Weasel Providers

Providers — what weasel can talk to

chimera weasel reuses Chimera’s standard provider stack (chimera/providers/factory.py), so any provider that mink, otter, or ferret can drive, weasel can drive. The difference is in defaults and the resolution chain: weasel’s resolver (chimera/weasel/providers.py) prefers the broadest reach — it will fall back to a local Ollama daemon if no hosted key is set, so the zero-config path works on a fresh laptop.

This page is the weasel-specific layer. For the deep, line-numbered tour of every adapter, see docs/mink/providers.md. Same code, deeper notes.

Resolution order

build_provider(args) walks this chain on every weasel invocation, first match wins:

1. Explicit args.model (CLI --model <id>, SDK model=).
2. $WEASEL_MODEL environment variable.
3. $ANTHROPIC_API_KEY set → defaults to claude-sonnet-4-6.
4. $OPENAI_API_KEY set → defaults to gpt-4o.
5. $OPENROUTER_API_KEY set → defaults to anthropic/claude-sonnet-4.
6. Local Ollama daemon reachable on $OLLAMA_HOST (default http://localhost:11434) → first installed tag.
7. Friendly error pointing at the env vars above.

Explicit beats env beats default. So chimera weasel --model gpt-4o -p "..." works even when $ANTHROPIC_API_KEY is set.

Anthropic (default for hosted)

The first-class hosted target. claude-sonnet-4-6 is the default when $ANTHROPIC_API_KEY is set: it streams, tool-calls cleanly, supports extended thinking, and has prompt caching for long sessions.

• Setup:

  uv sync --extra anthropic
  export ANTHROPIC_API_KEY=sk-ant-...

• Use:

  chimera weasel -p "review this PR"
  chimera weasel --model claude-opus-4 -p "long-form refactor"

• Wired: streaming, tool calls, async, extended thinking, prompt caching, vision. See chimera/providers/anthropic.py.
• Anthropic-compatible endpoints: the same provider also accepts ANTHROPIC_BASE_URL + ANTHROPIC_AUTH_TOKEN, so weasel can route through GLM-4.6, Moonshot, or any third-party gateway that speaks the Messages API:

  export ANTHROPIC_BASE_URL=https://api.z.ai/v1/anthropic
  export ANTHROPIC_AUTH_TOKEN=...
  chimera weasel --model glm-4.6 -p "..."

OpenAI

Use when you have an OpenAI key and want gpt-4o, o3, or any GPT-5-class model weasel recognizes.

• Setup:

  uv sync --extra openai
  export OPENAI_API_KEY=sk-...

• Use:

  chimera weasel --model gpt-4o -p "draft a release note"
  chimera weasel --model o3-mini -p "prove this invariant"

• Wired: native streaming, tool calls, async (AsyncOpenAI), reasoning-token tracking for o-series, prompt-cache hit accounting, vision (gpt-4o), JSON mode.

OpenRouter

OpenRouter is one of weasel’s first-class targets because the vendor/name model id convention (anthropic/claude-sonnet-4, google/gemini-2.5-pro, meta-llama/llama-3.3-70b) lets a single key fan out across providers.

Routing rule: when $OPENROUTER_API_KEY is set and the resolved model id contains a /, weasel hands it to the OpenAI-compatible adapter pointed at https://openrouter.ai/api/v1. A bare claude-sonnet-4-6 with both $OPENROUTER_API_KEY and $ANTHROPIC_API_KEY set still goes direct to Anthropic — the / separator is the explicit signal.

• Setup:

  export OPENROUTER_API_KEY=sk-or-...

• Use:

  chimera weasel --model anthropic/claude-sonnet-4 -p "..."
  chimera weasel --model google/gemini-2.5-pro -p "..."
  chimera weasel --model meta-llama/llama-3.3-70b -p "..."

• Wired: non-streaming complete() plus base-class shim for streaming. Tool calls forwarded as standard OpenAI deltas. No provider-side caching.

Ollama (local)

Same OllamaProvider mink and otter use, same /api/chat endpoint, same keep_alive: 60m, same :cloud tag handling. Weasel falls back to Ollama automatically when no hosted key is set, which makes the zero-config laptop story work:

brew install ollama
ollama serve &
ollama pull qwen3:32b
chimera weasel -p "explain this repo"   # picks qwen3:32b

• Use:

  chimera weasel --model qwen3:32b -p "summarize"
  chimera weasel --model glm-5.1:cloud -p "long-context refactor"
  chimera weasel --model kimi-k2.6:cloud -p "deep reasoning"

• Wired: native streaming over NDJSON, tool calls (tool_calls on done:false chunks), think:true for kimi* tags, per-request num_ctx, configurable OLLAMA_HOST for remote daemons.

llama.cpp (llama-server)

Direct integration with llama.cpp’s OpenAI-compatible HTTP server. Useful for running quantized GGUF models on the metal without going through Ollama.

