CanyouuseClaudeCodewithOllamaoranotherlocalmodel?

Yes, an Ollama Claude Code setup works. Ollama, LM Studio and llama.cpp each serve an Anthropic-style endpoint, and Claude Code lets you change its base URL, so a local model can sit behind it. Anthropic doesn’t support that, and the local model is the weak link. Here is what runs, what breaks and what developers report.

Claims below were checked against the linked docs on 2026-10-26. Local-model support in Claude Code changes fast, so re-check the version notes before you rely on a detail.

Does Claude Code support non-Anthropic or local models?

Not officially. Claude Code’s gateway documentation says any gateway that exposes a supported API format works, then adds: “Anthropic doesn’t endorse, maintain, or audit third-party gateway products, and doesn’t support routing Claude Code to non-Claude models through any gateway.” The mechanism exists, and the runtime vendors support their side of it. Anthropic does not.

What makes it possible is protocol, not intent. According to the gateway compatibility guide, Claude Code speaks three formats: Anthropic Messages (selected with ANTHROPIC_BASE_URL), Amazon Bedrock InvokeModel, and Google Cloud’s Agent Platform rawPredict. A server that implements Anthropic Messages at /v1/messages looks like the Claude API to Claude Code, which, per the same guide, “can’t tell which upstream you forward to.”

Three local runtimes now implement that endpoint:

Anything unofficial in this article is marked as such. Treat the whole local setup as unsupported by Anthropic and supported, to varying degrees, by the runtime.

How do you point Claude Code at Ollama, LM Studio or llama.cpp?

Set ANTHROPIC_BASE_URL to the local server and give Claude Code a placeholder credential. The Claude Code docs explain why you need the credential too: setting only the base URL “doesn’t replace the subscription”, so your saved claude.ai login stays the active credential. A token variable takes precedence over it.

Ollama. Per the Ollama Claude Code page, the manual setup is:

export ANTHROPIC_BASE_URL=http://localhost:11434
export ANTHROPIC_AUTH_TOKEN=ollama
export ANTHROPIC_API_KEY=""
claude --model qwen3.5

The same page documents ollama launch claude to start Claude Code against Ollama without exporting variables by hand, and says Claude settings can override your shell variables, so run /status to confirm the connection.

LM Studio. The LM Studio docs start the server on port 1234 (lms server start --port 1234) and set:

export ANTHROPIC_BASE_URL=http://localhost:1234
export ANTHROPIC_AUTH_TOKEN=lmstudio
export CLAUDE_CODE_ATTRIBUTION_HEADER=0
claude --model openai/gpt-oss-20b

If “Require Authentication” is on in LM Studio, put your LM Studio API token in ANTHROPIC_AUTH_TOKEN instead of the placeholder.

llama.cpp. The ggml-org post starts llama-server with a tool-capable GGUF and points Claude Code at http://127.0.0.1:8080. The post shows ANTHROPIC_BASE_URL alone, but a commenter there reports Claude Code still asked for a login until ANTHROPIC_AUTH_TOKEN and ANTHROPIC_MODEL were also set. A Hacker News commenter describes the same symptom with llama.cpp (comment, Jan 2026). That matches the Claude Code docs on credentials, so set the token. Another commenter reports that telemetry pings caused network failures in their setup and fixed it with CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1 (comment); that is a single community report, so check the Claude Code docs for that variable before relying on it.

Two details apply to every runtime:

• Don’t add /v1 to the base URL. Claude Code appends /v1/messages itself, and the compatibility guide lists that as the endpoint.
• Verify before you trust it. Run /status and check that the Anthropic base URL line shows your local address and an Auth token line is present. The connect page describes both checks.

Do you need a translation proxy?

Only when your server does not speak one of Claude Code’s three formats. Ollama, LM Studio and llama.cpp now do, so for them no proxy is needed. A server that exposes only an OpenAI-style /v1/chat/completions endpoint has nothing Claude Code can call, and you would need a gateway in front that translates the two. Any such proxy is outside Anthropic’s support statement, and this article doesn’t cover specific proxy products because we haven’t verified them.

Which local models handle agentic tool calls?

