Use Claude Code and other coding assistants with Outlier

A coding assistant is two things: the program in your terminal that reads files and runs commands, and the model that decides what to do next. The program is free to point somewhere else. Point it at Outlier and the deciding happens on your own Mac: your code is not uploaded to anyone, there is no per-token bill, and it keeps working on a plane.

Turn on Outlier's API

In Outlier, open Settings → Advanced → Developer → OpenAI-compatible API, switch it on and copy the API key. The same switch and the same key serve both API shapes. The server listens only on your Mac (127.0.0.1), so nothing else on your network can reach it.

Claude Code

Claude Code talks Anthropic's API, so it needs a server that speaks it. Outlier does. Set these in the terminal you start Claude Code from:

export ANTHROPIC_BASE_URL=http://127.0.0.1:8766
export ANTHROPIC_AUTH_TOKEN=your-outlier-api-key
export ANTHROPIC_MODEL=outlier-compact
export ANTHROPIC_DEFAULT_SONNET_MODEL=outlier-compact
export ANTHROPIC_DEFAULT_HAIKU_MODEL=outlier-compact
export ANTHROPIC_DEFAULT_OPUS_MODEL=outlier-compact
export CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1
export CLAUDE_CODE_AUTO_COMPACT_WINDOW=126976
claude

The three DEFAULT_*_MODEL lines matter more than they look. Claude Code makes small background calls under other model names; without them those calls ask Outlier for a model you did not pick. AUTO_COMPACT_WINDOW tells Claude Code how much room it has: it assumes 200,000 tokens for a model it does not know, while Outlier Pro gives a model up to 131,072 when your Mac has the memory for it. If Claude Code ever stops on a context overflow, lower the number to 28672.

What we measured

Outlier 1.11.999 with outlier-compact (Qwen3.6-27B), Claude Code 2.1.317, on a Mac Studio with an M1 Ultra and 64 GB of memory, with every outbound connection blocked except Outlier's port:

Task	Result	Time
Reply with one word	Passed ("Ready")	86.9 s
Write hello.py that prints 17 + 25, run it, report the output	Passed (42)	185.3 s

Most of that time is reading, not writing. Claude Code sends its own long instructions and its tool list with every request, and a 27B model on this Mac reads a long prompt at roughly 110 to 130 tokens a second (measured in the engine's own log). The first answer waits for that. While it waits, Outlier sends a keep-alive every 15 seconds, so the connection does not time out.

Where Claude in the cloud is better

Honestly: anything hard. Large changes across many files, long debugging sessions and work that needs a frontier model's judgment go better with Claude running in Anthropic's cloud, and much faster. Our 54-prompt comparison has the numbers. Use the local setup for private code, for working offline, and for the long tail of small jobs you do not want to pay per token for.

Claude Code is a product of Anthropic. Outlier is independent and is not affiliated with or endorsed by Anthropic.

OpenCode, Goose, Qwen Code and other OpenAI-style tools

These speak the OpenAI API. Give them the base URL http://127.0.0.1:8766/v1, your Outlier API key, and the model outlier-compact. We ran all three against Outlier in our offline head-to-head this month with exactly these settings.

OpenCode, in ~/.config/opencode/opencode.json:

{
  "provider": {
    "local": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "Outlier",
      "options": { "baseURL": "http://127.0.0.1:8766/v1", "apiKey": "your-outlier-api-key" },
      "models": { "outlier-compact": { "name": "outlier-compact" } }
    }
  },
  "model": "local/outlier-compact"
}

Goose:

export GOOSE_PROVIDER=openai
export OPENAI_HOST=http://127.0.0.1:8766
export OPENAI_BASE_PATH=v1/chat/completions
export OPENAI_API_KEY=your-outlier-api-key
export GOOSE_MODEL=outlier-compact

Qwen Code:

export OPENAI_BASE_URL=http://127.0.0.1:8766/v1
export OPENAI_API_KEY=your-outlier-api-key
export OPENAI_MODEL=outlier-compact

Open WebUI takes the same three values in its "OpenAI-compatible" connection settings: we ran it against Outlier on 5 November, and the Open WebUI guide walks through it. Continue and Zed have the same kind of OpenAI-compatible provider setting, but we have not run either one against Outlier yet. Cursor is the exception: it has no local mode, as this page explains.

Limits worth knowing

• One request runs at a time; others wait their turn, up to 10 minutes by default.
• An answer can ask for up to 262,144 tokens; above 16,384 it is capped at what the context leaves.
• Context is up to 131,072 tokens on Pro, depending on how much memory your Mac has free.
• A single request can be up to 8 MB.

What stays on your Mac

The model runs on your Mac, and Outlier sends nothing out unless you turn on web search. The assistant is separate software with its own habits: Claude Code checks in with Anthropic for updates and telemetry unless you set CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1. Our test blocked every outbound connection except Outlier's port, and both tasks still passed.

Common questions