Use with Ollama

There are two ways to drive Chimera with Ollama, both supported as of v0.7.1:

Path	Endpoint	Auth	What runs the model
A. Ollama Cloud direct	https://ollama.com	Authorization: Bearer $OLLAMA_API_KEY	Ollama’s GPUs (their datacenter)
B. Local daemon (passthrough or local)	http://localhost:11434	none (or ollama signin for cloud passthrough)	Local daemon → forwards :cloud ids to Ollama / keeps non-cloud ids local

Path A is simpler — no daemon, no ollama serve, just two env vars and any of the 7 CLIs. Path B is what you use if you also want truly local models on your own GPU.

---

Path A — Ollama Cloud direct (recommended for cloud models)

Setup

1. Sign up at ollama.com
2. Generate an API key at ollama.com/settings/keys
3. Set two env vars and you’re done:

export OLLAMA_API_KEY=your_key_here
export OLLAMA_HOST=https://ollama.com

That’s it. No daemon, no ollama serve, no ollama pull. Requests go straight to Ollama’s API with bearer-token auth.

Verify

List the models your key has access to:

curl -s -H "Authorization: Bearer $OLLAMA_API_KEY" https://ollama.com/api/tags | jq '.models[].name'

You should see ~30-40 cloud models including gpt-oss:120b, glm-5.1, qwen3-coder:480b, kimi-k2.6, deepseek-v4-pro.

Run any of the 7 CLIs

# Pick the codename for your task — chimera which --task "..." helps
chimera mink   --model gpt-oss:120b-cloud  -p "explain this repo"
chimera otter  --model glm-5.1:cloud       -p "summarize TODO comments"
chimera ferret --model deepseek-v4-pro     --sandbox workspace-write -p "fix tests"
chimera weasel --model gpt-oss:120b-cloud  -p "list py files in chimera/" --json
chimera shrew  --model qwen3-coder:480b-cloud -p "explain this directory"
chimera stoat  --model kimi-k2.6-cloud     -p "what is 8 - 2?"
chimera badger --model glm-5.1:cloud       -p "verify the test suite passes"

The -cloud and :cloud suffixes work interchangeably. When $OLLAMA_HOST points at ollama.com, Chimera strips the suffix automatically for the direct API. When pointed at a local daemon, it keeps the suffix so the daemon knows to forward.

Sessions to disk

Every run is event-sourced. Inspect and resume:

chimera mink runs list
chimera mink runs show mink-20260514T034436-46db8b28
chimera resume <run-id>          # auto-detects which CLI saved it

---

Path B — Local daemon (for truly local models + cloud passthrough)

Setup

1. Download Ollama and run ollama serve (or just let the app boot).
2. Optional: pull local models. ollama pull qwen3-coder-30b, ollama pull glm-4.7-flash, etc.
3. Optional: ollama signin to enable :cloud-suffixed passthrough.
4. Point Chimera at the daemon. Either drop in via the Anthropic-compatible adapter:

export ANTHROPIC_BASE_URL=http://localhost:11434
export ANTHROPIC_AUTH_TOKEN=ollama
chimera mink --model glm-5.1:cloud -p "..."

…or use the native Ollama provider (and don’t touch ANTHROPIC_*):

unset OLLAMA_HOST OLLAMA_API_KEY ANTHROPIC_BASE_URL
chimera shrew --model qwen3-coder:30b -p "..."   # local, free, your hardware
chimera shrew --model glm-5.1:cloud -p "..."     # daemon forwards to Ollama Cloud

Use this path when you want any of:

• A local model running on your own GPU / Apple Silicon
• The daemon’s keep_alive model caching for many small turns
• SSH-key based cloud auth (handled by ollama signin)

---

Recommended models (as of 2026-05-13, on ollama.com)

Model id	Hosting	Approx context	Best for
gpt-oss:120b-cloud	Cloud	128k	General-purpose, fast, free at time of writing
glm-5.1:cloud	Cloud	128k+	Coding / refactoring; Chimera benchmarks against it
kimi-k2.6:cloud	Cloud	200k+	Long sessions, tool-use heavy
kimi-k2-thinking	Cloud	1M+	Extended thinking, reasoning-heavy
qwen3-coder:480b-cloud	Cloud	131k	Pure coding tasks
deepseek-v4-pro	Cloud	262k	Latest DeepSeek; reasoning + code
minimax-m2.7:cloud	Cloud	1M+	Very long contexts
qwen3-coder	Local (~60GB)	131k	Local coding on Apple Silicon / GPU
glm-4.7-flash	Local	128k	Local coding, smaller footprint

Browse the full live catalog:

curl -s -H "Authorization: Bearer $OLLAMA_API_KEY" https://ollama.com/api/tags | jq '.models[] | {name, size_gb: (.size / 1e9 | round)}'

---

Per-CLI tips

• chimera mink — TUI-first, slash-command palette. Reads ~/.claude/settings.json for hooks / permissions / MCP. Best for interactive sessions where you want a REPL with rich rendering.
• chimera otter — Multi-client server. Three transports (one-shot, HTTP+SSE, ACP). Use when you want a long-running agent server other tools connect to.
• chimera ferret — Sandbox-first. --sandbox read-only|workspace-write|danger-full-access + --approval suggest|auto|yolo. Best for destructive tasks where you want a safety net.
• chimera weasel — Minimal harness. Four modes: interactive REPL, -p print, JSON-RPC over stdio, embedded SDK. Best for scripting / pipelines / programmatic use.
• chimera shrew — Tuned for small local models. Smaller default --max-steps, restricted default tools, output-parser repair for text-mode tool calls, quality monitor for empty/hallucinated responses.
• chimera stoat — Shell-mode toggle (Ctrl-X). Each REPL line either feeds the LLM or runs as a bash command. Mode-tagged history.
• chimera badger — Strict harness posture. Tighter --max-steps, --rerun-on-failure, parity tracking. Best for repeatable / regression-style work.

Run chimera agents to see this same matrix any time. Or chimera which --task "describe what you want" to get a recommendation.

---

Troubleshooting

HTTP 401 — Auth rejected by upstream

Either $OLLAMA_API_KEY is wrong, expired, or unset. Verify with the curl above.

HTTP 404 ... /v1/chat/completions

Your shell still has a stale $ANTHROPIC_BASE_URL pointing at localhost:11434 while no daemon is running. unset ANTHROPIC_BASE_URL ANTHROPIC_AUTH_TOKEN.

HTTP 503 — Upstream issue

Ollama’s upstream is having trouble. The provider returns a friendly error and a hint. Retry, or switch model: --model glm-5.1:cloud is usually up when others are down.

stoat: $MOONSHOT_API_KEY is required for kimi-* models (fixed in v0.7.1+)

Use the -cloud or :cloud suffix: chimera stoat --model kimi-k2.6-cloud. Bare kimi ids still need a Moonshot key.

Models are slow on first call

The daemon (path B) does cold-loads. Subsequent calls are fast due to keep_alive. Path A has no cold-load.

---

Cost & privacy

Path	Cost	Privacy
A: Cloud direct	Free at time of writing (Ollama’s loss-leader pricing). May change.	Calls + payloads go to Ollama’s servers.
B: Local daemon, non-cloud model	$0	Stays on your machine.
B: Local daemon, :cloud model	Same as A — the daemon just proxies.	Same as A.

For sensitive code, prefer Path B with a non-cloud model.

---

Next

• Architecture — how the 8 phases compose
• Per-CLI quickstarts — chimera mink deep-dive (and otter / ferret / weasel / shrew / stoat / badger)
• Inspirations — which upstream tool inspired each codename