Updated
Can you give coding agents canonical context without a cloud?
Yes. trovex runs entirely on your machine, vectors in a local SQLite file, embeddings from a local ONNX model, no API keys and no network calls, so your code and docs never leave. You still get the same saving, about 60% fewer tokens per lookup, because it comes from serving one canonical doc instead of rereading candidates, not from where the index lives.
Why local-first matters for agent context
Most ways to give an agent more context send your code somewhere: a hosted RAG service, a cloud embedding API, a third-party index. For a side project that's fine. For a real codebase, it means your proprietary source and internal docs leave your control and land on someone else's servers, which is a non-starter in regulated or security-conscious teams. Local-first removes the question entirely: the context layer is a process on your machine, and nothing is uploaded to be indexed.
How it works here
trovex indexes your repo's markdown into a local SQLite database and embeds it with a bundled ONNX model that runs on your CPU. Retrieval is a local query that returns the one canonical doc for a question, a path:line pointer with a freshness marker. There is no account, no key, and no outbound request, so the same setup works on a laptop, a locked-down work machine, or an air-gapped build server.
What you give up (the honest part)
Not much, for this job. You run the index yourself, a one-time command that takes about a minute and re-runs as docs change, rather than a service doing it for you. There's no hosted dashboard for a whole org out of the box. In exchange you keep your code on your machine, pay no per-token embedding bill, and get the same retrieval saving. For most teams that's the right trade.
How trovex does it
trovex is the open-source, local-first MCP server described above: index a repo, serve agents one current answer per query, write back what they learn so the next agent reads it. It reports the tokens it saved, so the number is measured on your repo. See the per-scenario math in what context costs your agents.
FAQ
Does my code leave my machine?
No. trovex indexes and serves your docs entirely on your machine: vectors in a local SQLite file, embeddings from a local ONNX model, no network calls and no API keys. Your code and docs never go to a third-party server.
Do I need an OpenAI or other API key?
No. Embeddings run locally via ONNX, so there is no embedding API to pay for or send text to. You point trovex at a repo and it indexes offline.
Does it work offline or in an air-gapped environment?
Yes. Because indexing and retrieval are local with no network dependency, trovex works offline and suits air-gapped or regulated environments where code can't leave the building.
Is local-first slower or less accurate than a cloud service?
No meaningful trade-off for this job. Retrieval is a local SQLite lookup, and the saving is the same, about 60% fewer tokens per lookup, because it comes from serving one canonical doc instead of rereading candidates, not from where the index lives.
Keep your code on your machine.
Point trovex at your docs and serve your agents one current answer per query, no cloud, no keys.
Open source. No cloud, no API keys. Your docs never leave your machine. In a regulated environment and rolling agents out across a team? tsukumo (the team behind trovex) helps you operate them safely.