Ollama

Local models served by Ollama on 127.0.0.1:11434. Recommended host: codex, not Claude Code.

Quick start

ollama serve &
ollama pull qwen3-coder:30b

agents harness create
# pick ollama, fill prompts, run smoke test
agents run my-profile "hello"

Required values

Var	Where to get it
Endpoint	Defaults to http://127.0.0.1:11434/v1 for the OpenAI-compatible API
Model id	Whatever you've pulled, e.g. qwen3-coder:30b
API key	Ollama doesn't require one — pass any non-empty string (ollama)

Generated profile shape

name: my-profile
host: { agent: codex }                     # Codex is OpenAI-native; talks to Ollama without a shim
env:
  # static vars (always set):
  OPENAI_BASE_URL: http://127.0.0.1:11434/v1
  # vars wizard collects:
  OPENAI_MODEL: qwen3-coder:30b
auth:
  envVar: OPENAI_API_KEY
  keychainItem: agents-cli.ollama.token

Known caveats

Use the Codex host, not Claude Code. Codex speaks OpenAI natively, which is the shape Ollama serves on /v1. Pointing Claude Code at Ollama requires a translation shim (CCR, LiteLLM, anyclaude) that converts Anthropic ↔ OpenAI — and every shim available today drops the tools array on the way through, breaking the tool_use round-trip. Use Codex and skip the shim.

Known-working model. qwen3-coder:30b works well end-to-end on Apple Silicon with Codex as the host. Other coding models work too — deepseek-coder-v2, qwen2.5-coder — but tool-call quality is model-dependent at this size.

Ollama doesn't authenticate, but Codex demands a key. Set OPENAI_API_KEY to any non-empty string (the keychain entry can be literally ollama). The value never leaves your machine.

Troubleshooting

Symptom	Cause	Fix
Tool calls silently ignored	Using Claude Code with an Anthropic→OpenAI translation shim (CCR / LiteLLM / anyclaude)	Switch the profile's host.agent to codex
connection refused on 127.0.0.1:11434	ollama serve not running	Start the daemon, or brew services start ollama
Model spins for minutes before first token	Model isn't loaded into memory yet	First request always pays the cold-start cost; subsequent runs are fast
model 'X' not found	Model not pulled	ollama pull <model>