# trovex

> trovex — one canonical doc for your coding agents, ~60% fewer tokens.
>
> trovex is the canonical doc store for AI coding agents. It indexes a repo's
> markdown and serves an agent the single current doc that answers a query — a
> path:line pointer with a freshness marker — instead of letting it reread the
> repo to guess which file is canonical. About 60% fewer tokens per doc lookup,
> measured at equal task-success (methodology at trovex.dev/measure). Open source, local-first (SQLite + ONNX), runs on your
> machine. Ships as an MCP server and a CLI.

## What it does
- Serves one canonical answer per query (with freshness: canonical / stale / duplicate), not a pile of candidate chunks to sift.
- Section-level reads: hands back the paragraph that answers, not the whole file, so a short answer costs short-answer tokens.
- Shared write path: one agent records what it learns once; every other agent and your teammates read it back — single source of truth by construction.
- Local-first: indexing and embeddings run on your machine in SQLite + ONNX. No cloud, no API keys, your code never leaves your machine.

## Who it's for
- Solo developers and small teams (1–20) running AI coding agents on real, doc-heavy repos.
- Anyone whose agents reread the same .md files every session and burn tokens guessing which doc is current.

## Try it
- [Free self-audit](https://trovex.dev/audit): a 5-minute interactive check that estimates how much your coding agent wastes rereading docs, and what trovex would save — no install required.

## Comparisons
- [trovex vs CLAUDE.md / AGENTS.md](https://trovex.dev/vs/claude-md/): a static instructions file vs one canonical doc routed per query.
- [trovex vs repomix / files-to-prompt](https://trovex.dev/vs/repomix/): packing the whole repo into a prompt vs retrieving the one answer.
- [trovex vs context-hub (CTX)](https://trovex.dev/vs/context-hub/): a ranked list of candidate chunks vs one canonical doc with a freshness marker.
- [All comparisons](https://trovex.dev/vs/)

## Answers
- [How do I manage context for my coding agents?](https://trovex.dev/answers/managing-agent-context/): control the read — one canonical doc per query, current, section-level, shared across the fleet.
- [How do I stop my coding agent rereading the same files every session?](https://trovex.dev/answers/stop-agent-rereading-docs/): serve one canonical answer per query instead of raw files to sift.
- [How do I reduce the token cost of giving a coding agent context?](https://trovex.dev/answers/reduce-agent-token-costs/): cut the repeated retrieval, not the one-time prompt.
- [What is canonical context for coding agents?](https://trovex.dev/answers/canonical-context-for-agents/): the single current, authoritative doc that answers a query.
- [Why do AI coding agents pick stale or outdated docs?](https://trovex.dev/answers/why-agents-pick-stale-docs/): similarity search ranks by relevance, not freshness; a freshness marker, not a bigger model, is the fix.
- [Does a bigger context window make rereading docs cheaper?](https://trovex.dev/answers/bigger-context-window-rereading/): no — it raises the ceiling, not the efficiency.
- [How do I keep multiple agents working from the same docs?](https://trovex.dev/answers/shared-source-of-truth-multiple-agents/): one shared read-and-write point, no drift.
- [Is there an MCP server that serves a repo's docs to coding agents?](https://trovex.dev/answers/mcp-server-for-repo-docs/): yes — trovex is an MCP server that returns the one canonical doc per query, not the whole repo.
- [What's the single source of truth for a multi-agent repo?](https://trovex.dev/answers/source-of-truth-multi-agent-repos/): one canonical, freshness-marked doc every agent reads and writes back, instead of each agent guessing from raw files.
- [What's the difference between agent memory and retrieval (RAG)?](https://trovex.dev/answers/agent-memory-vs-rag/): memory is an agent's recall of its own past; RAG is on-demand fetch from a corpus; neither gives a fleet one current source of truth — that's a separate shared canonical-context layer.
- [What's the missing layer in agent memory?](https://trovex.dev/answers/missing-layer-agent-memory/): the 7 memory types are per-agent recall; the missing layer is one current, authoritative source of truth the whole fleet reads and writes back to.
- [What is trovex?](https://trovex.dev/answers/what-is-trovex/): the open-source canonical doc store and context layer for AI coding agents, an MCP server and CLI that cuts about 60% of tokens per doc lookup, by tsukumo.
- [How much do agents spend rereading the same files?](https://trovex.dev/answers/agent-cost-rereading-files/): the hidden cost is re-reading to re-derive what the agent already knew; serve one canonical answer instead.
- [What does context cost your coding agents?](https://trovex.dev/answers/token-cost-of-agent-context/): the per-lookup token math, and how serving one canonical doc cuts it.
- [What is local-first context for coding agents?](https://trovex.dev/answers/local-first-context-for-agents/): index and serve your repo's canonical docs on your own machine over MCP, no cloud or API keys.
- [Why don't context files (CLAUDE.md / AGENTS.md) scale?](https://trovex.dev/answers/why-context-files-dont-scale/): one static file goes stale and can't route a query to the right doc; serve the current one per query instead.
- [Can AI agents correct their own mistakes?](https://trovex.dev/answers/can-ai-agents-correct-their-own-mistakes/): not reliably alone; a 2026 study (arXiv:2606.05976) found correction rose 23 to 93 points by relabeling the same claim as external, so the fix is a second role in the loop.
- [Do AI agents report false success?](https://trovex.dev/answers/do-ai-agents-report-false-success/): yes; a 2026 study (arXiv:2606.09863) measured false success up to 75.8% of self-assessing runs and an LLM judge caught it near coin-flip odds, so gate on a verified state change, not self-report.
- [Which multi-agent architecture should I use?](https://trovex.dev/answers/which-multi-agent-architecture-to-use/): the most accurate is usually the wrong production pick; a 2026 financial-doc benchmark (arXiv:2603.22651) found a hybrid kept 89% of the top-accuracy gain at 1.15 times the cost vs 2.3 times, so pick on the cost-accuracy frontier.
- [How do I debug a multi-agent failure?](https://trovex.dev/answers/how-to-debug-multi-agent-failures/): failure attribution needs the full execution trace, not output-only logs; a 2026 benchmark (arXiv:2604.22708) found full traces up to 76% more accurate, so capture inputs and keep runs reproducible.
- [How do I run AI agent fleets in production?](https://trovex.dev/answers/how-to-run-ai-agent-fleets-in-production/): an operations discipline, not a framework choice; four operating problems (reliability, observability, context, cost), each with an open tsukumo tool, and tsukumo consults on the model.
- [Is AI agent memory any good?](https://trovex.dev/answers/is-ai-agent-memory-good/): not in real group settings; a 2026 benchmark (arXiv:2605.14498) found the best memory system scored 46.0% avg and BM25 keyword search matched or beat most, so memory is a retrieval problem, not storage.
- [All answers](https://trovex.dev/answers/)

