Finn is a coding agent for the terminal, like Claude Code or OpenCode, written in Agentica. You give it a task in a folder; it reads files, edits them, runs commands and answers.

make install          # in a clone of agc-agent
bin/finn login        # set up a provider with its key
bin/finn              # chat on the full screen in this folder

Quickstart · Providers · CLI and TUI · Approvals and security

What Finn does #

  • One server, several clients. The agent runs as a server on localhost. The full-screen client (finn) and the line-mode client (finn chat) talk to it over HTTP and Server-Sent Events. A session goes on when a client closes.
  • Many model providers. Anthropic, OpenAI, OpenRouter, Fireworks, Groq, DeepSeek, a local llama.cpp server, and any OpenAI-compatible API of your config file.
  • Tools. Read, write and edit files, search them, and run bash commands. Code mode lets the model call the tools from a Python script in a sandbox.
  • Sessions you can go back in. Resume a session, go back to an earlier message, fork or clone it. Long histories fold into a summary.
  • Skills and MCP. SKILL.md folders and MCP servers add what the model can do.
  • Extensions. An Agentica program can serve Finn with its own tools, wrappers and services.
  • Approvals when you want them. Off by default; deny rules count either way. See Approvals and security.

Before you start #

Tools run without asking by default: the model can run any command and change any file that you can, except what a deny rule refuses. Run Finn in a sandbox or a container, or turn the approvals on with finn serve --approvals.