Skip to content

A model router on your own machine

Easy requests go to your cheap models. Hard ones go to your strong ones.

Your app sends its requests to delegation-kit instead of to one provider. A decider you pick rates how hard each request is, and the cheapest of your models that can handle it answers. Every answer says which model wrote it, and why.

Get started → How you use it ↓

Free and open source under MIT. Runs on your Mac, with your own keys.

You send Decider’s rating Answered by
Translate this paragraph into French Rated 1 of 4 · no minimum Answered by DeepSeek Flash, or GPT-6 Luna (max), or DeepSeek Flash, or GPT-6 Luna (max)
Summarise this contract Rated 2 of 4 · min 38.4 Answered by GPT-6.1 Terra, or GPT-6.1 Terra, or GPT-6.1 Terra, or Claude Opus 5.5
Find the race condition in this scheduler Rated 3 of 4 · code · min 50.4 Answered by Claude Opus 5.5, or Claude Sonnet 5.5 (max), or GPT-6.1 Sol (max), or Claude Opus 5.5
The answer column cycles through four example pools with explicitly measured efforts; in each, the cheapest model that reaches the request's minimum answers. The setup below uses the first: GPT-6.1 Terra (through Codex at high effort), DeepSeek Flash (an API key at max effort) and Claude Opus 5.5 (through Claude Code at max effort).

How you use it

Three steps. In your app, only the address and the model name change.

  1. Install it

    On macOS, with Node.js, jq, git and python3. It sets up the models it knows (Claude and GPT through Claude Code and Codex, Meta Muse through Muse Code or its API, DeepSeek, Qwen, GLM, Kimi, Grok, MiniMax). Keys go in ~/.config/delegation-kit/.env, a file only you can read. Any other model with an API compatible with OpenAI's can be added, a local one too.

    Terminal

    npx delegation-kit
delegation-config show  # your models

  2. Choose your models and the decider

    Name the models the router may pick from, the decider that rates each request, and the model that answers when the decider can't. Here: GPT-6.1 Terra, DeepSeek Flash and Claude Opus 5.5.

    ~/.config/delegation-kit/config.json, next to “profiles”

    "router": {
  "decider": { "kind": "llm", "profile": "deepseek-flash" },
  "pool": ["terra-builder", "deepseek-flash", "opus-reviewer"],
  "default": "deepseek-flash"
}

  3. Start it and point your app at it

    Anything that uses OpenAI's Chat Completions API works: set the base URL to the router and the model to "auto". The answer comes back as usual; its X-Router-Model header names the model. An agent loop that sends X-Router-Conversation stays on the model its first request got. Claude Code and Codex speak other formats: they can't send their requests here, but the router can pick them to answer.

    Terminal

    delegation-router serve

    Your app (Python, OpenAI SDK)

    from openai import OpenAI

client = OpenAI(base_url="http://127.0.0.1:4141/v1", api_key="unused")
reply = client.chat.completions.create(
    model="auto",
    messages=[{"role": "user", "content": "Rename foo to bar in utils.py"}],
)

Get started, step by step →

How it picks a model

  1. Measured efforts get a score

    Artificial Analysis publishes a versioned set of model tests. Unmeasured efforts stay unknown. Within 8 points of the day's best: frontier. Within 20: mid. The rest: small.

  2. Your decider rates the request

    It reads an excerpt, never your keys or your list of models, and answers: how hard is this, 0 to 4, half points allowed? Code, math, rare knowledge?

  3. The cheapest one strong enough answers

    The rating sets a minimum score: none for 0–1, 38.4 for 2, 50.4 for 3–4, in between for half points. Math or rare knowledge: 38.4 at least. The cheapest of your models above it answers; none above? Your strongest. No rating? Your default.

