Your coding agents talk past each other; this TUI listens

What it does

cass is a Rust TUI and CLI that slurps up chat history from the swarm of AI coding assistants now living on your machine—Cursor, Claude Code, Aider, Codex, and more than a dozen others—and dumps them into a single SQLite-backed timeline. You search across all of them at once, either by raw text or optional local semantic embeddings, without leaving your machine. It also speaks JSON when called from scripts or other agents, so it doubles as an archive for automated workflows.

The interesting bit

The project treats your chat logs like a database that needs an ops team: lexical indexes are published with atomic swaps and versioned backups, corrupt assets are quarantined instead of deleted, and a built-in doctor command runs read-only diagnostics before it ever touches a file. That level of paranoia makes sense when the “database” is a pile of vendor-specific JSONL files scattered across your home directory.

Key highlights

• Hybrid search defaults to fast lexical matching, with local semantic embeddings (all-MiniLM, Snowflake Arctic, Nomic) joining in the background only when explicitly installed and ready.
• SQLite is the sole source of truth; every derived index can be rebuilt from it, and the tool will recover or finalize interrupted publishes on its own.
• Full TUI with a three-pane layout (filter, results, preview) plus a --robot mode that emits structured JSON for agent-to-agent consumption.
• Supports roughly 20 providers including Cursor, Claude Code, Gemini CLI, GitHub Copilot Chat, Aider, ChatGPT, and several lesser-known agents.
• Air-gapped friendly: embedders are opt-in manual downloads, with an offline path for importing model files manually.

Caveats

• Marked alpha; the JSON API contracts are tested but still evolving, and schema changes require deliberate golden-file regeneration.
• Semantic search is opportunistic—without explicitly installed local embedders, queries silently fall back to lexical-only.
• The optional MCP “Agent Mail” coordination requires a separate Python server and careful working-directory discipline to avoid “project not found” errors.

Verdict

Grab it if you juggle multiple AI agents and need a single, offline search layer across their histories. Skip it if you only talk to one coding assistant, or if you expect a polished, stable API while the project is still alpha.