Local AI coding agents

How each agent connects to a local model

Only agents whose own documentation describes a local setup are listed; each link goes to that page. Checked Oct 6, 2026.

Why run a coding agent locally

Privacy. With a local model, your code and prompts are not sent to a model provider. (Check the agent's own telemetry settings too.) That matters for client code, regulated work, or a company policy that forbids sending source to third parties.

Cost. There is no per-token bill. You pay once, in hardware, and in electricity.

Offline work. Once the model is downloaded, open-source agents can keep working without a connection. Extensions of commercial products may still need you to be signed in.

What to expect

Local models are smaller than the frontier models behind commercial agents, and coding agents are demanding: they read many files, keep a long context and call tools in loops. In practice that means:

• Hardware decides what you can run. A laptop with 16 GB of memory runs small models that can handle simple edits and explanations. Models that can drive a multi-step agent loop reliably need much more memory, typically a workstation GPU or a recent Mac with a large amount of unified memory.
• Context length matters as much as model size. Local servers often start with a short default context window; agents need a long one to see enough of the codebase. Check your server's context setting before judging a model.
• Tool calling must be supported by the model and the server, or the agent cannot run commands or edit files.

A common setup is hybrid: a local model for routine or sensitive work, a hosted model for the hard tasks. Agents that let you switch models mid-session make that easy.

Agents that do not run locally

Some popular agents work only with their vendor's hosted models, or reach a model on your machine only through a public tunnel. For example, Claude Code documents that it does not route to non-Claude models, and Gemini CLI connects only to Google's backends. Their pages on this site list which models they support.