One Hacker News commenter reports Qwen3-Coder 30B-A3B at Q8 with a 32k context, split across a 24 GB and a 16 GB GPU, running well behind Claude Code on Ollama (comment, Jan 2026); another found Devstral 24B and quantized Qwen3 30B-A3B fine for simple tasks but unreliable at tool use (comment, Dec 2025). So use a model that is explicitly trained or packaged for tool calling, and treat vendor claims as claims. Claude Code is a tool-calling loop: every file read, edit and shell command is a tool call, so a model that emits malformed calls fails visibly.

The runtime docs name a handful of models. None of them publish a Claude Code comparison, so this table records what each source actually says.

To compare tool-calling ability across models with numbers rather than model-card prose, the Berkeley Function Calling Leaderboard is the neutral reference. It measures function calling, including multi-turn use in V4. It doesn’t measure Claude Code sessions. For a coding-specific shortlist, see best local LLM for coding in 2026.

What breaks when the model is local?

A developer on an Apple M4 Max with 128 GB says local coding finally feels usable, but Claude Code’s large bootstrap context slows everything down, so they switched to a lighter agent (comment, Jul 2026). Another calls Claude Code on Ollama “doable, but not really viable” for larger tasks (comment, Jun 2026). Context length breaks first, then protocol gaps, then speed. Roughly in that order:

1. Context window. Ollama’s context docs set defaults by VRAM: 4k under 24 GiB, 32k at 24 to 48 GiB, 256k at 48 GiB or more. They say agents and coding tools should run “at least 64000 tokens”, set with OLLAMA_CONTEXT_LENGTH=64000 ollama serve. The Ollama Claude Code page repeats the 64k advice. LM Studio is more modest, recommending more than about 25k. A 4k default under a tool-heavy agent loop will fail in confusing ways.
2. Compaction doesn’t know your limit. For a model ID it doesn’t recognize, Claude Code assumes a 200k window (it warns “context assumed to be 200k tokens”). The gateway guide says to set CLAUDE_CODE_AUTO_COMPACT_WINDOW to the server’s limit, but values are clamped to at least 100,000 tokens. With a 64k local window you can’t match it, so /compact by hand becomes your recovery.
3. Missing Anthropic features. Ollama’s compatibility page lists as unsupported: token counting, tool_choice, prompt caching via cache_control, the Batches API, citations and PDF blocks. Without a count endpoint, Claude Code falls back to a character-based estimate, so /context counts are approximate.
4. Hosted features. The Ollama Claude Code page says hosted WebSearch and advanced tool controls are “not fully supported” with Ollama’s cloud endpoint. Claude Code’s own docs say Remote Control is disabled while ANTHROPIC_BASE_URL points at a non-Anthropic host.
5. Prompt prefix and caching. LM Studio’s docs add CLAUDE_CODE_ATTRIBUTION_HEADER=0. Claude Code’s gateway guide explains the block it controls: from v2.1.181 it is stable within a conversation behind a custom base URL, and before that it changed every request and hurt prompt-cache reuse. On a current version you may not need the flag.
6. Quality. There is no verified head-to-head number to quote here. Aider’s docs warn it “may not work well with less capable models”, and Anthropic states it doesn’t support non-Claude models. Test how your model handles tool calls on your own repository before relying on it. For a fuller comparison, see local LLM vs Claude for coding.

On speed: neither the runtime docs nor Claude Code’s docs publish tokens-per-second figures for this setup, so this article doesn’t either. Your hardware sets it. The runtime tradeoffs are covered in Ollama vs llama.cpp vs vLLM.

What do developers report in practice?

These are community reports from Hacker News, not our tests, and each is one person’s setup. Treat them as leads to check.

The pattern across independent reports: hardware that fits a 30B-class coding model at 32k or more makes it work, tool-call reliability varies by model and quant, and Claude Code’s own prompt overhead is a cost on local hardware. Those reports back up the context and quality caveats above, but none of them is a benchmark.

What are the alternatives built for local models?

If the goal is a local coding agent rather than Claude Code specifically, tools designed around local models need less workaround. All four below have documented Ollama support.

Notice what these have in common: each documents a context-window setting. That is the main failure mode for local agents regardless of the front end. For an editor-side setup, see self-hosted Copilot alternatives.

FAQ