No reviews yetBe the first to leave a review for Make Claude Code cheaper and smarter
Maker
📌
DedrooM
Cuts your AI coding agent's token bill without changing how you work.
Why
Two things quietly inflate an agent session's token spend: the agent retrying a failing command in a near-identical loop, and tool output (file listings, logs, diffs) piling into context that the model has effectively already seen. Neither shows up as a single big line item they show up as your bill being higher than the work should have cost.
DedrooM catches both automatically. One command wraps your existing agent (dedroom wrap claude), and from then on repeated failing calls get cut off before they compound, and redundant tool output gets compressed before it reaches the model — no changes to how you invoke your agent, no new workflow to learn.
Report
how does it actually decide when to mutate vs just retry, and does that logic adapt over time or stay fixed?
The mutation decision is based on a combination of:
Repeat count (adaptive threshold that tightens on errors)
Recent trajectory (oscillation detection)
Error signals from the tool response
When we decide to mutate, we generate 3–4 strategy variations (parameter tweak, tool swap, decomposition, rephrase) and pick the most promising one using a fast heuristic scorer (with optional local LLM judgment).
The logic does adapt over time through our learning system successful mutations are remembered per project, so similar situations get better strategies in future sessions.
Happy to share more details if you're curious!
Report
A dashboard is great, but consider exposing the mutation rules as configurable so users can tweak the retry and strategy logic to fit their own workflows instead of relying solely on the defaults.
Configurable mutation rules is high on our roadmap. Right now we ship with balanced defaults that work well for most users, but we plan to add a dedroom rules command + config file so power users can define custom strategies and priorities.
Would love to hear what kind of rules you'd want to configure.
Report
one thing i'd love here is a way to mark certain tool calls as "expected to be expensive" so the proxy knows not to compress them. sometimes i actually need the full output for debugging and it's annoying when it gets trimmed down before i even see it.
We're adding a "protect" list in the next update you'll be able to mark certain tools or patterns as "do not compress" so you always get full output when you need it for debugging.
For now, you can use dedroom doctor to see what's being compressed and adjust via config if needed.
how does it actually decide when to mutate vs just retry, and does that logic adapt over time or stay fixed?
@erdiakkoc Great question!
The mutation decision is based on a combination of:
Repeat count (adaptive threshold that tightens on errors)
Recent trajectory (oscillation detection)
Error signals from the tool response
When we decide to mutate, we generate 3–4 strategy variations (parameter tweak, tool swap, decomposition, rephrase) and pick the most promising one using a fast heuristic scorer (with optional local LLM judgment).
The logic does adapt over time through our learning system successful mutations are remembered per project, so similar situations get better strategies in future sessions.
Happy to share more details if you're curious!
A dashboard is great, but consider exposing the mutation rules as configurable so users can tweak the retry and strategy logic to fit their own workflows instead of relying solely on the defaults.
@berke746137 Excellent suggestion thank you.
Configurable mutation rules is high on our roadmap. Right now we ship with balanced defaults that work well for most users, but we plan to add a dedroom rules command + config file so power users can define custom strategies and priorities.
Would love to hear what kind of rules you'd want to configure.
one thing i'd love here is a way to mark certain tool calls as "expected to be expensive" so the proxy knows not to compress them. sometimes i actually need the full output for debugging and it's annoying when it gets trimmed down before i even see it.
@yarenk8za Totally valid point.
We're adding a "protect" list in the next update you'll be able to mark certain tools or patterns as "do not compress" so you always get full output when you need it for debugging.
For now, you can use dedroom doctor to see what's being compressed and adjust via config if needed.
Thanks for the feedback!