Run Claude Code with Ollama: ollama launch claude, manual setup, context length, models

Prerequisites

• Ollama 0.14.0 or newer (ollama --version).
• Claude Code (curl -fsSL https://claude.ai/install.sh | bash).
• Memory: measured with a 32K context, gpt-oss:20b occupies 12 GB and qwen3-coder 19 GB (see the model section). On Linux that means GPU memory for full speed, otherwise Ollama falls back to the CPU and everything becomes slow.

ollama launch claude: the short way

Recent Ollama releases include a launcher that configures Claude Code for you:

ollama launch claude                       # pick a model from the menu, Claude Code starts
ollama launch claude --model gpt-oss:20b --yes      # non-interactive, pulls the model if needed
ollama launch claude --model gpt-oss:20b --yes -- -p "how does this repository work?"
ollama launch claude --config              # configure without starting
ollama launch claude --restore             # back to the defaults

The launcher sets the environment variables described below for that session, so you do not have to. It does not set the context length — that is a server setting, and the next sections explain it — and it hides what happens, which is why the manual way is worth knowing.

Step 1: Pull a model that can call tools

ollama pull gpt-oss:20b      # 14 GB download, 12 GB loaded with 32K context; 128K native
ollama pull qwen3-coder      # 19 GB loaded, 256K native; the better coder, wants 32 GB
ollama pull qwen3.8          # 18 GB download, dense 27B; strong, needs 32 GB or more

Claude Code drives the model through tool calls (read file, edit file, run command). All three support that (ollama show <model> lists tools under Capabilities), while many small chat models do not, and with those the agent talks but never acts. What fits where, measured on a 24 GB Mac with a 32K context: gpt-oss:20b loads entirely on the GPU and completes edit tasks. qwen3-coder loads with 16 of its 19 GB on the GPU and the rest on the CPU, which works but is slower. qwen3.8:27b runs out of GPU memory as soon as Claude Code's 20K-token prompt arrives. On 16 GB, use gemma4 (the 12B at about 8 GB) or qwen3.5:9b with a 16K context.

Step 2: Set the context length

Ollama chooses a default context from available GPU memory: 4K below 24 GiB, 32K between 24 and 48 GiB, 256K above. Claude Code's system prompt and tool definitions alone are about 20K tokens, so on a laptop the 4K default fails on the first request with a context-limit error. Start the server with an explicit context value:

OLLAMA_CONTEXT_LENGTH=32768 ollama serve

On macOS with the Ollama app, move the context-length slider in the app's settings instead (the app runs the server for you), or quit the app and run ollama serve from a terminal as above. On Linux with the systemd service, add the variable to the service:

sudo systemctl edit ollama
# in the editor, add:
# [Service]
# Environment="OLLAMA_CONTEXT_LENGTH=32768"
sudo systemctl restart ollama

On Windows, Ollama runs from the system tray and reads the variable from your account's environment: quit Ollama, open Edit environment variables for your account (search for "environment variables" in Settings), add OLLAMA_CONTEXT_LENGTH with the value 32768, and start Ollama again from the Start menu. I have not run the Windows setup myself, so this follows Ollama's documentation.

Verify with ollama ps: the CONTEXT column must show 32768 (or more) once a model is loaded. Ollama's own recommendation for coding agents is 64K, which is worth using when the memory allows, while 32K is the floor at which Claude Code works.

Step 3: Point Claude Code at Ollama

export ANTHROPIC_BASE_URL=http://localhost:11434
export ANTHROPIC_AUTH_TOKEN=ollama          # required by Claude Code, ignored by Ollama
export ANTHROPIC_DEFAULT_HAIKU_MODEL=qwen3-coder
export ANTHROPIC_DEFAULT_SONNET_MODEL=qwen3-coder
export ANTHROPIC_DEFAULT_OPUS_MODEL=qwen3-coder
export CLAUDE_CODE_MAX_CONTEXT_TOKENS=32768    # the context you set on the server
claude --model qwen3-coder

In PowerShell the same lines read $env:ANTHROPIC_BASE_URL = "http://localhost:11434" and so on, or put them in .claude/settings.json as shown below, which is identical on every platform.

