Context
Local-first, instant, plug-and-play documentation for AI coding agents, fixing confidently wrong answers caused by outdated training data.
Evidence shows a local-first MCP docs server: queries hit local SQLite with an architecture diagram explaining data flow; auth credentials are documented as stored with 0600 permissions; the maintenance-watch workflow uses minimal, scoped permissions and is detection-only; context remove requires explicit version selection before deleting. Deductions: packages are auto-downloaded from a community registry on first query without user confirmation, and no implementation source code is in the evidence set — download integrity checks, the auth storage implementation, and the security boundary of --host 0.0.0.0 HTTP exposure are README claims only, so dependency security and confirmation score conservatively.
CI runs lint/build/test with --frozen-lockfile and pinned pnpm/Node versions; the watchdog workflow shape-checks every API response and fails loudly on the unexpected, explicitly avoiding silent 'all-clear' failures — thoughtful error design. Deductions: runtime behavior (download failures, SQLite corruption) is only partially evidenced via CLI warning examples (e.g., the few-sections Google-search hint); actual runtime error message quality cannot be confirmed statically.
README covers 9 major MCP clients, cross-platform paths (Linux/macOS/Windows), Docker/HTTP multi-client deployment, language-agnostic doc formats (md/adoc/rst/html), and per-project package scoping via --libs — very thorough scenario fit. Deductions: capability boundaries are only implied ('use context add for anything outside the registry'), with little on poor-quality or very large doc sources; trigger precision cannot be judged as no MCP tool descriptions are provided.
Clear information architecture (quick start → CLI reference → architecture → FAQ → development); install notes per client; concrete FAQ with contribution guide; full Apache-2.0 text; changesets-based versioning with name@version git tags; the maintenance-watch workflow demonstrates active maintenance responsibility (unanswered PRs, release drift, failed nightly publishes). Deductions: no CHANGELOG file in evidence; naming stability depends on the registry/ directory, which is not in the evidence set.
The product targets a real pain point (AIs confidently citing outdated docs); local SQLite + FTS5 + token-budget filtering gives clear marginal value and cost advantages (free, offline, sub-10ms). Deductions: output usability (retrieval quality, token filtering effectiveness) is asserted without evidence, and performance figures like '<10ms' have no benchmark support.
README claims are corroborated by the CI workflows, changesets setup, and watchdog script (e.g., release-drift detection matches the changesets flow); facts and inferences are mostly separable. Deductions: core claims (performance, 100+ packages, query privacy) are single-source README statements; registry definitions, implementation code, and tests are outside the evidence set and cannot be cross-checked.
- First query auto-downloads third-party-built .db files from a community registry — a supply-chain surface; no evidence of integrity checks or signing was available in this review, and enterprise users should verify package provenance.
- context serve --http --host 0.0.0.0 exposes the MCP server to the local network with no documented authentication; confirm before deploying (the 127.0.0.1 default is safe).
- context auth writes real cookies/credentials to ~/.context/auth.; verify local file permissions and avoid shared environments.
- This is a static review without implementation source or test code in evidence; all security and performance claims are unverified by execution.
What does this agent do, and when should you use it?
Context (GitHub: neuledge/context) is a local-first MCP documentation server that solves the problem of AI agents being trained on outdated library docs. It is backed by a community-driven package registry with 100+ popular libraries already pre-built (Next.js, React, Prisma, LangChain, and more). When an agent needs documentation, Context automatically searches the registry, downloads the right SQLite docs package into ~/.context/packages/, and answers queries via SQLite FTS5 full-text search in under 10ms. After the initial download everything is offline, private, and free of subscriptions or rate limits. Beyond the registry, context add builds packages from private repos, local directories, llms.txt websites, or arbitrary URLs, and a Docker image can expose Context as an HTTP server for multi-client deployments.
