setup-ollama-local
Skill388 starsChanged 7 months agoMITView on GitHub
Interactively guide setup for connecting the Claude Code CLI to a local Ollama LLM on Mac. NOTE — This is for the standalone Claude Code CLI itself, NOT MulmoClaude (MulmoClaude does not currently support Ollama as a backend). Covers Ollama install, model pull, env switching, and verification. Respond in the user's language.
- Pipes a download into a shell
- Reads credentials
- Installs packages
In this repository13 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- archive-shipped-plansSkill
- e2e-live-dockerSkill
- e2e-live-mediaSkill
- e2e-liveSkill
- e2e-live-wikiSkill
- make-e2e-liveSkill
- publish-mulmoclaudeSkill
- release-appSkill
- setup-mulmobridge-lineSkill
- setup-mulmoclaudeSkill
- setup-relaySkill
- setup-wizardSkill
The file
We split this skill into the rules it gives an agent
4 rules. About 0 are common, 1 is distinctive, and 3 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈1
Distinctive · estimate
This author's own choices. Worth a look.
3
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 45
Empty / connection refused: start it.
- Very specific: line 73
Avoid: qwen3:14b (40k training limit), qwen2.5-coder:14b (older runner ignores OLLAMA_CONTEXT_LENGTH), gemma4:26b (Content block parse …
- Very specific: line 147
If the user wants a one-liner, suggest an alias in ~/.zshrc.
- Very specific: line 153
After source ~/.zshrc, usage is:
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Setup Claude Code × Ollama (local LLM)
Scope / 適用範囲
This skill sets up the standalone
claudeCLI to talk to a local Ollama server. It is independent of MulmoClaude; MulmoClaude itself does not currently support Ollama (seeplans/feat-mulmoclaude-ollama-support.mdfor a tentative plan).このスキルは
claudeCLI 単体をローカルの Ollama サーバに接続するセットアップです。MulmoClaude とは独立しており、MulmoClaude 本体は現在 Ollama 接続をサポートしていません(実装案はplans/feat-mulmoclaude-ollama-support.mdを参照)。
For detailed findings and pitfalls, see docs/tips/claude-code-ollama.md (Japanese) / docs/tips/claude-code-ollama.en.md (English).
Prerequisites / 前提知識
- Ollama v0.14.0 or later is required (Anthropic Messages API compatibility was added in that version).
- Claude Code sends roughly 50,000–57,000 tokens per request, so the model needs at least a 64k context window.
- 3B-class small models effectively cannot drive Claude Code (no tool calling, broken templates).
- Even on supported models, the first turn often takes 10+ minutes on a MacBook Air; subsequent turns benefit from KV cache and drop to 1–3 minutes.
Step 1: Verify / install Ollama
1-1. Check existing install
which ollama && ollama --version- Installed and v0.14.0+: proceed to 1-2.
- Older version:
brew upgrade ollamaand thenbrew services restart ollama(Homebrew installs). - Not installed: suggest one of:
- Official installer (recommended): https://ollama.com/download/mac
- Homebrew:
brew install ollama
1-2. Verify the server is running
curl -s http://localhost:11434/api/tags | head -c 200- Got JSON back: server is up, go to Step 2.
- Empty / connection refused: start it.
- Official app: click the Ollama menu-bar icon.
- Homebrew:
brew services start ollamaorollama serve.
Step 2: Verify Claude Code
which claude && claude --version- Installed: continue to Step 3.
- Not installed: install via npm (recommended), the official script, or Homebrew:
npm install -g @anthropic-ai/claude-codecurl -fsSL https://claude.ai/install.sh | shbrew install anthropic/tap/claude-code
Step 3: Choose and pull a model
Confirm the user's RAM and use case before recommending. Verified working models on a MacBook Air M4 32GB are summarized in docs/tips/claude-code-ollama.md. Quick picks:
| RAM | Recommended | Size | Notes |
| 8–16GB | (Claude Code × Ollama is impractical here) | — | Cold start exceeds 10 min timeout |
| 32GB | qwen3.5:9b | 6.6GB | Most practical, lightest fit |
| 32GB | qwen3.6:35b-a3b | 23GB | MoE (3B active), heavier but works |
| 16–32GB | gemma4:e4b | 3GB on disk (~10.9 GiB resident) | Verified on 32GB; thinking blocks render correctly |
| 24GB+ (NVIDIA) | glm-4.7-flash | 19GB | 198k context, untested on Mac |
Avoid: qwen3:14b (40k training limit), qwen2.5-coder:14b (older runner ignores OLLAMA_CONTEXT_LENGTH), gemma4:26b (Content block parse errors — note: gemma4:e4b is fine), gpt-oss:20b (Ollama template bug). See findings doc for details.
