Skip to content

[providers]

Model providers: name, address, key source, routing role, limits and declared models.

Key Type Default Description
providers[].console.label string label is the name the provider list shows; empty shows the provider’s name.
providers[].console.models array of string models are the model ids discovery accepted, kept for display and to publish the provider at startup before discovery answers.
providers[].console.created_at integer created_at is when the console added the provider, in Unix seconds.
providers[].console.created_by string created_by is the integration id of the agent token that added the provider; empty means the owner.
providers[].console.revision integer revision counts the console’s saves of this provider; a save names the revision it read and is refused when another save came first.
providers[].context_length integer Provider-wide limits for APIs that omit model metadata, e.g. Sakana (0 = unspecified). Both model catalogs label these values with source=config. Live provider/catalog facts win; per-model bootstrap declarations fill gaps before these defaults. Values survive each discovery refresh.
providers[].max_output_tokens integer max_output_tokens is the provider-wide output limit for models whose metadata omits it (0 = unspecified). Provider and catalog values win.
providers[].stream_open_keepalive_ms integer Nil inherits the server default; explicit zero disables keepalives.
providers[].proxy_url string proxy_url overrides the global egress proxy for this provider only.
providers[].name string name identifies the provider in model ids (provider::model), reports and logs. Known names (openai, anthropic, google, openrouter, groq, deepseek, together, fireworks, deepinfra, tokenbroker, meta, typesafe) bring a default base_url and key variable; any other name is a custom provider.
providers[].api_key_env string api_key_env names the environment variable that holds the provider’s API key. Empty uses the provider’s built-in variable (for example OPENAI_API_KEY); a local server needs none. The key itself never goes in this file.
providers[].api_key_ref string api_key_ref is a secret reference (op://vault/item/field) resolved by [secrets]. Mutually exclusive with api_key_env.
providers[].base_url string base_url is the provider’s API address. Optional for known providers. For OpenAI-compatible servers both the server root and the /v1 form work.
providers[].self_hosted bool Self-hosted server: longer stream-open and attempt timeouts. Inferred true for loopback, RFC 1918, 100.64.0.0/10 and .local hosts.
providers[].no_wallet bool Explicit zero-wallet attestation for managed spend authority. Local or loopback networking alone is never treated as free.
providers[].local_cost_microunits integer Fixed planned cost of one request, in millionths of a local cost unit, for quality-first cost planning. 0. .1000000000000; only on a provider that is self_hosted, routing_role = “local” or private_trusted.
providers[].private_trusted bool Alternatively, private_trusted = true opts a provider whose transport is NOT private (e.g. a tailscale/CGNAT overlay address) into private-routing eligibility – and also silences the warning by making intent explicit:
providers[].stream_protocol string Wire protocol override; normally inferred from the provider name: openai_compat | anthropic_messages | ollama_native | tei_buffered | typesafe_systemone
providers[].openai_quirk string OpenAI-compatible dialect quirks: openrouter | together
providers[].request_policy_profile string Optional: profile name in request_policies.json for this provider’s chat completions body rules. When omitted, Prism picks a built-in default from the provider name at config load (e.g. google -> google_chat_strict_v1).
providers[].routing_role string Optional routing topology override. If omitted, Prism auto-detects direct, aggregator, or local from provider name/base_url. direct | aggregator | local
providers[].cost_premium number Optional aggregator premium. Ignored for direct/local roles.
[[providers[].bootstrap_models]] array of tables bootstrap_models declare models for providers that publish no model list, and fill metadata that discovery omits. Discovery replaces them when it works; a declared model it still lists keeps the declared fields discovery did not report.
providers[].bootstrap_models[].input_price_per_1m number input_price_per_1m and output_price_per_1m are the declared prices in USD per million input and output tokens.
providers[].bootstrap_models[].output_price_per_1m number input_price_per_1m and output_price_per_1m are the declared prices in USD per million input and output tokens.
providers[].bootstrap_models[].model_id string model_id is the provider’s own id of the declared model.
providers[].bootstrap_models[].task_types array of string task_types lists what the model serves: chat, responses, embedding, rerank or systemone. A model enters an automatic pool only for its declared tasks.
providers[].bootstrap_models[].supports_tools bool supports_tools, supports_json, supports_streaming and supports_responses_api declare the model’s tool calls, JSON mode, streaming and Responses API support. true is an operator statement (source config), which the provider’s own model list and what Prism observes in traffic outrank; false or unset declares nothing, and the flag stays unknown until a stronger source states it.
providers[].bootstrap_models[].supports_json bool supports_tools, supports_json, supports_streaming and supports_responses_api declare the model’s tool calls, JSON mode, streaming and Responses API support. true is an operator statement (source config), which the provider’s own model list and what Prism observes in traffic outrank; false or unset declares nothing, and the flag stays unknown until a stronger source states it.
providers[].bootstrap_models[].supports_streaming bool supports_tools, supports_json, supports_streaming and supports_responses_api declare the model’s tool calls, JSON mode, streaming and Responses API support. true is an operator statement (source config), which the provider’s own model list and what Prism observes in traffic outrank; false or unset declares nothing, and the flag stays unknown until a stronger source states it.
providers[].bootstrap_models[].supports_responses_api bool supports_tools, supports_json, supports_streaming and supports_responses_api declare the model’s tool calls, JSON mode, streaming and Responses API support. true is an operator statement (source config), which the provider’s own model list and what Prism observes in traffic outrank; false or unset declares nothing, and the flag stays unknown until a stronger source states it.
providers[].bootstrap_models[].context_window integer context_window is the model’s context size in tokens.
providers[].bootstrap_models[].max_output_tokens integer max_output_tokens is the model’s output limit in tokens.
providers[].bootstrap_models[].enforced_max_output_tokens integer enforced_max_output_tokens is an output limit Prism applies to every request to this model; 0 applies none.
providers[].bootstrap_models[].embedding_dimensions integer embedding_dimensions is the vector size of an embedding model, published in the model catalogs.
[[providers[].pricing]] array of tables pricing is an operator-declared per-model rate table for sellers that do not publish prices on /v1/models. It does not seed catalog rows; use bootstrap_models when the model must exist before discovery.
providers[].pricing[].model string model is the provider’s model id the prices apply to.
providers[].pricing[].input_price_per_1m number input_price_per_1m and output_price_per_1m are the prices in USD per million input and output tokens.
providers[].pricing[].output_price_per_1m number input_price_per_1m and output_price_per_1m are the prices in USD per million input and output tokens.