i built an open source coding agent harness that runs on hardware you control

Description and use

Autopilot lets developers run coding agents on their own hardware, keeping API keys secure and supporting many LLM providers.

Problem

Developers want to run coding agents on hardware they control rather than relying on cloud providers, while keeping API keys secure.

Outcome

Autopilot enables secure, local or self‑hosted coding agent execution with key isolation, supporting multiple LLM providers and sandboxed task execution.

Workflow

• Install the CLI with npm install -g directionallylocal dl
• Run dl in a GitHub repo to start an agent session
• Use dl run "task description" to execute a task in a sandboxed environment
• The agent returns a PR or output, and the sandbox is destroyed
• Close the terminal; the agent continues running and can be reattached with dl

Evidence

you bring your own key. (if you used kimiflare before, that was autopilot's old name.)

Source

subagents and background tasks run in hotcell sandboxes (docker, firecracker or apple virtualization microVMs). your real api key never goes inside them.

Source

1 caveat: the gateway only swaps credentials for HTTP APIs. it can't protect something like a postgres password.

Source