Running Claude Code Locally Just Got Easier with ollama-code

Back in May, I walked through how to run Claude Code locally by emulating Anthropic’s API and using a local LLM with the same prompt structure. It worked, but it was a little clunky. You had to manually map Claude’s behavior onto a local model, spin up your own fake endpoint, and tweak system prompts to get decent results. Now, thanks to ollama-code, you don’t have to.

This repo makes it dead simple to replicate Claude Code’s experience using open models like codellama, deepseek-coder, or codestral, locally via Ollama.

Here’s what changed.

What is `ollama-code`?

ollama-code is a lightweight wrapper around the Ollama ecosystem that mimics Claude Code’s linear, REPL-style development experience. It’s opinionated, and that’s a good thing.

Handles prompt engineering and system message injection
Loads your preferred code-capable model (like codellama:34b)
Maintains an in-memory “project context” for incremental development
Optionally integrates with shell commands or interpreters

It’s built by @tcsenpai, who’s been shipping fast, useful dev tools. This one nails it.

Why It’s Better Than My Last Setup

The original local Claude Code workaround involved:

Emulating Anthropic’s API with a local Flask server
Crafting custom system prompts to replicate the Claude Code persona
Manually managing conversation and code state

ollama-code eliminates that overhead.

Instead of faking endpoints or building your own wrapper, you just:

git clone https://github.com/tcsenpai/ollama-code
cd ollama-code
ollama run codellama:34b

Then start coding with natural language.

It auto-handles:

Role separation (user/system/assistant)
Updating a virtual in-memory codebase
Partial file edits based on your request
Refactors vs. rewrites intelligently

This gives you a more Claude-like dev experience, the LLM behaves more like a smart pair programmer, not a chatbot.

Works With Your Local Setup

If you’re already using Ollama and have a model like codellama:34b, deepseek-coder, or codestral, you’re good to go. No API keys. No Anthropic paywalls.

It supports multi-turn planning and context updates. You can say:

Add error handling to the fetchData function and log failures to a file.

…and it’ll adjust just the lines that matter. No full-file overwrites.

Use Cases

If you’re building a local dev agent or testing workflows, ollama-code is a solid base:

Replace Claude Code in your CLI workflow
Prototype projects incrementally via REPL
Integrate into automations (e.g. Raycast, n8n, Bash scripts)
Fine-tune Claude-style prompts

And since it’s open source, you can tweak whatever you need, context handling, file routing, etc.

What’s Still Missing

It’s early, and there’s room to grow:

No GUI or editor plugin (yet)
Limited to one-file or basic multi-file memory
Performance depends on your local Ollama model

Still, for a tool that removes the glue code between natural language and working software, it’s a leap forward.

Final Thoughts

Running Claude Code locally used to be a hack. Now it’s just ollama-code.

If you’re building AI-assisted dev workflows or exploring local coding agents, this is one of the fastest ways to get started.