HakashiKatake/docorbit

Version-aware documentation intelligence for coding agents.
DocOrbit discovers authoritative documentation, resolves it against your project's dependency versions, retrieves task-specific context, and verifies generated code against documentation contracts.
Quick Start •
Project vs Global Storage •
Why DocOrbit? •
See It in Action •
MCP Integration •
15 MCP Tools •
Benchmarks •
Code Verification •
Security •
Architecture
---
---
Quick Start
DocOrbit requires **Node.js ≥ 22.5.0** and has **zero external runtime dependencies**.
> [!IMPORTANT]
> **Project-Based by Default**: If you do not specify any flags, DocOrbit **always defaults to project-based storage** ( /.docorbit/docorbit.db). When you delete or branch a project, its documentation cache is isolated and cleans up automatically — **zero orphaned files, zero global disk bloat, and zero cross-project version collisions**.
1. Run via npx (Zero Installation)
2. Connect to Your Coding Agent (MCP)
When started by an agent (Cursor, Claude Desktop, Windsurf, Zed), DocOrbit **automatically resolves to the active project workspace's SQLite database** (.docorbit/docorbit.db):
Cursor (.cursor/mcp.json)
Claude Desktop (claude_desktop_config.json)
Claude Code (Terminal CLI)
*DocOrbit detects when standard input is a machine pipe and starts the MCP stdio transport automatically in project-local mode. If you prefer a shared user-wide store across all projects, pass -g: ["-y", "docorbit", "mcp", "-g"].*
3. How to Work with DocOrbit in Your Prompts
Once mounted, you never need to manually copy-paste documentation pages into your agent chat. Simply give your agent the official documentation link and tell it to use DocOrbit:
**What DocOrbit does behind the scenes:**
1. **Tree Crawling & SQLite FTS5 Indexing**: The agent automatically invokes docorbit.ingest_doc({ url: "https://docs.stripe.com/payments/checkout" }). DocOrbit crawls the documentation with SSRF protection and indexes endpoints and types in ~24ms.
2. **Context Assembly (Zero JSON Bloat)**: The agent calls docorbit.get_implementation_context(...) and receives 412 tokens of clean, pure GitHub Markdown grounded in active API contracts.
3. **AST Contract Verification**: The agent calls docorbit.check_api(...) to verify code against AST contracts before writing to disk, preventing deprecated methods (such as legacy stripe.charges.create) and runtime parameter crashes.
---
Project-Scoped Storage & Flags
DocOrbit gives developers full control over where documentation is persisted:
| Storage Mode | Location | Flag | Best Used For |
| :--- | :--- | :---: | :--- |
| **Project-Local (Default)** | /.docorbit/docorbit.db | -p, --project | **Default for all workflows.** Zero disk leaks. Clean git isolation (.gitignore). Deletes when project is deleted. No version collisions between projects. |
| **Global Store** | ~/.docorbit/docorbit.db | -g, --global | Shared documentation across multiple ad-hoc scripts or global toolchains without repository workspaces. |
Interactive Selection Prompt
When running docorbit init or docorbit add in an interactive terminal without flags on a new project, DocOrbit will ask:
* Pressing **Enter** directly accepts the default ([1] Project-local).
* Non-interactive environments (CI, agents, pipes, scripts, --json) **automatically default to Project-local** without hanging.
* Passing -p or -g immediately selects that target and skips the prompt.
---
Why DocOrbit?
AI coding agents (Claude Code, Cursor, Windsurf, Devin) frequently produce broken code not because models lack reasoning, but because **the documentation context they receive is flawed**:
* **Wrong Version Collisions**: An agent in a Next.js 14 codebase gets fed Next.js 15 documentation and uses await params, breaking production builds.
* **Bloated HTML**: Generic scrapers dump navigation headers, footers, cookie banners, and script tags, wasting 70%+ of the agent's context window.
* **Missing API Contracts**: Models guess query parameters and request bodies because documentation lack structured OpenAPI/Swagger schemas.