DocOrbit is a version-aware documentation intelligence tool for coding agents, developed by HakashiKatake under the MIT license. Its core goal is to solve the problem of AI coding agents generating incorrect code due to flawed documentation context.
Purpose
DocOrbit is used to provide coding agents with accurate, version-matched documentation context and to verify that generated code conforms to documentation contracts. It targets the following common problems: version conflicts (such as mistakenly using Next.js 15 documentation in a Next.js 14 project), HTML redundancy (wasting the context window), missing API contracts, outdated examples and deprecated interfaces, and prompt injection risks in third-party documentation.
Main Capabilities
- Documentation intelligence: automatically detects the target domain and ranks available formats (llms-full.txt > openapi.json > llms.txt > skill.md > Markdown > HTML sitemap); dynamically ranks sources based on intent (navigation, concept, API, example, implementation); semantic AST slicing preserves heading hierarchy, breadcrumbs, and indivisible code blocks.
- Version intelligence: scans workspace manifests across 8 ecosystems (npm, cargo, go, pypi, composer, rubygems, pub, maven); deterministically resolves project versions through a SemVer confidence ladder; generates a deterministic docs.lock to lock documentation versions.
- Implementation knowledge: parses OpenAPI 3.x and Swagger 2.0 (parameters, JSON request/response schemas, authentication, pagination headers, safely handling $ref cycles); extracts and indexes code examples categorized by language and framework; indexes breaking changes, server/client boundaries, rate limits, and security pitfalls; assembles implementation blueprints with evidence levels.
- Closed-loop verification: a static AST code verifier that evaluates JavaScript, TypeScript, Python, and cURL code without calling an LLM; strict contract checks (endpoint paths, HTTP methods, required parameters, deprecated APIs); returns insufficient_evidence for dynamic expressions to avoid false positives.
- Change intelligence: documentation diff comparison (diff_docs) detects added, modified, deleted, and deprecated endpoints; workspace impact analysis (analyze_impact) scans repository source files to locateaffected lines.
- Agent integration: provides 15 MCP tools, supports stdio and Streamable HTTP; returns token-budgeted Markdown designed for agent consumption; generates export files such as AGENTS.md, CLAUDE.md, skill.md, llms.txt, and docs-map.md.
Usage
DocOrbit requires Node.js ≥ 22.5.0 and has zero external runtime dependencies. It can be run with zero installation via npx. By default it uses project-level storage (<projectRoot>/.docorbit/docorbit.db), and the cache is automatically isolated and cleaned up when a project is deleted or branched. When connecting to coding agents (Cursor, Claude Desktop, Windsurf, Zed), DocOrbit automatically resolves to the SQLite database of the active project workspace; if cross-project shared storage is needed, the -g flag can be passed. In prompts, users only need to give the agent the official documentation link and instruct it to use DocOrbit, and the agent will automatically call ingest_doc to crawl and index, get_implementation_context to assemble context, and check_api to verify code. Running docorbit init or docorbit add <url> in an interactive terminal prompts you to choose project-level or global storage; in non-interactive environments it automatically defaults to project-level.
Applicable Scenarios
Suitable for development teams and individuals using AI coding agents, especially in projects with multi-version dependencies, that require strict API contract verification, or that are concerned about documentation security.
Limitations
The documentation does not provide specific pricing, cost, or platform support information. On security, DocOrbit treats external documentation as untrusted input and provides SSRF protection, non-destructive security annotations, and resource limits (default 10MB streaming byte limit). On performance, cold start is < 40ms and indexing is about 24ms. The repository includes a reproducible 17-task benchmark suite.
