# Quecto

> Two small, open source tools for a language model you host yourself, by Jesús Gulfo (https://jgulfo.com). Both are MIT licensed.

## Quecto Harness

A local coding agent that talks to a model server you host: a llama-server, a llama.cpp router, or anything that speaks the OpenAI chat API. One Go binary with the web UI compiled in. The transcript, the tools and the state stay on your machine.

- Every write and every command is asked about until you change the permission mode. Allow always on a shell command adds its prefix to the project's allowlist.
- Plan mode reads, searches and asks, then submits a plan to approve. Execute mode does the work.
- Any number of named agents, each with its own address, protocol (llama.cpp, OpenAI chat or OpenAI responses), model and settings. An agent can hand part of a task to another agent.
- A changes panel shows every file the session wrote as a diff. A context meter shows how full the window is, and the session compacts itself when it runs short.
- Binds 127.0.0.1:7331 and runs as a launchd or systemd service if you want it to.

Page: https://quecto.jgulfo.com/harness/
Source: https://github.com/jesdga95/quecto-harness

## Quecto Proxy

A small proxy for exposing a self-hosted OpenAI-compatible model server to the internet. One 11MB static binary and one SQLite file.

- API keys with expiry, per-key limits on requests and tokens per minute, and lifetime token quotas.
- Usage logs are metadata only: key, model, status, token counts and timings. Prompts and responses are never stored.
- A dashboard on its own port for keys, usage, logs and settings. The API port serves the API and nothing else.
- Streamed replies are still counted. Responses carry OpenAI's X-RateLimit headers. Roughly 50 microseconds added per request.
- Runs with Docker Compose in front of llama.cpp, vLLM or Ollama, behind a reverse proxy or tunnel that terminates TLS.

Page: https://quecto.jgulfo.com/proxy/
Source: https://github.com/jesdga95/quecto-proxy