A tool to get Claude Code-style reliability from fully local models
22d 3h ago by lemmy.world/u/aclarke in Ollama
I've been experimenting with local models via Ollama for a long time now, but one thing kept frustrating me:
Smaller local models are actually a LOT more capable than people think, but they struggle with reliability.
They lose focus, drift over long sessions, stop halfway through problems, and aren't great at breaking apart larger tasks.
So, to try and fix that, I built Coyote.
Loki is a local-first command line tool and runtime for building and running LLM workflows locally. It's model agnostic, works well with Ollama, and includes things like:
- Agents and delegation
- Workflow scripting
- MCP servers
- RAG
- Roles/personas
- Custom tools
- Macros
A lot of the architecture is specifically designed around improving the reliability of smaller models. So for example:
- Auto-continuation to push models towards completion and prevent them from stopping halfway through a problem
- Parallel agent delegation to reduce context overload and keep scope small
- Workflow-based execution for more deterministic automations
- Scoped agents to reduce context drift
My long term goal with this project is basically:
Get as close as possible to Claude Code-style reliability using fully local models.
I'd love feedback from other Ollama users experimenting in this space and any tips/pointers/ideas of other things to add to make it function even better!
Very neat, I'll check this out. Also have had the same issues.
You'll never get uploads here btw, check out some of the communities over on db0, it's an AI friendly instance
You mean upvotes? Yeah was wondering why the downvotes
because people are sick of ai, and corporate slop
if I keep screaming and forcing string beans in your face constantly, you're gonna grow to hate string beans... except string beans have proven to be useful unlike ai
I get that but isn’t Ollama about ai?