Updated
What is trovex?
trovex is an open-source canonical doc store and context layer for AI coding agents: an MCP server and CLI that hands an agent one current answer per question instead of making it reread the repo to work out which doc is right. That cuts about 60% of the tokens per doc lookup. It is built and run by tsukumo, the studio that ships it as part of an open suite.
One canonical answer, not a pile of candidates
Point trovex at a folder of markdown and it builds a small index. When an agent asks a question, trovex returns a pointer to the single doc that answers it, plus that doc, instead of dumping the top few files for the agent to read and reconcile. The agent stops paying tokens to figure out which of three near-duplicate docs is the current one. It runs locally as an MCP server and a CLI, with a local embedder and no API key required to start.
Why that saves tokens
The waste in agent context is not the answer, it is the search. An agent rereading the top candidate docs to decide which is current burns tokens on files it then throws away. Resolving the question to one canonical doc skips that. Measured on our own repo at equal task-success, it cut a median of 69% of the tokens per markdown lookup across 26 queries, range 41 to 81%, judged by an LLM with no gold answers. We publish about 60% as the conservative floor and ship the benchmark as a command so you get your own number. Run uvx trovex bench on your repo.
Where trovex fits tsukumo
trovex is one of the open tools tsukumo ships, alongside wrai.th and yoru. They are how the studio proves it runs what it sells: real software in production, with the numbers disclosed. trovex is the top of the funnel, not the business. If your team is spending too much to run agents and you want help fixing the operating model rather than a tool to install, that is the consulting. If that is you, here is what getting help running agent fleets in production looks like.
FAQ
What is trovex in one line?
An open-source canonical doc store and context layer for AI coding agents, delivered as an MCP server and CLI that cuts about 60% of the tokens per doc lookup. Built by tsukumo.
How does trovex save tokens?
It resolves a question to one canonical doc instead of making the agent reread the top candidate files to work out which is current. Measured on our own repo at equal task-success, that cut a median of 69% of tokens per markdown lookup across 26 queries, range 41 to 81%. We publish about 60% as the conservative floor. Run uvx trovex bench for your own number.
Is trovex an MCP server or a CLI?
Both. The same tool exposes an MCP server so coding agents can query it directly, and a CLI so you can index a folder, search it, and run the savings benchmark from the terminal. It runs locally with a local embedder and needs no API key to start.
Is trovex free and open source?
Yes. trovex is open source and free to run. It is built and maintained by tsukumo, the Swiss studio and AI consultancy that ships it as part of an open suite.
How is trovex different from putting docs in AGENTS.md or CLAUDE.md?
A context file is a static dump every agent reads in full, current or not. trovex indexes your docs and returns the one that answers the question asked, so the agent reads what is relevant instead of everything. The two compose: keep the short standing instructions in a context file, let trovex serve the long tail on demand.
Who makes trovex?
tsukumo, a Swiss dev studio and AI consultancy that runs AI agents in production. trovex is part of the open suite the studio ships to prove it runs what it sells.
See it on your own repo.
Index a folder of markdown and serve your agents one current answer per query, in about a minute.
uv tool install trovex
Open source. No cloud, no API keys. Your docs never leave your machine.