callback callback rule, runtime
MCP client registration, and local OAuth QA with PINDOC_FORCE_OAUTH_LOCAL.
Development
# Run Go tests. Integration tests that need Postgres are skipped unless
# PINDOC_TEST_DATABASE_URL is set.
go test ./...
# Web checks.
cd web
pnpm install --frozen-lockfile
pnpm typecheck
pnpm test:unit
pnpm build
# Full image build.
docker build -t pindoc-server:local .
To test the OAuth bearer path locally while still connecting through
127.0.0.1, set PINDOC_FORCE_OAUTH_LOCAL=true; the daemon will warn on boot
and require Bearer tokens for loopback /mcp calls.
On Windows hosts without a local C toolchain, run Go tests through Docker:
docker run --rm -v "${PWD}:/work" -w /work golang:1.25 go test ./...
Documentation
- Documentation Hub
- Public Demo Plan
- Public Demo Story Path
- Record-worthy Artifact Policy
- Public Release Checklist
- Contributing
- Security
- Design source notes
Status
Pindoc is in active dogfood. The local self-host path, Reader UI, project/area model, artifact proposal flow, task queue, revision history, summaries, and real embedding provider path are implemented. The public OSS launch track is focused on first-run reliability, a read-only dogfood demo, CI, security docs, and clearer collaborative positioning.
License
Apache License 2.0. See LICENSE.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。