Providers

tm routes turns through a provider fabric with twenty-one backends. Run tm provider detect (add --json for machine-readable output) to see which are configured in your environment, tm provider list for the role-to-candidate routing from the project’s providers.toml, and tm provider test <slug> for a real (potentially billed) smoke test.

tm init scaffolds providers.toml (provider role routing) and a separate harness.toml (harness behavior). A generated providers.toml follows the environment: if DEVPASS_* is set, coder.fast routes to DevPass; a hand-edited file is never touched. tm provider reset regenerates it (keeping the old one as providers.toml.bak). Local model names are never guessed; select one with /model <provider>/<model>.

A providers.toml role candidate may name any slug below:

[coder.fast]
candidates = [
    { provider = "openai", model = "gpt-4o-mini", max_concurrency = 20 },
]

Anthropic🔗

• Slug: anthropic
• Env vars: ANTHROPIC_API_KEY (required)
• Base URL: https://api.anthropic.com
• Free tier: no
• Capabilities: completion, streaming, tool use

OpenAI🔗

• Slug: openai
• Env vars: OPENAI_API_KEY (required); OPENAI_BASE_URL, OPENAI_ORGANIZATION, OPENAI_PROJECT (optional)
• Base URL: https://api.openai.com/v1
• Free tier: no (new-account trial credit only, not a standing free tier)
• Capabilities: completion, embedding, streaming, tool use
• Known gap: reasoning-model families (o1/o3/gpt-5-style) require max_completion_tokens instead of max_tokens; the shared compat core always sends max_tokens, so those specific models are not correctly reachable through this backend yet (see the module’s own doc comment).

OpenRouter🔗

• Slug: openrouter
• Env vars: OPENROUTER_API_KEY (required); OPENROUTER_BASE_URL, OPENROUTER_HTTP_REFERER, OPENROUTER_APP_TITLE (optional)
• Base URL: https://openrouter.ai/api/v1
• Free tier: yes — several :free-suffixed routed models
• Capabilities: completion, streaming, tool use

GitHub Models🔗

• Slug: github-models
• Env vars: GITHUB_TOKEN (required — a GitHub PAT or Actions token scoped for GitHub Models, or GITHUB_MODELS_TOKEN); GITHUB_MODELS_BASE_URL (optional)
• Base URL: https://models.github.ai/inference
• Free tier: yes — free within GitHub-imposed per-model rate limits
• Capabilities: completion, streaming, tool use

Groq🔗

• Slug: groq
• Env vars: GROQ_API_KEY (required); GROQ_BASE_URL (optional)
• Base URL: https://api.groq.com/openai/v1
• Free tier: yes — free developer tier with rate limits
• Capabilities: completion, streaming, tool use

Cerebras🔗

• Slug: cerebras
• Env vars: CEREBRAS_API_KEY (required); CEREBRAS_BASE_URL (optional)
• Base URL: https://api.cerebras.ai/v1
• Free tier: yes — free tier with daily request/token caps
• Capabilities: completion, streaming, tool use

DeepSeek🔗

• Slug: deepseek
• Env vars: DEEPSEEK_API_KEY (required); DEEPSEEK_BASE_URL (optional)
• Base URL: https://api.deepseek.com
• Free tier: no (pay-as-you-go only)
• Capabilities: completion, streaming, tool use

Mistral🔗

• Slug: mistral
• Env vars: MISTRAL_API_KEY (required); MISTRAL_BASE_URL (optional)
• Base URL: https://api.mistral.ai/v1
• Free tier: yes — a rate-limited free “La Plateforme” experiment tier
• Capabilities: completion, embedding, streaming, tool use

xAI🔗

• Slug: xai
• Env vars: XAI_API_KEY (required); XAI_BASE_URL (optional)
• Base URL: https://api.x.ai/v1
• Free tier: no (new-account trial credit only)
• Capabilities: completion, streaming, tool use