./llama-server -m ./models/qwen3-32b.Q4_K_M.gguf --port 8080
chimera weasel \
  --model qwen3-32b \
  --base-url http://localhost:8080/v1 \
  -p "list files"

• Wired: non-streaming + streaming chat completions, tool calls (when the model supports them), no caching, no thinking.
• The OpenAI-compatible adapter (chimera/providers/compatible.py) is what does the work; --base-url is the only extra flag.

Modal

Modal-hosted vLLM container exposing OpenAI-shape /v1/chat/completions. Weasel inherits the same adapter as mink — useful when you’ve stood up an open-weight model on Modal.

• Setup:

  pip install modal httpx
  modal token new

• Use: --model only auto-routes Anthropic / OpenAI / OpenRouter / Ollama. To call Modal, build the provider in Python:

  from chimera.providers import create_provider
  provider = create_provider(
      "modal", model="meta-llama/Llama-3.3-70B",
      base_url="https://your-org--llm-app-serve.modal.run/v1",
  )

Custom providers

Implement Provider (chimera/providers/base.py) and call register_provider("my-name", factory) (chimera/providers/registry.py). Factory signature: factory(model=..., api_key=..., base_url=..., **kw). After registration, create_provider("my-name", model=...) works identically to the built-ins.

To make weasel pick your provider automatically:

1. Hand it an explicit model with a recognizable prefix (extend _infer_provider), or
2. Construct the provider yourself and pass it to the embedded SDK Agent constructor:

   from chimera.weasel.sdk import Agent
   from chimera.providers import create_provider
   
   provider = create_provider("my-name", model="my-1.0")
   agent = Agent(provider=provider)

Self-registration on import (e.g. in your package __init__.py) mirrors the built-ins.

Choosing a provider

Concern	Pick
Default, “just works”	anthropic (claude-sonnet-4-6)
One key, many vendors	openrouter (anthropic/..., google/..., meta-llama/...)
Privacy / local	ollama (qwen3:32b) or llama.cpp
Cheap + fast	compatible against Groq / DeepSeek / Together
Vision-heavy	anthropic (claude) or openai (gpt-4o)
Long context (>200k)	google (Gemini 1M), anthropic (200k), Kimi (262k)
Reasoning-tuned	openai (o3, o3-mini) or anthropic (claude-opus + thinking)

Mixing providers in one session

Weasel holds one provider per process. To swap mid-session, exit and re-launch with a different --model or different env. The REPL /model slash command cycles through the model list passed via --models <a>,<b>,<c> — it rebuilds the provider on each switch.

In RPC mode, the host process can spawn a fresh chimera weasel --mode rpc subprocess per provider it needs to drive.

Auth storage

Weasel shares Chimera’s credential storage with the rest of the CLI:

Path	Source	Mode
~/.chimera/credentials.json	OAuth-issued tokens, refresh tokens	0o600
~/.chimera/auth.json	AuthManager.set_token()	default

CredentialStore._write chmods to 0o600 after each save (chimera/auth/store.py).

Proxy mode for teams

For an org-wide gateway sitting in front of Anthropic-shaped providers:

export ANTHROPIC_BASE_URL=http://proxy.internal:8000
export ANTHROPIC_AUTH_TOKEN=team-issued-jwt
chimera weasel --model claude-sonnet-4-6 -p "..."

AnthropicProvider honors both env vars. The dedicated proxy provider (chimera/providers/proxy.py) is the alternative when your gateway speaks its own JSON shape rather than the Anthropic wire protocol.

--list-models

Weasel ships a --list-models flag that asks each configured provider for its catalogue and prints the union:

chimera weasel --list-models
chimera weasel --list-models --json | jq '.[] | select(.provider == "anthropic")'

The list is provider-driven (Anthropic and OpenAI return live catalogues; Ollama returns installed tags) and updates every time the CLI is launched — there’s no static model file to keep current.

Troubleshooting

Symptom	Likely cause / fix
weasel: no provider configured	Set one of $ANTHROPIC_API_KEY, $OPENAI_API_KEY, $OPENROUTER_API_KEY, or pass --model <id> / $WEASEL_MODEL. Or ollama serve.
401 / 403	Wrong key. printenv | grep -E '(ANTHROPIC|OPENAI|OPENROUTER)_' to verify.
OpenRouter not used despite key	Model id needs the vendor/name / separator.
ImportError: pip install chimera-run[anthropic]	uv sync --extra anthropic to pull the SDK.
Cannot infer provider from model name '...'	Pass --model <id> with a known prefix (claude-*, gpt-*, gemini-*, glm-*, kimi-*, qwen*, …).
Streaming hangs on first call	Cloud cold start. Anthropic / OpenAI typically warm in <2s; Ollama Cloud needs keep_alive: 60m.
tool_calls always empty on Ollama	You hit /v1/chat/completions instead of /api/chat. Set OLLAMA_HOST to the daemon root, not .../v1.
llama.cpp returns 404	Confirm the OpenAI compat path. llama-server exposes /v1/chat/completions by default; pass that as the base URL.

See also

• modes.md — provider behavior is identical across modes.
• docs/mink/providers.md — line-numbered deep dive into every adapter.
• security-and-trademarks.md — auth storage + redaction posture.