Claude Code with Ollama (2026): Local Models Setup and Memory

4 min read

How to run Claude Code with Ollama local or cloud models: the ollama launch command, the manual env vars, the 64k context rule, and how memory works. Ollama docs read 2026-11-17.

Claude Code with Ollama means running Anthropic’s coding agent against a model served by Ollama, either on your own machine or in Ollama’s cloud. Ollama’s docs say it connects Claude Code to local and cloud models through an Anthropic-compatible API, and the quickest start is one command: ollama launch claude. The model does the thinking, but Claude Code still reads your files and runs commands, so the quality of the result depends on the model you pick.

Quick setup: the launch command

Ollama’s Claude Code page (docs.ollama.com/integrations/claude-code, read 2026-11-17) gives this as the way to start:

ollama launch claude

For scripts, Docker or CI, skip the menus and name the model. The docs say --yes pulls the model when needed and requires --model. Anything after -- goes straight to Claude Code:

ollama launch claude --model gemma4:cloud --yes -- -p "how does this repository work?"

Manual setup with a local model

If you want to point Claude Code at Ollama yourself, install Claude Code first (how to install Claude Code), then set three variables. These are the values from Ollama’s manual setup section:

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

ANTHROPIC_AUTH_TOKEN is the bearer token. ANTHROPIC_API_KEY is set empty so Claude Code does not send a real Anthropic key. Settings in Claude’s own config files can override these shell variables. If Claude Code connects to the wrong place, check the settings files, then run /status inside Claude Code.

Another way to send Claude Code to other models is a local router. See Claude Code Router for what it does and what Anthropic’s docs say about it.

Local or cloud: which to pick

Local model (localhost:11434)Cloud model (Ollama’s cloud)
InstallOllama on your machine, model pulledOnly Claude Code and an Ollama API key
Setup from Ollama’s docsManual variables above, or ollama launch claudeANTHROPIC_BASE_URL=https://ollama.com and ANTHROPIC_AUTH_TOKEN set to OLLAMA_API_KEY
ContextSet 64k tokens or more for larger reposNo download, larger models
Known gapsSpeed and quality depend on your hardwareHosted WebSearch and some advanced tool controls are not fully supported, per Ollama’s docs

Ollama’s docs list what works in Claude Code through Ollama: chat, command line runs with the permission flow, tool calling with compatible models, file edits, subagents, web search, web fetch and images. Tool calling depends on the model. Pick one that supports tools.

Memory with a local model

Claude Code has two built-in memory systems. CLAUDE.md files hold instructions you write, and auto memory holds notes Claude writes for itself in ~/.claude/projects/. Both are plain files that Claude Code reads at the start of a session (Claude Code memory guide). Ollama’s docs do not describe memory at all, so do not assume anything about how a local model uses these files. Test it on your own repo.

If you want memory that lasts across sessions and machines, a separate memory store is the usual route. Hindsight is one option. It is MIT-licensed and has a Coding Agents package that covers Claude Code. Install it from your repo folder with npx @vectorize-io/hindsight-coding-agents install claude-code, per the Hindsight coding agents docs, read 2026-11-12. The package gives each repo its own memory bank.

Hindsight’s README says its own server can use a local model through Ollama, LM Studio or llama.cpp, so the memory side can also run without a cloud LLM. That is a claim about the server’s options. This page did not test how well a small local model extracts facts for memory, so check the result on your own sessions.

Where Hindsight fits, with its sourced strengths:

  • Long histories. On the BEAM benchmark it scores 73.9 at 1M tokens and 64.1 at 10M. Mem0’s Platform scores 64.1 at 1M and 48.6 at 10M. Those numbers come from Mem0’s own comparison page, and they are vendor-run.
  • Open source. Hindsight’s server is MIT-licensed with the full feature set, so you can run the whole stack on your own machine.

The limits: each write calls an LLM, so a local model adds time to every save. Hindsight also returns more context per recall than Mem0 by default, so set a lower max_tokens if local context is tight. For a single short project, the built-in CLAUDE.md files may be enough.

Which memory to start with

  1. Get Claude Code running on Ollama with the launch command. Confirm it reads your repo.
  2. Set the context window to 64k tokens or more before you judge quality.
  3. Keep CLAUDE.md for rules you always want followed.
  4. Add a memory store only when sessions need to remember each other. Hindsight is the one this site covers in depth.

Sources: Ollama, Claude Code integration (read 2026-11-17); Claude Code docs; Hindsight README; Hindsight coding agents docs (read 2026-11-12).