ollama pull <model>
ollama listStep 4: Start Ollama with the right context window
Claude Code requires ≥64k context. The default is 32k, so always extend it when launching Ollama for Claude Code:
brew services stop ollama # if running under brew services
OLLAMA_CONTEXT_LENGTH=65536 ollama serveThis terminal must stay open for the duration of the session. For longer sessions add OLLAMA_KEEP_ALIVE=30m so the KV cache survives idle gaps.
Step 5: Warm up the model
In a second terminal, load the model into memory and confirm it responds at all:
ollama run <model> "hello"Expect a response within a few seconds. If this hangs, the model is unsuitable for Claude Code.
Step 6: Run Claude Code against Ollama
In a third terminal, set the env vars and launch:
export ANTHROPIC_AUTH_TOKEN="ollama"
export ANTHROPIC_API_KEY=""
export ANTHROPIC_BASE_URL="http://localhost:11434"
claude --verbose --model <model>Role of each variable:
| Variable | Value | Purpose |
ANTHROPIC_AUTH_TOKEN | "ollama" | Enables Ollama mode |
ANTHROPIC_API_KEY | "" (empty) | Disables the cloud API key (prevents collision) |
ANTHROPIC_BASE_URL | http://localhost:11434 | Routes API calls to the local server |
Send a simple message (e.g. "Hello, what model are you?") to confirm. The first turn can take 10+ minutes; subsequent turns drop to 1–3 minutes once the KV cache is warm.
While waiting, watch the Ollama log in another terminal to see what's happening:
tail -f /opt/homebrew/var/log/ollama.log # Homebrew install
# or just watch the terminal where `ollama serve` is runningKey log signals:
KvSize:65536✓ — context is correctly extendedtruncating input prompt limit=XXXXX✗ — model/runner ignores the env var; switch modelPOST /v1/messages 200✓ — successful turnPOST /v1/messages 500✗ — template incompatibility; switch model
Step 7: Switching back to cloud Claude
The local mode is scoped to the terminal where the env vars were set:
- Easiest: close that terminal and open a fresh one — back to cloud.
- Or unset explicitly:
unset ANTHROPIC_AUTH_TOKEN ANTHROPIC_API_KEY ANTHROPIC_BASE_URL
Step 8 (optional): Convenience alias
If the user wants a one-liner, suggest an alias in ~/.zshrc. Do not put bare export ANTHROPIC_BASE_URL=... lines in a startup file — that breaks normal cloud usage everywhere.
alias claude-local='ANTHROPIC_AUTH_TOKEN="ollama" ANTHROPIC_API_KEY="" ANTHROPIC_BASE_URL="http://localhost:11434" claude'After source ~/.zshrc, usage is:
claude-local --model qwen3.5:9b # local
claude # cloud, unchangedKey pitfalls to highlight
- Ollama < v0.14.0 has no Anthropic API compatibility — always check the version first.
ANTHROPIC_API_KEYmust be explicitly empty; otherwise an existing cloud key may collide.- 3B-class models cannot reliably emit Claude's tool-use JSON, so file edits and shell commands fail.
- Even tool-capable open models (Gemma 4, gpt-oss) are not fully aligned with Anthropic's response format — complex skill chains misbehave.
- Large models (20B+) eat memory; watch with
vm_stator Activity Monitor. - Permanent
export ANTHROPIC_BASE_URL=...in.zshrc/.bashrcwill silently break normal cloud Claude usage. Use an alias instead. - The first-turn 10-minute Claude Code timeout is unavoidable, but Ollama keeps processing in the background, so a retry usually succeeds via cache reuse.
Reference links
- Ollama Claude Code integration: https://docs.ollama.com/integrations/claude-code
- Ollama Anthropic compatibility blog: https://ollama.com/blog/claude
- Claude Code official docs: https://code.claude.com/docs/en/overview
- Local findings:
docs/tips/claude-code-ollama.md(ja) /.en.md(en)
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