Context runs as an MCP server (context serve) that exposes documentation lookup to any MCP-compatible AI agent. The core flow: on a doc request, the server searches the community registry, automatically downloads the pre-built SQLite package into ~/.context/packages/, and on subsequent queries uses SQLite FTS5 full-text search to locate relevant sections, applying smart filtering to stay within the token budget. The CLI offers browse (list registry versions), install (fetch a specific registry/name version such as npm/next or pip/django), add (build/install from a git repo, local directory, llms.txt site, blog URL, or .db file), list, remove, auth (per-domain cookie/header storage for subscriber-only content), query (direct command-line lookup), and serve (stdio or HTTP transport). Supported doc formats include Markdown, AsciiDoc, reStructuredText, and HTML; the registry uses YAML definitions covering npm, pip, and maven ecosystems.
- Developers using fast-moving libraries like AI SDK or Next.js whose AI assistant keeps suggesting deprecated APIs (e.g., v5's Experimental_Agent instead of v6's ToolLoopAgent) and needs version-accurate docs injected.
- Teams with internal libraries or private design systems that aren't in any public registry, who want to build doc packages via context add and share portable .db files with teammates.
- Developers who work offline on flights or in coffee shops but still need their AI agent to query library documentation locally.
- Dev teams deploying in Kubernetes or multi-client environments who want Context running as a Dockerized HTTP server shared by several MCP clients on the network.
- Developers in non-JavaScript ecosystems (Django, Spring Boot, Rust) who need to turn reStructuredText or AsciiDoc docs folders into AI-queryable packages.
- Users with many globally installed packages who want per-project scoping via context serve --libs react [email protected].
What are this agent's strengths and limitations?
- Fully local: docs are stored as local SQLite databases, queries return in under 10ms, work offline, and queries never leave your machine.
- Plug and play: 100+ popular libraries have pre-built registry packages that agents download automatically on demand.
- Model- and client-agnostic: as a standard MCP server it works with Claude Code, Cursor, Codex, VS Code, Windsurf, Zed, Goose, and OpenCode.
- Custom data sources: context add supports private git repos, local directories, llms.txt sites, blog URLs, and multiple doc formats (Markdown/AsciiDoc/rst/HTML), with portable .db exports for team sharing.
- No ongoing cost: no subscriptions, rate limits, or usage caps, and no dependence on any cloud service's availability.
- Requires a Node.js runtime and a global npm install as a resident local process; the Docker HTTP deployment must be built from the monorepo root.
- Libraries not in the registry require manual context add builds, and quality depends on how the target repo organizes its docs (e.g., React's docs live in a separate repo — picking the wrong repo triggers a low-sections warning).
- The first use of a new library still requires network access to download its package; full offline use presupposes packages are pre-installed.
- Fetching subscriber-only content requires manually configuring context auth cookies/headers, stored in plaintext at ~/.context/auth. (0600 permissions).
- Advanced capabilities (e.g., --libs session scoping, multi-client HTTP transport) require explicit CLI options; the default stdio mode serves a single client.
How do you install or deploy this agent?
Install globally via npm: npm install -g @neuledge/context. No API keys required. Then connect Context to any MCP-compatible agent, e.g., Claude Code: claude mcp add context -- context serve; for Cursor, add {"mcpServers":{"context":{"command":"context","args":["serve"]}}} to .cursor/mcp.. Other documented integrations include Claude Desktop, OpenAI Codex, VS Code (GitHub Copilot), Windsurf, Zed, Goose, and OpenCode, each with exact config snippets in the README. For HTTP deployment, build from the repo root: docker build -t context:local -f packages/context/Dockerfile . && docker run --rm -p 8080:8080 context:local; the endpoint is http://localhost:8080/mcp.
How do you use this agent?
After installing and connecting, just ask your AI agent (e.g., "How do I create middleware in Next.js?") — the agent automatically searches the registry and downloads the docs package, no manual context install needed for registry packages. Common CLI operations: context browse npm/next to list available versions; context install npm/next 15.0.4 for a specific version; context add https://github.com/vercel/next.js/tree/v16.0.0 to build from a tag (supports --path, --name, --pkg-version, --save); context add https://svelte.dev to fetch a site's llms.txt; context add ./my-project --name my-lib --save ./packages/ to build and export a shareable .db; context list to show installed packages; context query nextjs 'middleware authentication' for direct lookup; context serve --http 3000 for multi-client HTTP mode; context serve --libs react [email protected] to lock a session to specific libraries. For subscriber-only content, first configure credentials with context auth add substack.com --cookies "...".