which-llm — model selection for agents

For agents that need repeatable routing

Pick the right model.
Keep the reasons.

which-llm is a deterministic model-selection API for agents. Submit constraints, receive a stable ranking, pay only when you use it, and keep structured explainability for retries, audits, and downstream automation.

Explore capabilities View pricing Read agent metadata Open docs

Wallet signing stays local. The API verifies public payment proof only. The intended default is per-request approval on the host, while lower-friction execution depends on host configuration.

Typical paid request flow

POST /decision/optimize
402 Payment Required
pay exact USDC on supported chain
retry with X-Payment-* headers
receive deterministic ranking + reasons

Supported networks

Base, Ethereum, Arbitrum, Optimism, Avalanche

Credit loop

Report results to https://api.which-llm.com/decision/outcome for a discount token.

Core flow

Optimize then report outcome

Payment

HTTP 402 + exact USDC

Execution

Deterministic, no LLM call

Wallet model

Local signing only

What it is

A ranking service, not a model host

which-llm helps agents choose a model under hard constraints like maximum cost and minimum quality. It returns stable rankings, structured explainability, and fallback options without making an LLM call on your behalf.

Why agents use it

Consistent decisions under retries, caching, and audit

The same inputs produce the same ordering. That makes it easier to cache results, compare runs, review behavior changes, and automate routing without opaque heuristics.

Essential endpoints

Start with three endpoints

Use https://api.which-llm.com/decision/optimize for ranking, https://api.which-llm.com/decision/outcome for outcome credits, and https://api.which-llm.com/capabilities for machine-readable discovery.

/decision/optimize

Choose a model

Submit your goal and constraints to get a recommendation, expected tradeoffs, and a fallback plan.

/decision/outcome

Report execution results

Send actual cost, latency, quality, and success to earn a credit token for the next paid call.

/capabilities

Discover behavior

Fetch machine-readable service capabilities for agent tooling, generated docs, or static validation.

Payment and wallet guardrails

Paid endpoints return 402 Payment Required and accept exact USDC payments on Base, Ethereum, Arbitrum, Optimism, Avalanche. Use a separate low-balance wallet. Prefer an encrypted keystore over a raw private key. The API never needs your signing secret and only verifies the transaction hash, payer address, and related public proof.

Address verification

Verify payment addresses through multiple retrieval paths

Use more than one published path before automating payments. These are multiple publication or retrieval channels, but they may still share operational control, so treat them as corroborating signals rather than fully separate trust anchors.

/.well-known/agent.json

Agent metadata path that can include payment details.

HTTP 402 accepts[].pay_to

Request-specific address surfaced during the paid flow.

ENS: which-llm.eth

Additional retrieval path outside the main JSON endpoints.