Together AI🔗

• Slug: together
• Env vars: TOGETHER_API_KEY (required); TOGETHER_BASE_URL (optional)
• Base URL: https://api.together.xyz/v1
• Free tier: yes — a handful of always-free serverless models
• Capabilities: completion, embedding, streaming, tool use

Fireworks AI🔗

• Slug: fireworks
• Env vars: FIREWORKS_API_KEY (required); FIREWORKS_BASE_URL (optional)
• Base URL: https://api.fireworks.ai/inference/v1
• Free tier: yes — new-account free credit, not a standing free tier
• Capabilities: completion, embedding, streaming, tool use

Hugging Face🔗

• Slug: huggingface
• Env vars: HF_TOKEN (required); HF_BASE_URL (optional)
• Base URL: https://router.huggingface.co/v1
• Free tier: yes — rate-limited free inference for many hosted models
• Capabilities: completion, streaming (no tool use through this router)

Google Gemini🔗

• Slug: gemini
• Env vars: GEMINI_API_KEY (required, falls back to GOOGLE_API_KEY if unset); GOOGLE_API_KEY (optional fallback name); GEMINI_BASE_URL (optional)
• Base URL: https://generativelanguage.googleapis.com/v1beta
• Free tier: yes — a free tier with per-model rate limits
• Capabilities: completion, embedding, streaming, tool use

Ollama (local)🔗

• Slug: ollama
• Env vars: OLLAMA_HOST (optional, all env vars optional — see caveat below)
• Base URL: http://localhost:11434
• Free tier: yes — runs entirely locally, no API key
• Capabilities: completion, embedding, streaming, tool use
• No guessed default model: select a pulled model explicitly, for example /model ollama/qwen2.5-coder.
• Caveat: since every one of this backend’s env vars is optional, ProviderInfo::is_configured() always reports it as “configured”, regardless of whether an Ollama server is actually reachable at the configured host. Autodetection here means “constructible”, not “reachable”; confirming reachability needs an actual probe (tm provider test), not Registry::autodetect().

LM Studio (local)🔗

• Slug: lm-studio
• Env vars: LM_STUDIO_HOST (optional)
• Base URL: http://localhost:1234
• Free tier: yes — runs entirely locally, no API key
• Capabilities: completion, embedding, streaming, tool use
• No guessed default model: select a model exposed by LM Studio explicitly.
• Same “configured means constructible, not reachable” caveat as Ollama above.

llama.cpp (local)🔗

• Slug: llama-cpp
• Env vars: LLAMA_CPP_HOST (optional); LLAMA_CPP_API_KEY (optional, only needed if llama-server was started with --api-key)
• Base URL: http://localhost:8080
• Free tier: yes — runs entirely locally, no API key required by default
• Capabilities: completion, embedding, streaming, tool use
• No guessed default model: select the model served by llama-server explicitly.
• Same “configured means constructible, not reachable” caveat as Ollama above.

Azure OpenAI🔗

• Slug: azure-openai
• Env vars: AZURE_OPENAI_API_KEY (required); AZURE_OPENAI_ENDPOINT (required, e.g. https://<resource>.openai.azure.com, no path); AZURE_OPENAI_DEPLOYMENT (required, the deployment name, not the underlying model id); AZURE_OPENAI_API_VERSION (optional, default 2024-10-21)
• Free tier: no (Azure subscription required; some free credit programs exist but no standing free tier)
• Capabilities: completion, embedding, streaming, tool use

AWS Bedrock🔗

• Slug: bedrock
• Env vars: AWS_ACCESS_KEY_ID (required); AWS_SECRET_ACCESS_KEY (required); AWS_SESSION_TOKEN (optional, for temporary credentials); AWS_REGION (required, e.g. us-east-1); BEDROCK_MODEL_ID (required, Bedrock’s own model id string)
• Free tier: no (AWS subscription required)
• Capabilities: completion, streaming (no tool use, no embedding through this module)

