Claude Code and Codex on your own models

Run Claude Code or Codex on a model you've downloaded in Dyno. The model answers on your Mac instead of Anthropic's or OpenAI's servers. The tools themselves may still contact their makers for sign-in, updates or usage data, depending on how you've set them up; dyno launch turns off Claude Code's optional traffic.

dyno launch claude --model qwen3.8-27b
dyno launch codex --model qwen3.8-27b

Dyno's model server speaks each tool's own API, so there's no proxy to install and no config file to edit.

Before you start

1. A model. Download one in Discover. A 27B model or larger works best: we tested Qwen3.8-27B (MLX, 4-bit). An 8B model answers simple questions and uses the tools, but it gets lost on multi-step tasks.
2. The tool. Install it if you haven't:
   • Claude Code: npm install -g @anthropic-ai/claude-code
   • Codex: npm install -g @openai/codex
3. The dyno command. Dyno adds it when you open the app, so it works in a new Terminal window. If you have Homebrew, it goes in /opt/homebrew/bin with no password. Otherwise Dyno asks once to add /usr/local/bin/dyno, and macOS asks for your password; choose Not Now and you can do it later in Router → Install the dyno command…. If you move or update Dyno, the command follows it. Dyno never replaces another program called dyno. If you installed Dyno's Python package with pip, you already have the command.

The Router tab shows these commands for the model you pick, with a Copy button for each.

Run it

dyno models

This lists the models you've downloaded, with their size, context window and the port serving them, if any.

dyno launch claude --model qwen3.8-27b

--model takes the model's full name (lmstudio-community/Qwen3.8-27B-MLX-4bit) or any part of it that matches only one model. Dyno tells you if a name matches several.

dyno launch then:

1. Uses the model if it's already running in a Dyno server, or starts it on a free port between 8970 and 8979. Loading takes a moment.
2. Starts the tool pointed at it. Every model the tool would use, including Claude Code's background model and subagents, is the one you chose. The tool's context window is set to the model's own.
3. Stops the model it started when you quit the tool, or close the terminal. Add --keep to leave it running.

Anything after -- goes to the tool:

dyno launch claude --model qwen3.8-27b -- -p "The tests fail. Fix them."
dyno launch codex --model qwen3.8-27b -- exec "The tests fail. Fix them."

What we saw with Qwen3.8-27B on an M5 Max: given a small project with a failing test, Claude Code ran the test, found the bug, fixed it and reran the test in 74 seconds, including loading the model. Codex did the same in 52 seconds.

Web search

Both tools can search the web while they run on your model. Claude Code's WebSearch and Codex's web search normally run on Anthropic's or OpenAI's servers. With Dyno, Dyno runs them:

• The model gets a search tool. When it searches, Dyno runs the search, gives the model the results, and the tool shows the search as usual.
• Dyno searches with SearXNG in Docker on your Mac (free), or with a paid search API you choose in Settings → Web search: Brave, Tavily, Exa or Serper. A paid API needs a key; Dyno keeps it in your macOS Keychain. Test runs one search so you can check it works.
• Claude Code's allowed and blocked domains are applied to the results.

Search queries leave your Mac: they go to the search engines (SearXNG) or to the provider you chose. See Dyno's assistant for what that means.

Without dyno launch

Any Dyno model server (dyno serve, or a model started in Models) answers both APIs, on the same port as its OpenAI API:

To point Claude Code at a model running on port 8971 yourself:

ANTHROPIC_BASE_URL=http://127.0.0.1:8971 ANTHROPIC_AUTH_TOKEN=dyno claude --model Qwen3.8-27B-MLX-4bit

And Codex:

codex -c model_provider=dyno -c 'model_providers.dyno.name="Dyno"' \
  -c 'model_providers.dyno.base_url="http://127.0.0.1:8971/v1"' \
  -c 'model_providers.dyno.wire_api="responses"' -m Qwen3.8-27B-MLX-4bit

The model name can be anything the server doesn't list (Claude Code asks for claude-… names): the server then uses the model it's serving, and never downloads one.

Limits

• Your model, not Claude or GPT. The tools work, but a local model makes more mistakes with them, especially on long tasks.
• Not translated: Anthropic's and OpenAI's other hosted tools (code execution, file search), PDFs, and your MCP servers in Codex, which Codex only offers to OpenAI's models. Claude Code's MCP servers work: they're ordinary tools.
• The first reply is slow. Before it answers, the model reads the tool's instructions, about 16,000 to 24,000 tokens. That took 75 seconds on an M5 Max and over 5 minutes on a slower Mac. Later replies read only what's new, and so does a new session while the model is still loaded: about 3,000 tokens instead of 16,000, in any folder. So keep the model loaded between sessions: start it in Models, or add --keep to dyno launch. Dyno keeps the connection alive meanwhile, and dyno launch raises the tools' timeouts to an hour (API_TIMEOUT_MS for Claude Code). If you point Claude Code at Dyno yourself, set API_TIMEOUT_MS=3600000 too. A smaller model reads faster.
• Token counts are estimates (about four characters a token). Claude Code uses them to decide when to compact a conversation.
• Only this Mac. The APIs answer programs on this Mac and refuse web pages, so a site you visit can't use your model through them.
• A model started in Dyno before 0.6.9 doesn't have these APIs. dyno launch then starts another copy of the model and says so; restart the model in Models to avoid that.