gremlord
Claude Code on any model, with a budget
7 followers
Claude Code on any model, with a budget
7 followers
gremlord wraps Claude Code in a thin local router. Same TUI, same tools, but the model behind it can be Anthropic, OpenAI, xAI, Ollama, or any OpenAI-compatible endpoint. Every token is priced, and hard budgets stop the spend. Free, MIT licensed.
Free
Launch Team
Framer AI AgentsDesign and publish professional sites with AI
Promoted
Thanks! The animation is made with Charmbracelet VHS. The recipe and synthetic demo data are checked in as `assets/hero.tape` and `assets/hero-demo-home.sh`, so it is reproducible.
On the model/harness boundary: I do not try to infer it from normal sessions. `gremlord eval` runs the baseline and test model through the same non-interactive Claude Code harness, on the same coding task and repo, with an isolated home and no user/project settings, MCP servers, subagents, or web tools. It captures the full patch plus verifier evidence; the optional judge sees blinded outputs and never sees model names, cost, or run order. That holds the harness constant and changes the model.
For prompt visibility, the router receives Claude Code's Messages request at the API boundary. Anthropic traffic is passed through; other providers get protocol translation, not a different agent loop or hand-written replacement prompt. `gremlord context` also breaks the request down into system prompt, tool schemas, and conversation, which made this much less mysterious for me.
That animation is freaking great. How did you make it?
As for gremlord: I'm interested to know how you managed to measure the differences of different models on the same harness. The claim that the harness is great but you wanted to try the different models, makes me believe you have a clear distinction on where the harness and the model begins.
I know in theory how this works but in practice because of very little visibility I don't *really" know what goes into the prompt.
This is great, since my company started enforcing strict cost limit it makes it hard to correlate that to actual tasks and optimizing cost (and not token usage).
It is really a struggle as every AI company have their own fake currency conversion ratio to worry about. Now I can make conscious decisions about the tradeoffs of cost vs quality on a restricted resource (real world budget)