Docs / Autopilot

Autopilot

Autopilot is a terminal coding agent that runs on your own provider key: OpenRouter, Anthropic, OpenAI, Gemini, Bedrock, Ollama and more.

What Autopilot is

Autopilot is a terminal coding agent. It sends every model call to the provider you choose, using your own key. Nothing is proxied or marked up. OpenRouter is the default provider, and you can switch to Anthropic, OpenAI, Gemini, Bedrock, Vertex AI, DeepSeek, a local Ollama or LM Studio model, or any OpenAI-compatible endpoint.

  • npm package: autopilot-ai
  • Command: autopilot
  • License: MIT
  • Formerly: kimiflare. Old settings, sessions and KIMIFLARE_* environment variables still work.
  • Source: github.com/directionally-local/autopilot

If you want agents that run tasks in the background and return pull requests, install Directionally Local instead (npm install -g directionallylocal, then dl; see /docs/install/). It runs Autopilot inside Hotcell sandboxes for you.

Parts

Part What it does Docs
Interactive terminal UI The default when you run autopilot. Chat, approve tool calls, switch models. Quick start
One-shot print mode autopilot -p "<prompt>" streams one answer to stdout and exits. Commands
Providers and models 28 provider ids, each with its own key and model catalog. Models and providers
Permission modes Optional plan, edit and auto modes (off by default, so every turn runs in auto), plus config rules for headless runs. Modes and permissions
Subagents Read-only Autopilot instances that research in local Hotcell sandboxes. Subagents
Hooks Shell commands that fire on lifecycle events and can veto actions. Hooks
Logs, traces and evals JSONL logs, optional OpenTelemetry export, local session traces, an eval runner. Logs, traces and evals
SDK and RPC createAgentSession from autopilot-ai/sdk, or JSONL over stdio. SDK and RPC
Headless server autopilot serve for unattended runs, jobs and timers. Running unattended

Built-in capabilities

  • File, shell, search and web tools, GitHub access and a headless browser fetch.
  • Image understanding: include image paths (PNG, JPG, WebP, GIF, BMP, up to 5 MB) in a prompt.
  • Language Server Protocol (LSP) for hover, go-to-definition, references and diagnostics.
  • Model Context Protocol (MCP) servers for external tools.
  • Local structured memory (SQLite and embeddings) through the remember, recall and forget tools.
  • Per-turn cost in the status bar, and /cost totals by session, day, month and all time.
  • Message queuing: messages you send while the agent works are triaged as interrupt, steer, queue or aside.

Next steps

  1. Install Autopilot.
  2. Follow the quick start.
  3. Pick a provider and model.