How to use an Anthropic-compatible client with Ollama

Ollama can answer requests shaped like the Anthropic Messages API, which lets tools that expect Anthropic endpoints talk to a local Ollama server. The compatibility layer is useful when an application exposes only Anthropic-style base URL and token settings.

The local server still ignores the placeholder token, but many clients require one syntactically. Point the client at http://localhost:11434 and keep the model name set to an Ollama model that supports the requested behavior.

Compatibility does not make every Anthropic feature identical. Start with a minimal messages request, then add streaming, tools, or longer context after the basic route returns final text.

1. Set the Anthropic-compatible environment variables for local tools.

   $ export ANTHROPIC_AUTH_TOKEN=ollama
   $ export ANTHROPIC_BASE_URL=http://localhost:11434

   The token value is required by some clients but ignored by the local Ollama server.

2. Send a minimal Messages API request.

   $ curl -s http://localhost:11434/v1/messages \
     -H 'Content-Type: application/json' \
     -H 'x-api-key: ollama' \
     -H 'anthropic-version: 2023-06-01' \
     -d '{
       "model": "gpt-oss:20b",
       "max_tokens": 80,
       "messages": [{"role":"user","content":"Return only OK."}]
     }'
   {"type":"message","content":[{"type":"thinking"},{"type":"text","text":"OK"}],"stop_reason":"end_turn"}

3. Read final text from content entries with type text.

   $ python3 - <<'PY'
   content = [{"type":"thinking"}, {"type":"text", "text":"OK"}]
   print(''.join(part.get('text','') for part in content if part.get('type') == 'text'))
   PY
   OK

4. Increase max_tokens if a thinking model stops before final text.

   $ curl -s http://localhost:11434/v1/messages -d '{"model":"gpt-oss:20b","max_tokens":20,"messages":[{"role":"user","content":"Return OK."}]}'
   {"content":[{"type":"thinking"}],"stop_reason":"max_tokens"}

   For gpt-oss models, too small a token cap can spend the budget on reasoning.

5. Run the same client with the application model name after the smoke test passes.

   $ export ANTHROPIC_MODEL=gpt-oss:20b

   Use the model name from ollama list or api/tags.