What Is an AI Model Router and Should You Use Workweave Router?

What is Workweave Router?

Workweave Router is an ai model router built by Workweave Inc., the company behind the Weave engineering intelligence platform. It acts as a drop-in proxy that speaks the Anthropic Messages, OpenAI Chat Completions, and Gemini native APIs — so tools like Claude Code, Codex, and Cursor can point at it without changing their code.

The router uses a cluster-scoring algorithm based on the Avengers-Pro research paper to decide which model handles each prompt. Per its GitHub description, routing decisions take under 50ms and the project claims 40–70% cost reduction versus sending every request to a single frontier model.

The project is written in Go and was first published on GitHub on April 27, 2026. As of June 28, 2026, it has 470 stars, 23 forks, and 41 open issues. It is licensed under the Elastic License 2.0 (ELv2), which permits self-hosting and modification but restricts using the software to offer a competing hosted service.

How AI model routers save money on coding agents

Coding agents like Claude Code and Codex generate hundreds of API calls per session. Not every call needs a frontier model — a subagent that reads a file doesn't need Claude Opus, but the final code-generation step might. Without a router, developers either overpay by using one expensive model for everything or manually switch models, which breaks flow.

A model router automates this. It inspects each prompt and routes it to an appropriate model based on complexity, cost, and capability. Workweave Router's approach uses an on-box embedder rather than a separate classifier LLM, which keeps latency low. For self-hosted setups, you supply your own API keys for each provider (Anthropic, OpenAI, Gemini, OpenRouter) and the router chooses among them. The hosted version at weaverouter.com bundles provider access and billing.

Who should use Workweave Router?

• Teams already paying for multiple LLM APIs who want to stop manually switching models in their coding tools. If you have Anthropic, OpenAI, and OpenRouter keys, the router picks the right one per prompt.
• Claude Code and Codex users who want to use cheaper models like DeepSeek or Gemini for routine tasks while reserving Claude or GPT for complex work. The one-command installer (npx @workweave/router) wires everything up.
• Developers who want observability into which model handled which request. The router emits OTLP traces viewable in Honeycomb, Datadog, or Grafana.

It may not be a fit for teams that use only one LLM provider and are happy with their current costs, or for production services requiring an OSI-approved open-source license (ELv2 is source-available, not OSI-approved). The project is also two months old — expect rough edges and API changes.

Workweave Router vs OpenRouter vs Portkey vs LiteLLM

Workweave Router enters a market with several established options. Here is how they compare across the factors that matter most when choosing an ai model router for coding agents.

OpenRouter is a commercial service with 400+ models from 70+ providers behind a single API. It charges a 5.5% platform fee on paid usage and offers a free tier with 50 requests per day. It is a fully managed service — you do not self-host it. It is the easiest to start with but adds a per-request markup.

Portkey Gateway (12,223 GitHub stars, MIT license) is a TypeScript-based AI gateway with built-in guardrails, fallbacks, load balancing, and retries. It supports 1,600+ LLMs and offers a free tier of 10,000 logged requests per month. Portkey positions itself as a production-grade control plane rather than a lightweight router for individual developers.

LiteLLM (51,786 GitHub stars) is the most popular open-source option — a Python SDK and proxy server that normalizes 100+ LLM APIs into an OpenAI-compatible format with cost tracking and guardrails. It is the most mature and feature-rich but heavier to self-host than Workweave Router's single Go binary. Its license is listed as "Other" on GitHub.

How to install and set up Workweave Router

The fastest path is the hosted installer, which requires Node.js 18 or later:

1. Run npx @workweave/router in your terminal.
2. The installer asks which tool you use (Claude Code, Codex, or opencode) and whether to configure it per-user or per-project.
3. It creates a router key and patches the tool's config file to point at the router endpoint.

For self-hosting, clone the repository, set your provider API keys in a .env.local file, and run make full-setup. This starts a Postgres database and the router on port 8080 with a dashboard at /ui/. You can then point Claude Code at it with make install-cc or configure Cursor by overriding the OpenAI Base URL in Settings.

To switch the router on and off without uninstalling, use npx @workweave/router off --claude (or --codex, --opencode) and on to re-enable. Claude Code also gets /router-off and /router-on slash commands.

Limitations and what to watch for

The project is young — first commit April 27, 2026 — with 41 open issues as of this writing. The Elastic License 2.0 means you cannot offer a competing hosted service using the code, which may matter for teams building internal platforms that serve multiple teams. Cursor support is described as "early beta" in the README with a note that performance may not be optimal.

The router's cost-savings claim (40–70%) is stated by the project and not independently benchmarked. Results will vary depending on your prompt mix — if most of your prompts genuinely require a frontier model, routing saves less. The self-hosted version requires managing your own provider API keys and a Postgres database.

At a glance

AI model router comparison: Workweave Router vs alternatives (June 2026)

FAQ

Is Workweave Router free?

The self-hosted version is free to run on your own infrastructure — you pay only your provider API costs (BYOK). The hosted version at weaverouter.com offers a free tier to start; paid plan details are not publicly listed on the pricing page as of June 2026.

How does Workweave Router compare to OpenRouter?

Workweave Router is designed specifically for AI coding agents (Claude Code, Codex, Cursor) with per-prompt routing decisions. OpenRouter is a general-purpose model API with 400+ models and a 5.5% platform fee. Workweave Router can be self-hosted with your own keys; OpenRouter is a managed service only.

Is Workweave Router open source?

It is source-available under the Elastic License 2.0 (ELv2). You can view, modify, and self-host the code, but you cannot use it to provide a competing hosted service. It is not under an OSI-approved open-source license like MIT or Apache 2.0.

Related reading

ToolBistro Radar

Sources

• Workweave Router GitHub repository
• Workweave Router official site
• Workweave Router HN launch discussion
• OpenRouter pricing page
• Portkey Gateway GitHub repository
• LiteLLM GitHub repository
• Avengers-Pro routing research paper