trovex · open source

One canonical answer, not a pile to reread.

Your coding agent rereads several docs every session to work out which one is current, then answers from a guess. trovex serves it the one canonical doc for the question instead. About 60% fewer tokens per lookup, one source of truth, always current. Runs on your machine over MCP, no cloud, no account to start.

Get started

Star on GitHub

Open source, MIT · ~60% fewer tokens per lookup (measured) · local-first, MCP

Works with

Claude Code · Cursor · Windsurf · any MCP client

What it does.

One answer, not a pile

Your agent asks a question; trovex returns the single canonical doc that answers it, not ten candidates to open and rank. One lookup, one source of truth.

~60% fewer tokens per lookup

Serving the canonical doc instead of the reread pile drops a typical lookup from roughly 720 to roughly 280 tokens. Measured on a real repo, reproducible with the bench command.

Always current, never stale

trovex tracks which doc is canonical and flags the stale ones, so your agent reads what is true now instead of last quarter's runbook.

Local-first, your machine

Indexes and serves your docs locally over MCP. Your docs never leave your box; no cloud, no account to start.

One command. Then point your agents at it.

Install the CLI, index your docs, and trovex serves the canonical answer over MCP. No account, no keys to start. In a hurry, try it with no install: uvx trovex search "...".

uvx trovex   # or: pip install trovex

GitHub ↗trovex.dev ↗

Straight answers.

How is this different from CLAUDE.md or AGENTS.md?: Those pin one static blob into context; it goes stale and cannot route a specific question to the specific doc and section that answers it. trovex indexes your docs and serves the one current answer per query.
Do my docs leave my machine?: No. trovex runs locally and serves over MCP; your docs stay on your box. No cloud, no telemetry.
Where does the ~60% come from?: A real measurement on our own repo: the canonical lookup against the reread-the-pile baseline. Run the bench yourself and reproduce it at trovex.dev/measure.

Follow the build

When we ship or learn something real about cutting the tokens agents burn on context. No filler, no fixed schedule.

Running a fleet of agents across a team?

trovex is what we run our own agents on. If your team wants one source of truth for a fleet of agents, that's the work we do.

Talk to us

One canonical answer, not a pile to reread.

Open source, MIT · ~60% fewer tokens per lookup (measured) · local-first, MCP

What it does.

One answer, not a pile

Your agent asks a question; trovex returns the single canonical doc that answers it, not ten candidates to open and rank. One lookup, one source of truth.

~60% fewer tokens per lookup

Serving the canonical doc instead of the reread pile drops a typical lookup from roughly 720 to roughly 280 tokens. Measured on a real repo, reproducible with the bench command.

Always current, never stale

trovex tracks which doc is canonical and flags the stale ones, so your agent reads what is true now instead of last quarter's runbook.

Local-first, your machine

Indexes and serves your docs locally over MCP. Your docs never leave your box; no cloud, no account to start.

Straight answers.

How is this different from CLAUDE.md or AGENTS.md?

Those pin one static blob into context; it goes stale and cannot route a specific question to the specific doc and section that answers it. trovex indexes your docs and serves the one current answer per query.

Do my docs leave my machine?

No. trovex runs locally and serves over MCP; your docs stay on your box. No cloud, no telemetry.

Where does the ~60% come from?

A real measurement on our own repo: the canonical lookup against the reread-the-pile baseline. Run the bench yourself and reproduce it at trovex.dev/measure.

Follow the build

When we ship or learn something real about cutting the tokens agents burn on context. No filler, no fixed schedule.