FAQ: Providers · rousseau-agent docs

Overview

Fifteen provider-focused questions covering all five LLM backends.

Questions

1. Which provider is default?

claudecli. See setDefaults in internal/config/config.go.

2. Do I need an API key for the default?

No. claudecli shells out to the local claude binary and inherits its auth.

3. How do I switch to Anthropic API directly?

Set provider: anthropic and export ANTHROPIC_API_KEY. See Migrations: Provider migration.

4. Does Bedrock need explicit keys in the config?

No. Rousseau uses the standard AWS credential chain: env vars, ~/.aws/credentials, IMDS, IRSA. Set bedrock.profile for a named profile.

5. Does Vertex work with Application Default Credentials?

Yes. Leave vertex.credentials_file empty and set GOOGLE_APPLICATION_CREDENTIALS, or run gcloud auth application-default login.

6. Which OpenAI-compatible endpoints work?

Any that speak the Chat Completions shape. Tested: OpenRouter, Ollama, LM Studio, vLLM. openai.base_url is the knob.

7. Does Ollama need an API key?

No. setDefaults seeds ollama.api_key: not-required.

8. Does prompt caching work?

Yes on anthropic. The client writes cache markers on the last two messages. Other providers depend on their own caching layer.

9. What model do you recommend?

claudecli: whatever claude is configured to use.
anthropic: claude-sonnet-4-6 (the default).
bedrock: anthropic.claude-sonnet-4-6-20250101-v1:0.
vertex: claude-sonnet-4-6@20250101.
Local: llama3.1:70b-instruct minimum for reliable tool-use.

10. Can I switch providers mid-session?

Yes — providers don't leak model-specific state into the session. Edit config.yaml, restart. History is preserved.

11. Which provider is cheapest?

Depends on scale. claudecli with a subscription plan is flat-rate. Ollama is free after model download. Pay-per-token providers vary; enable prompt caching + compression to reduce spend.

12. Which provider is fastest?

Anthropic direct usually wins on TTFT. Ollama depends on your GPU. Bedrock/Vertex add STS/IAM latency but stream fine.

13. Does tool-use work on all providers?

Anthropic / claudecli / Bedrock / Vertex: excellent. OpenAI-compatible: depends on the model — GPT-class is fine, small local models often fail structured tool inputs.

14. Can I use two providers simultaneously?

Not from one config. Run two rousseau processes with distinct configs and state paths.

15. Where do provider errors show up?

Structured logs with err attribute. Grep journalctl … | grep 'provider.error' or the equivalent event name for your provider.