LLM providers

Dapper runs on Anthropic’s Claude by default and is built and optimized for it. An experimental router mode can route requests through alternative providers (OpenAI, OpenRouter, Ollama) via claude-code-router, for model experimentation only.

  1. Anthropic (default, recommended)
  2. Router mode (experimental)
    1. Supported providers
    2. OpenAI
    3. OpenRouter
    4. Ollama (local)
  3. How routing is wired
  4. Caveats

The default and supported path. Provide an Anthropic credential in .env — either an API key or a Claude Code OAuth token — and run normally (no ROUTER flag).

# .env — recommended
ANTHROPIC_API_KEY=sk-ant-...

# or, instead of the API key:
# CLAUDE_CODE_OAUTH_TOKEN=your-oauth-token-here

./dapper start URL=https://example.com REPO=repo-name

Get an API key from console.anthropic.com. See Environment variables for the full credential reference.

Router mode (experimental)

Experimental and unsupported. Dapper is built on the Anthropic Agent SDK and optimized for Claude. Alternative providers may produce inconsistent results — including failing early phases such as Recon — depending on the model and routing setup. Bug reports against router mode are best-effort.

Router mode is intended for trying Dapper with non-Claude models. Enable it by configuring one provider in .env, then passing ROUTER=true:

./dapper start URL=https://example.com REPO=repo-name ROUTER=true

When ROUTER=true, the ./dapper script starts the claude-code-router container and points the Agent SDK at it. The router reads ROUTER_DEFAULT (format: provider,model) to pick the active provider and model.

Supported providers

Provider Models API key Notes
OpenAI gpt-5.2, gpt-5-mini OPENAI_API_KEY Good tool use; balanced cost/performance.
OpenRouter google/gemini-3-flash-preview (and others OpenRouter exposes) OPENROUTER_API_KEY Access to Gemini 3 family and many models via one API.
Ollama any locally pulled model (e.g. llama3.3, qwen2.5, deepseek-r1) none Free local inference; configurable via OLLAMA_BASE_URL.

OpenAI 

# .env
OPENAI_API_KEY=sk-your-openai-key
ROUTER_DEFAULT=openai,gpt-5.2

./dapper start URL=https://example.com REPO=repo-name ROUTER=true

OpenRouter 

# .env
OPENROUTER_API_KEY=sk-or-your-openrouter-key
ROUTER_DEFAULT=openrouter,google/gemini-3-flash-preview

./dapper start URL=https://example.com REPO=repo-name ROUTER=true

Ollama (local)

No API key is required. By default the router reaches Ollama at http://host.docker.internal:11434; override with OLLAMA_BASE_URL if it runs elsewhere.

# .env
ROUTER_DEFAULT=ollama,llama3.3
# OLLAMA_BASE_URL=http://host.docker.internal:11434   # default; change if needed

./dapper start URL=https://example.com REPO=repo-name ROUTER=true

Ollama satisfies the credential check on its own — with ROUTER_DEFAULT=ollama,<model> you do not need an Anthropic or other provider key set.

How routing is wired

With ROUTER=true, the script:

  1. Starts the router profile container (claude-code-router) if it is not already running.
  2. Sets ANTHROPIC_BASE_URL=http://router:3456 and an auth token so the Agent SDK talks to the router instead of Anthropic directly.
  3. The router (configs/router-config.json) maps the request to the provider/model named in ROUTER_DEFAULT.

./dapper stop (and ./dapper stop CLEAN=true) also bring the router container down.

Caveats

  • Quality varies. Some providers fail early phases such as Recon.
  • No official support. Router mode is best-effort.
  • Costs and rate limits differ per provider — track your own quota.
  • Optimized for Claude. For reliable results, use Anthropic models.