Self-hosted AI coding agents give teams control over where their code goes, what models they use, and what their infrastructure looks like. For teams with data residency requirements, security policies, or a preference for not paying cloud-agent pricing on top of model costs, self-hosted options cover most of what managed services offer.
This list covers the best options across both agent tools and the orchestration layer that manages them.
The most complete self-hosted autonomous coding agent. Docker-based, MIT licensed, supports Claude, GPT-4, and local models. Browser UI for monitoring. Strong SWE-bench performance.
Best for: Teams that want a fully self-hosted autonomous coding agent with a browser UI and broad LLM support.
Lightweight, no Docker required, pip installable. Self-hosted in the sense that it runs entirely on your machine against whichever model endpoint you configure, including local Ollama models.
Best for: Developers who want the lowest-overhead self-hosted coding agent with no container dependency.
Runs locally on the developer's machine. No cloud agent infrastructure — it calls the Anthropic API directly for model inference and executes everything locally. Frequently updated by Anthropic.
Best for: Teams on Anthropic's API that want a capable locally running agent with direct filesystem and git access.
Open-source VS Code extension. Runs entirely on the developer's machine against any configured model endpoint including local models.
Best for: Developers who want a self-hosted editor-integrated agent with explicit approval control.
OpenAI's open-source terminal agent. Sandboxed execution locally, Apache 2.0 licensed. Configurable to use any OpenAI-compatible endpoint including self-hosted models.
Best for: Teams on the OpenAI stack who want a self-hosted terminal agent with sandboxing.
Runs on your own hardware against configured model endpoints. Self-hostable and open-source, good for teams that want to benchmark or customize the agent loop.
Best for: Teams that want a research-grade self-hosted agent they can benchmark and extend.
Fleet is designed specifically for the self-hosted case. It installs as a single Go binary with no Docker and runs on your own infrastructure. It orchestrates Claude Code agents running on your own machines, using a SQLite database and a fabric event bus over a Unix socket. For teams with data residency requirements or a preference for owning their infrastructure, Fleet plus Claude Code provides the full autonomous software delivery stack without any code leaving your environment except for model API calls.
Pick OpenHands for the most complete self-hosted autonomous agent with Docker sandbox.
Pick Aider for the lightest-weight self-hosted option with no container.
Pick Claude Code for a capable locally running agent with Anthropic support.
Pick Fleet for the self-hosted orchestration layer that coordinates a team of agents with governance and cost controls.
Not necessarily. Most self-hosted coding agents still make API calls to a model provider (Anthropic, OpenAI, etc.) for inference. Truly air-gapped operation requires a local model via Ollama or similar — which most agents support but at lower capability than hosted frontier models.
Requirements vary by tool. Aider and Claude Code run on any machine with Python or Go installed. OpenHands requires Docker. Fleet requires a machine with tmux and the Go binary — no Docker or Kubernetes. For production multi-agent use, a persistent server (Linux, 4GB RAM minimum) is recommended for running multiple parallel agents.
One binary. Five minutes. See every agent, coordinate every handoff, and keep a full audit trail of what your fleet did.
