Python · SQLite · sqlite-vec · DuckDB · fastembed · MCP

convo-index — searchable history for AI coding sessions

A local-first index and browser for Claude Code, Codex, Opencode, Antigravity and pi conversations, with semantic search and an MCP server.

View project ↗

Context

I work across several AI coding agents, and each keeps its history in its own format and folder. Finding "that session where we fixed the Kubernetes crash" meant grepping JSONL files by hand.

What I built

One index across Claude Code, Codex CLI, Opencode, Antigravity and the pi coding agent, with a two-pane browser UI and a CLI:

  • Search: full-text plus semantic search, using in-process embeddings (fastembed + sqlite-vec), so nothing leaves the machine.
  • Summaries and tags: generated through whichever LLM CLI is already installed. Hosted models stay off until you allow them, and you can exclude directories from indexing.
  • Related work: sessions linked by ticket, PR and branch, plus token usage with API-equivalent cost estimates.
  • MCP server: agents can search their own past conversations.
  • Continue from the browser: run another turn in the agent that owns the session, with hooks, plugins and project settings locked down, and read-only by default.

Outcome

A single curl | sh installs it, keeps the UI running from login and refreshes the index daily, using a LaunchAgent on macOS, a systemd user service on Linux and Task Scheduler on Windows. I use it every day to find and resume past work.