Google Vertex AI🔗

• Slug: vertex
• Env vars: VERTEX_ACCESS_TOKEN (required — a pre-minted OAuth2 bearer token; service-account token minting is not implemented by this module); VERTEX_PROJECT (required, GCP project id); VERTEX_LOCATION (required, e.g. us-central1)
• Free tier: no (GCP subscription required; GCP’s general free-credit programs are not a Vertex-specific standing free tier)
• Capabilities: completion, embedding, tool use — not streaming (this module’s complete always sends a non-streaming request; see the module’s own doc comment for why)

Cloudflare Workers AI🔗

• Slug: cloudflare
• Env vars: CF_API_TOKEN (required, scoped for Workers AI); CF_ACCOUNT_ID (required, part of the URL path)
• Base URL: https://api.cloudflare.com/client/v4/accounts/<CF_ACCOUNT_ID>/ai/v1 (built from CF_ACCOUNT_ID, not overridable — there is no CF_BASE_URL)
• Free tier: yes — a daily free-usage allotment per Cloudflare account
• Capabilities: completion, embedding, streaming, tool use

DevPass (generic OpenAI-compatible)🔗

• Slug: devpass over CompatProvider::from_env_prefix)
• Env vars: DEVPASS_API_KEY (required); DEVPASS_BASE_URL (required, no trailing slash); DEVPASS_MODEL (required — the default model; a request routed to another devpass candidate, or a /model devpass/<name> choice, is served by that model instead)
• Free tier: depends entirely on the operator-controlled gateway this points at
• Capabilities: completion, embedding, streaming, tool use
• CompatProvider::from_env_prefix itself (not registered under its own slug) is the generic half of this backend: any other bespoke OpenAI-compatible endpoint can reuse it with a new env var prefix without writing a new module.

DevPass as the default provider🔗

When DEVPASS_API_KEY, DEVPASS_BASE_URL and DEVPASS_MODEL are all set to non-empty values, DevPass becomes the default provider for coder.fast — the one role tm’s interactive session (bare tm, tm -p <prompt>) and its scheduler dispatcher (tm run/tm sched run) actually drive an agent-loop turn as. Concretely:

• RoleTable::default_table’s coder.fast primary candidate becomes devpass/$DEVPASS_MODEL instead of anthropic/claude-sonnet-5; the existing Anthropic fallback candidate is left in place, untouched.
• crates/tm-cli/src/agent.rs’s build_fabric registers a DevPassProvider instead of requiring ANTHROPIC_API_KEY — this is the point of the feature: a real end-to-end tm session can run against a cheap/free OpenAI-compatible backend without touching Anthropic quota or needing an Anthropic account at all. AnthropicProvider::from_env is still attempted best-effort in this mode (registered if it happens to succeed, silently skipped if not), so any other role a ticket names still gets Anthropic service if a key also happens to be present.
• A partial set of the three env vars (e.g. only DEVPASS_API_KEY) does not activate this preference at all — it falls straight through to the unchanged, Anthropic-only default. There is no half-activated state.
• This does not touch any other role: vision.frontier/planner.frontier/architect.frontier (tm genesis’s bootstrap stages, resolved through a separate, Anthropic-only provider-selection path — see crates/tm-cli/src/project.rs’s resolve_genesis_provider) and embedder (background indexing) are all unaffected, on purpose: genesis’s own resolution hardcodes AnthropicProvider regardless of what a candidate’s provider slug names, so pointing a frontier role’s default at devpass there would silently misconstruct a provider rather than actually use DevPass; embedding needs a model that actually serves embeddings, which a DevPass-testing model is not guaranteed to.
• Absent all three env vars, behavior is byte-for-byte unchanged from before this feature existed.

See docs/decisions/D-005-devpass-default-provider.md for the reasoning and what this does not cover.

Free-tier summary🔗

Free-tier claims above are as of this document’s writing and are not verified against each vendor’s live terms by this crate; treat them as a starting point for choosing a role’s fallback candidates, not a billing guarantee.