Show HN: Smart model routing directly in Claude, Codex and Cursor

SAN FRANCISCO WorkWeave used Hacker News to pitch a router that sits between coding agents and model APIs, then chooses which model handles each request. The company says the goal is to keep frontier models for hard work and shift routine turns to cheaper ones.

What did WorkWeave announce?

The June 26, 2026 Show HN post said the router plugs into Claude Code, Codex, Cursor, and other agent clients, then sends each inference request to the best model for the task. The post also said the system is trained on tens of thousands of agent traces and saved 40 percent on tokens internally, without a visible drop in quality or velocity.

The submission was filed by adchurch, drew 216 points, and generated 113 comments. The launch post said the router was source available under Elastic License 2.0 and could be self hosted or used through a hosted version.

Source	What it said	Why it matters
Show HN post	40 percent token savings, no visible quality loss, RL routing on tens of thousands of traces	These are the headline product claims at launch
HN metadata	June 26, 2026, 216 points, 113 comments	Shows the post had immediate attention
License commit	Elastic License 2.0, with a ban on hosted or managed service use	Sets the legal boundary for self hosting

How does the router work in practice?

The repository now says the router makes a decision per action, meaning one upstream API request at a time. It speaks Anthropic Messages, OpenAI Chat Completions, and Gemini, and the docs say provider keys stay on the user's machine.

The current docs also frame the router as cache aware. Once it starts using one model, the threshold to switch rises because the system weighs the savings from changing models against the cost of a cache miss.

That is the central operational promise. The router is trying to preserve the fast path for simple work, while reserving costly frontier models for planning, rescue steps, or harder edits.

What did critics push back on?

Critics focused on cache misses, agent awareness, and training freshness. One common objection was that coding agents already know when they are exploring, planning, or implementing, so an extra proxy could break the control loop and burn more money through cache churn.

Another challenge was durability. Commenters asked how a router trained on one organization's traces would keep up with new models and new task patterns, and they asked for public benchmarks on tools such as Terminal-Bench or SWE-Bench style tests.

The poster replied that the router is cache aware, so the system raises the switching bar when a model change would be expensive. The response also said the team would keep growing the dataset and use production reward signals, then publish evals later.

What do the license and privacy terms mean?

The launch license is permissive for self hosting, but it is not an open field. The Elastic License 2.0 text in the repository commit says users may not provide the software as a hosted or managed service, and they may not move, change, disable, or circumvent the license key functionality.

Privacy is better described as local control with clear edges. The current configuration docs say prompts go from the router straight to the chosen provider, provider keys remain on the operator's box, and OTLP tracing is opt in, while the optional HMM sidecar can call an external embedding provider at startup to verify the package's embedding space.

That means the default scorer path is local, but the optional HMM path can depend on an external embedding service, such as Google or an OpenAI compatible endpoint, during startup checks. The sidecar README says those assets are verified before readiness and are not fetched on a live request.

What should readers watch next?

The launch post made a clear business claim, lower spend without a quality hit, but the public record still lacks an independent benchmark set from WorkWeave itself. The repo and docs are stronger on implementation detail than on third party validation.

That leaves the next test simple: whether the router can keep saving money once it faces more users, more models, and more codebases. The launch showed demand, but the market will want public evals before it trusts the savings number.