## Setup (per MCP client)
- [trovex for Claude Code](https://trovex.dev/for/claude-code/): `claude mcp add --transport http trovex http://localhost:8765/mcp`.
- [trovex for Cursor](https://trovex.dev/for/cursor/): add it to `.cursor/mcp.json`, enable in Settings → MCP.
- [trovex for Windsurf](https://trovex.dev/for/windsurf/): `serverUrl` in `~/.codeium/windsurf/mcp_config.json`.
- [trovex for Cline](https://trovex.dev/for/cline/): a remote streamable-HTTP server in `cline_mcp_settings.json`.
- [All clients](https://trovex.dev/for/)

## Access
- Status: public beta. The repository is public; install it and run it on your own repo.
- Install: `uv tool install trovex` (a GitHub star helps other devs find it).
- Community: questions or comparing setups, join the Slack at https://tsukumo.ch/slack.

## Links
- [GitHub repository](https://github.com/TsukumoHQ/trovex): public; install it and a star helps.
- [Community Slack](https://tsukumo.ch/slack): questions and setup help.
- [Website](https://trovex.dev): the landing page.
- [Token-savings calculator](https://trovex.dev/savings): a free client-side tool to estimate how many tokens your coding agents waste rereading the repo, and what one canonical answer saves (~60% fewer per doc lookup). No install, nothing leaves the page.
- [5-minute agent-docs self-audit](https://trovex.dev/audit): a free self-audit to spot whether your coding agents are burning tokens rereading docs, and what to fix.
- [How we measure the ~60%](https://trovex.dev/measure): the methodology and the receipt behind the only number trovex claims — measured 69% at equal task-success across 26 pre-registered queries (range 41-81%, LLM-judged), with the limits stated and a harness to reproduce it on your own repo. We headline a conservative ~60%.
- [Repository](https://github.com/TsukumoHQ/trovex): source (public; core AGPL-3.0, CLIs MIT).
- [llms-full.txt](https://trovex.dev/llms-full.txt): the full extractable content in one file.
- Markdown twins: every /answers, /vs, and /glossary page has a chrome-free markdown version at its URL + `index.md` (e.g. https://trovex.dev/answers/canonical-context-for-agents/index.md) — token-cheap content built for AI ingestion.

## Facts
- Category: context / memory infrastructure for AI coding agents.
- Interface: Model Context Protocol (MCP) server (local, streamable-HTTP at http://localhost:8765/mcp) + CLI.
- Works with: Claude Code, Cursor, Windsurf, Cline, Zed, and any MCP client.
- License: core AGPL-3.0, CLIs MIT.
- Access: public beta — the repository is public at github.com/TsukumoHQ/trovex.
- Install: `uv tool install trovex`, then `trovex index /path/to/repo` and `trovex serve`.