CLAUDE_CODE_MAX_CONTEXT_TOKENS matters since Claude Code 2.1.193: for a model name it does not know it assumes a 200K window (and says so with a [claude-code:unrecognized_model] line), while with the real value it compacts the conversation before Ollama rejects it. Ollama's documentation also sets ANTHROPIC_API_KEY="" so that no stale key competes with the token. The alias mapping matters: Claude Code uses its haiku alias for background work, and without the mapping that call asks Ollama for an Anthropic model name it does not have. (Ollama's documentation offers ollama cp qwen3-coder claude-3-5-sonnet as an alternative — copying the model under an Anthropic name — but the environment variables are cleaner.) The same variables can live in .claude/settings.json under "env" for a per-project setup, and the LM Studio guide shows the JSON.

/status inside Claude Code confirms the base URL in use. Add DISABLE_TELEMETRY=1, DISABLE_ERROR_REPORTING=1 and CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1 if the session should not talk to anything but Ollama.

Step 4: Check the endpoint directly if something fails

curl http://localhost:11434/v1/messages \
  -H "content-type: application/json" -H "x-api-key: ollama" \
  -d '{"model":"qwen3-coder","max_tokens":64,
       "messages":[{"role":"user","content":"Say hello in one line."}]}'

The model on another machine

A laptop with 16 GB and a desktop with a GPU is a common pair. Ollama on the desktop listens on all interfaces when started with OLLAMA_HOST=0.0.0.0 ollama serve (or the same variable in the systemd service). On the laptop, point Claude Code at that host:

export ANTHROPIC_BASE_URL=http://desktop.local:11434     # or a Tailscale/VPN name
export ANTHROPIC_AUTH_TOKEN=ollama
export OLLAMA_HOST=http://desktop.local:11434            # so that `ollama ps` etc. also go there

The rest is identical, and I run this setup over Tailscale, where the only difference is the hostname. Keep such a server off the open internet: Ollama has no authentication of its own.

What Ollama's endpoint supports — and what it does not

Supported: streaming, system prompts, multi-turn conversations, tool calling, extended thinking, base64 images. Not supported: forced tool_choice, prompt caching, the token-counting endpoint, URL images, PDFs, batches. In practice the missing token counter means Claude Code's context indicator can be off, and the missing prompt cache means every turn re-processes the full prompt — the main reason local sessions feel slower than the API even on fast hardware.

Troubleshooting

400 with a context or token limit in the message

The server runs with the default 4K context. Set OLLAMA_CONTEXT_LENGTH as above and confirm with ollama ps.

No answer at all, or "stream disconnected" in the server log

The model plus its KV cache does not fit the GPU memory, so the Metal or CUDA runner aborts (Insufficient Memory in ollama serve's output) and Claude Code shows nothing or exits. A dense 27B model with a 32K context does this on 24 GB. Use a smaller model or a shorter context.

"claude.ai connectors are disabled because ANTHROPIC_API_KEY or another auth source is set"

Expected with ANTHROPIC_AUTH_TOKEN set, which turns the claude.ai-bound features off for the session and is harmless.

404 model not found, often mid-session

The haiku alias is unmapped. Set ANTHROPIC_DEFAULT_HAIKU_MODEL.

Connection refused

ollama serve is not running, or the app runs on a different port.

Answers arrive but no file changes

The model does not support tool calling, so use qwen3-coder or gpt-oss:20b.

Very slow on Linux

The model does not fit in GPU memory and runs partly on the CPU, and ollama ps shows the CPU/GPU split. Use a smaller model or quantisation.

Related: Claude Code with LM Studio on Apple Silicon and OpenAI Codex CLI with a local model.

Sources: Ollama, Claude Code integration · Ollama, Anthropic compatibility · Ollama blog, Claude Code with Anthropic API compatibility (16 January 2026) · Ollama, context length · Claude Code, model configuration.