Intelligence Index v5.0 · 2026-10-28

Artificial Analysis scores, split into three tiers Artificial Analysis Intelligence Index v5.0 scores from 0 to 60 on 2026-10-28, one dot per model, using the example effort for ringed models and the best measured setting for other models. The best score of the day is Claude Opus 5.5 at 58.4. Within 8 points of it (50.4 and up): frontier, the strongest tier. Within 20 points (38.4 and up): mid. Below: small. The three models in the examples: Claude Opus 5.5 is frontier, DeepSeek Flash (32.5) is small; GPT-6.1 Terra (42.3) is mid, 3.9 above the mid line. small mid frontier 50.4 38.4 0 10 20 30 40 50 60 8 points 20 points claude-opus-5-5 · 58.4 · frontier gpt-6-astra · 56.9 · frontier gpt-6.1-sol · 54.5 · frontier claude-sonnet-5-5 · 54.2 · frontier gemini-4-argon · 50.4 · frontier claude-fable-5-1 · 49.7 · mid muse-wonder-full-1.0 · 49.7 · mid kimi-k3.1-pro · 46.0 · mid muse-wonder-1.0 · 43.3 · mid gpt-6.1-terra · 42.3 · mid muse-spark-1.3 · 36.4 · small claude-haiku-5-5 · 36.2 · small grok-4.7 · 35.0 · small qwen3.8-max · 33.6 · small deepseek-flash · 32.5 · small glm-5.3 · 31.7 · small gemini-3.8-flash · 30.5 · small minimax-m3.1-flash · 28.9 · small gpt-6-luna · 27.8 · small mistral-large-4 · 25.3 · small deepseek-v4-pro · 23.1 · small Claude Opus 5.5 58.4 best of the day DeepSeek Flash 32.5 GPT-6.1 Terra 42.3, 3.9 above mid Artificial Analysis scores, split into three tiers Artificial Analysis Intelligence Index v5.0 scores from 0 to 60 on 2026-10-28, one dot per model, using the example effort for ringed models and the best measured setting for other models. The best score of the day is Claude Opus 5.5 at 58.4. Within 8 points of it (50.4 and up): frontier, the strongest tier. Within 20 points (38.4 and up): mid. Below: small. The three models in the examples: Claude Opus 5.5 is frontier, DeepSeek Flash (32.5) is small; GPT-6.1 Terra (42.3) is mid, 3.9 above the mid line. frontier · 50.4 mid · 38.4 small 0 10 20 30 40 50 60 claude-opus-5-5 · 58.4 · frontier gpt-6-astra · 56.9 · frontier gpt-6.1-sol · 54.5 · frontier claude-sonnet-5-5 · 54.2 · frontier gemini-4-argon · 50.4 · frontier claude-fable-5-1 · 49.7 · mid muse-wonder-full-1.0 · 49.7 · mid kimi-k3.1-pro · 46.0 · mid muse-wonder-1.0 · 43.3 · mid gpt-6.1-terra · 42.3 · mid muse-spark-1.3 · 36.4 · small claude-haiku-5-5 · 36.2 · small grok-4.7 · 35.0 · small qwen3.8-max · 33.6 · small deepseek-flash · 32.5 · small glm-5.3 · 31.7 · small gemini-3.8-flash · 30.5 · small minimax-m3.1-flash · 28.9 · small gpt-6-luna · 27.8 · small mistral-large-4 · 25.3 · small deepseek-v4-pro · 23.1 · small Claude Opus 5.5 58.4 best of the day DeepSeek Flash 32.5 GPT-6.1 Terra 42.3 mid: 3.9 above the line

21 models plotted; glm-5.3-flash, mistral-medium-3.5, mistral-small-2603, muse-spark-1.3-contributor, qwen3.8-flash have no score yet.

Ringed dots use the efforts in this example; other dots use each model's best measured setting. Unmeasured results stay off the chart. The amber line is a request's minimum score: the cheapest ringed model at or above it answers. Refresh the scores with your own Artificial Analysis key, or set any model's tier yourself.

With your three models

You send Decider’s rating Answered by Why
Translate this paragraph into French Rated 1 of 4 · no minimum Answered by DeepSeek Flash small ~$0.0013 a request Why No minimum: the cheapest of your models, DeepSeek Flash.
Summarise this contract Rated 2 of 4 · min 38.4 Answered by GPT-6.1 Terra mid ~$0.0054 a request Why Minimum 38.4: Terra (42.3) clears it and is the cheapest that does.
Find the race condition in this scheduler Rated 3 of 4 · code · min 50.4 Answered by Claude Opus 5.5 frontier ~$0.0217 a request Why Minimum 50.4: only Opus 5.5 (58.4) reaches it.
Rename foo to bar in utils.py Rated 1 of 4 · code · no minimum Answered by DeepSeek Flash small ~$0.0013 a request Why No minimum: the cheapest code-capable model in this pool, DeepSeek Flash.
What is 17% of 2,340? Rated 1 of 4 · math · min 38.4 Answered by GPT-6.1 Terra mid ~$0.0054 a request Why Easy, but math raises the minimum to 38.4: Terra (42.3) still clears it.
Any request, when the decider times out Rated no rating Answered by DeepSeek Flash small ~$0.0013 a request Why No rating: your default model answers, and the response says so.
Costs are the router's estimates at API prices, used to compare models; a model you reach through a subscription is billed by your plan.

Get started → Router reference →