Show HN: Context Harness – Local first context engine for AI tools

[__parallaxis]

Context Harness is a single Rust binary that gives AI tools like Cursor and Claude project-specific memory. It ingests docs, code, Jira tickets, Slack threads, and anything else into a local SQLite database, indexes them with FTS5 and optional vector embeddings, and exposes hybrid search via CLI and an MCP-compatible HTTP server.

I built this because I kept hitting the same problem: AI tools are powerful but have no memory of my complex multi-repo project. They can't search our internal docs, past incidents, or architecture decisions. Cloud RAG services exist, but they're complex, expensive, and your data leaves your machine. I wanted something I could point at my sources and just run `ctx sync all`.

Quick start:

    # Install (pre-built binaries available for macOS/Linux/Windows)
    cargo install --git https://github.com/parallax-labs/context-harness.git

    # Create config and initialize
    ctx init

    # Sync your data sources (filesystem, Git, S3, or Lua scripts)
    ctx sync all

    # Search from CLI
    ctx search "how does the auth service validate tokens"

    # Or start the MCP server for Cursor/Claude Desktop
    ctx serve mcp

What it does differently from other RAG tools:

- *Truly local*: SQLite + single binary. No Docker, no Postgres, no cloud. Local embeddings (bundled or pure-Rust) so semantic and hybrid search work with zero API keys. Back up your entire knowledge base with `cp ctx.sqlite ctx.sqlite.bak`.

- *Hybrid search*: FTS5 keyword scoring + cosine vector similarity with configurable blending. Works without embeddings too (keyword-only mode); with local embeddings you get full hybrid search offline.

- *Lua extensibility*: Write custom connectors, tools, and agents in Lua without recompiling anything. The Lua VM has HTTP, JSON, crypto, and filesystem APIs built in.

- *Extension registry*: `ctx registry init` installs a Git-backed community registry with 10 connectors (Jira, Confluence, Slack, Notion, RSS, Stack Overflow, Linear, etc.), 4 MCP tools, and 2 agent personas.

- *MCP protocol*: Cursor, Claude Desktop, Continue.dev, and any MCP-compatible client can connect and search your knowledge base directly.

Embeddings: you can run *fully offline* — the default build uses local embeddings (fastembed with bundled ONNX on most platforms, or a pure-Rust tract path on Linux musl and Intel Mac). No API key required. Optional: Ollama (local LLM stack) or OpenAI if you prefer. Keyword-only mode needs zero deps. There's no built-in auth layer; it's designed for local or trusted network use.

Stack: Rust, SQLite (WAL mode), FTS5, mlua (Lua 5.4), axum, MCP Streamable HTTP. MIT licensed.

GitHub: https://github.com/parallax-labs/context-harness

Docs: https://parallax-labs.github.io/context-harness/

Community Registry: https://github.com/parallax-labs/ctx-registry

If you find it useful, a star on GitHub is always appreciated.

Would love feedback on the search quality tuning (hybrid alpha, candidate counts) and the Lua extension model.

[fidorka]

Hey Parker, this is really cool! Thanks for sharing. Have you tried using the entire CLI? It might also be a tool which you could compose into your workflow to have better memory of what the agents themselves did in the repo.

Btw, I built something similar to solve the context problem for most of my laptop-based activity.

It's slightly more heavyweight (electron app ingesting screenshots) - that being said I took many similar design decisions (local embeddings, sqlite with vector search and FTS hybrid, MCP extension to claude). Feel free to check it out:

https://github.com/deusXmachina-dev/memorylane

[__parallaxis]

Hey, thanks for the kind words — and for the CLI/agent-memory idea. I haven’t fully leaned into “index what the agents did in the repo” yet; that’s a great direction. The MCP server gets most of the love so far, but the CLI is the same pipeline (sync → chunk → embed → search), so wiring it in so agents can search their own history is totally doable. I’ll keep that in mind.

I have mostly used the entire API surface so far. Check out the usage in this github action script: https://github.com/parallax-labs/context-harness/blob/main/s...

This is used to build the search index on the website (below)

This tool is made for not only local, but embedded in a ci context

MemoryLane looks really cool — same problem, different surface. Local embeddings + SQLite + hybrid FTS/vector + MCP into Claude is basically the same stack; the screenshot-ingestion and Electron UX are a neat take for “everything I’ve seen on this machine.” I’ll definitely poke around the repo. If you want to see how we’re using custom agents on top of that pipeline, a couple of blog posts go into it: Chat with your blog

- https://parallax-labs.github.io/context-harness/blog/enginee... (persona grounded in your own writing, inline + Lua agents) this is in the same vein. Allowing agents to write into the vector store with an MCP tool is on the road map.

- https://parallax-labs.github.io/context-harness/blog/enginee... Unified context for your engineering team (shared KB, code-reviewer and on-call agents).

[fidorka]

Nice, thanks! Checking it out now.