trovex — Claude Code MCP server: one current doc instead of rereading the repo
Indexes your repo's markdown and hands the agent a single current document per question. Cuts tokens spent on doc lookups.
claude mcp add trovex -- uvx trovexWhat trovex does
Coding agents reread the repo each session to figure out which .md still holds, then answer from a guess. trovex builds one canonical index over your markdown and runs an MCP server: a question comes back as a path:line pointer to the one current doc, tagged canonical, stale or duplicate, plus just the section that answers it rather than the whole file.
Six MCP tools ship with it. Three only read (trovex, trovex_read, trovex_search); three write solely into trovex's own store (trovex_write, trovex_tag, trovex_delete). No shell runs, your source files stay untouched. Vectors live in SQLite and ONNX computes embeddings on the machine, so answering a query needs neither a network call nor an API key. On the authors' own repo the measured saving was about 60% at equal task success; your repo will differ, and trovex search prints what it saved.
Install comes from PyPI with uv tool install trovex, then trovex setup wires the skill, hooks and MCP server into Claude in one pass. No uv yet? It has its own one-line installer. The server starts with trovex serve and listens at http://localhost:8765/mcp, with the savings dashboard at http://localhost:8765/savings. Point a client at the endpoint before trovex serve is running and the connection simply fails.
What it helps with
- surface the one current document that answers an agent's question
- stop paying tokens to reread the repo every session
- tell a stale doc from the live one when the text looks alike
- serve the answering section instead of the full file
- pool what several agents and teammates learn into one store
Example prompts for the agent
- How do we roll back a deploy? Find the canonical doc with trovex.
- Index the markdown in this repo and show the tokens the search saved.
- Store the rollback runbook with trovex_write so the next agent reads it back.
How to install trovex
Platform CLI
claude mcp add trovex -- uvx trovexuvx runs the command named after the package; if the README names another one, use uvx --from trovex <command>
Platform CLI
codex mcp add trovex -- uvx trovexuvx runs the command named after the package; if the README names another one, use uvx --from trovex <command>
Config file
{
"mcpServers": {
"trovex": {
"command": "uvx",
"args": [
"trovex"
]
}
}
}uvx runs the command named after the package; if the README names another one, use uvx --from trovex <command>
Config file
{
"servers": {
"trovex": {
"type": "stdio",
"command": "uvx",
"args": [
"trovex"
]
}
}
}uvx runs the command named after the package; if the README names another one, use uvx --from trovex <command>
Platform CLI
gemini mcp add trovex uvx trovexuvx runs the command named after the package; if the README names another one, use uvx --from trovex <command>
Step by step, with commands for each platform: How to add an MCP server to Claude Code, Codex, Cursor, VS Code, Gemini CLI and more
Questions about trovex
How do I install trovex in Claude Code?
Run uv tool install trovex, then trovex setup, which registers the skill, hooks and MCP server in one step. Restart Claude afterwards so it picks up the skill and hooks.
What does trovex do?
It indexes your repo's markdown and returns one current document per agent question as a path:line pointer with a freshness marker. Only the answering section is served, not the whole file.
Does trovex need an API key or a cloud service?
No. Vectors sit in SQLite, ONNX handles embeddings and search runs locally. A key is only needed for trovex bench --eval, where an LLM judges the answers.
