Updated 22 June 2026

trovex for a fleet of agents

Run several coding agents on one repo — Claude Code, Cursor, your own — and each one re-derives the same context from scratch. Point the whole fleet at one trovex store and what one agent learns, the rest read back. About 60% fewer tokens, and no re-derivation.

trovex is open source and in public beta. The five steps below run in about a minute — no account, no keys to start.

Quickstart: one store, the whole fleet

1. Install trovex.
   uv tool install git+https://github.com/TsukumoHQ/trovexcopy
2. Index your repo — the shared store.

   One index every agent reads from. Embeds your markdown locally (SQLite + ONNX), about a minute.

   trovex index /path/to/repocopy
3. Point every agent at it.

   Run one server; each agent connects over MCP at http://localhost:8765/mcp. Wire each client with its guide — Claude Code, Cursor, Windsurf.

   trovex servecopy
4. See the shared savings.

   One agent figures something out and writes a record once. Every other agent — and your teammates — read it back instead of working it out again. The per-query savings (~60% fewer tokens) compound across the whole fleet. That compounding is the point.

5. Run it for a team.

   Local and keyless by default. For a team, add an API key so one shared trovex server backs every developer's agents — same store, same canonical answers. See the README for the team setup.

Why a fleet wants it

The cost of running many agents is re-derivation: each session, each agent, each teammate re-reads the repo and works out the same answer again. trovex keeps the answer canonical and serves the one current doc — down to the section, with a freshness marker — and lets any agent write a record the rest read back. So the fleet pays for an answer once, not once per agent. See one source of truth for a fleet of agents for the full pattern.