Configure a provider

Iteron resolves provider routing only from operator-controlled sources. A cloned repository cannot redirect your model traffic or credential to a new host.

For a built-in provider, start with the Setup and BYOK wizard. This page covers route selection and advanced custom endpoints.

Built-in providers

Provider id	Wire adapter	Credential environment variable
glm	OpenAI-compatible Chat	GLM_API_KEY
anthropic	Anthropic Messages	ANTHROPIC_API_KEY
openai	OpenAI Responses	OPENAI_API_KEY
deepseek	OpenAI-compatible Chat	DEEPSEEK_API_KEY
minimax	OpenAI-compatible Chat	MINIMAX_API_KEY
fireworks	OpenAI-compatible Chat	FIREWORKS_API_KEY

The built-in default is glm. For its exact standard Chat endpoint, Iteron ships a versioned static manifest whose documented default is glm-5.2. That manifest proves endpoint schema compatibility, not that a particular account is funded or entitled to every listed model. Without a credential, its leaves are visible but unavailable.

The GLM schema/capability snapshots and Anthropic effort beta metadata are dated world data. Iteron emits one preflight notice per selected provider showing the age of every static snapshot used by that run and whether an operator replacement changed its provider revision. See static provider metadata for the offline refresh path; no network response silently rewrites these claims.

Select a built-in provider

For one run:

iteron --provider anthropic --model PROVIDER_MODEL_ID

Or set trusted user defaults in ~/.iteron/config.json:

{
  "schema_version": 2,
  "provider": "anthropic",
  "model": "PROVIDER_MODEL_ID"
}

Credential values stay in environment variables. Do not add them to JSON.

Define a compatible provider instance

Operator-defined providers also live only in ~/.iteron/config.json:

{
  "schema_version": 2,
  "provider": "gateway",
  "providers": [
    {
      "id": "gateway",
      "display_name": "Internal compatible gateway",
      "adapter": "openai_chat",
      "error_profile": "custom",
      "api_root": "https://gateway.example.invalid/v1",
      "key_env": "GATEWAY_API_KEY",
      "enabled": true,
      "catalog": false,
      "models": ["declared-model-id"],
      "model_capabilities": {
        "declared-model-id": {
          "context_window_tokens": 1048576,
          "image_input": true
        }
      }
    }
  ]
}

model_capabilities is optional. Iteron has no portable way to discover a context window or image support. Fireworks' typed supportsImageInput catalog field and Iteron's static vendor snapshots are used when present; custom gateways can declare the exact model capability here. Without a window, compaction falls back to the absolute compaction_trigger_tokens instead of a share of the window, and the pre-flight admission check cannot run. Declaring it is you stating a number you read in your provider's documentation; it is not evidence of entitlement, and an official vendor snapshot for the same route still wins. See the configuration reference for the bounds.

The example domain is deliberately non-routable. Replace it with a trusted HTTPS API root. Exact loopback hosts may use HTTP; other HTTP roots, embedded credentials, queries, and fragments are rejected.

Supported adapters are anthropic_messages, openai_responses, and openai_chat. A declared model list is an operator manifest, not account discovery evidence.

Project configuration cannot reroute traffic

A repository may place a bare model id in .iteron/config.json, but it remains constrained to the independently selected provider. Repository values for provider, base_url, providers, MCP processes, hooks, effort, or a grant of code execution are ignored or rejected as appropriate.

See the configuration reference and provider matrix for the complete contract.