**One endpoint. Every model. Always the right one.**
A drop-in proxy for Anthropic, OpenAI, and Gemini that picks the best model
for *every* request: using a tiny on-box embedder, not a vibes-based prompt.
[](https://app.workweave.ai/reports/repository/org_QWsHDcRQWQEs6RpkdEZrlFK8/https%3A%2F%2Fgithub.com/1222789989)
[](go.mod)
[](https://github.com/weave-os/router/actions/workflows/test.yml)
[](LICENSE)
[](https://github.com/weave-os/router/deployments)
[](https://www.producthunt.com/posts/weave-router-2-0?embed=true&utm_source=badge-top-post-badge&utm_medium=badge&utm_campaign=badge-weave-router-2-0)
*Built by [Weave](https://www.workweave.ai): The #1 engineering intelligence platform,
loved by Robinhood, PostHog, Reducto, and hundreds of others.*
---
## What it does
Point Claude Code, Codex, Cursor, or your own app at `localhost:8080`. The router:
- 🎯 **Routes per action.** A cluster scorer derived from
[Avengers-Pro](https://arxiv.org/abs/2508.12631) [^1] picks the right
model from your enabled providers, for every upstream API request.
(See [docs/SEMANTICS.md](docs/SEMANTICS.md) for the canonical terminology:
the router routes per **action**, not per **turn**.)
- 🔌 **Speaks everyone's API.** Anthropic Messages, OpenAI Chat Completions,
Gemini native. Streaming, tools, vision, the works.
- 🧠 **Knows OSS too.** DeepSeek, Kimi, GLM, Qwen, Llama, Mistral via
OpenRouter (or any OpenAI-compatible endpoint).
- 🔒 **BYOK by default.** Provider keys stay on your box, encrypted at rest.
- 📊 **Observable.** OTLP traces out of the box. See them in the Weave dashboard (http://localhost:8080/ui/dashboard) or drop in Honeycomb, Datadog,
Grafana, whatever.
## 30-second quickstart
The fastest way: point Claude Code, Codex, opencode, or pi at the **hosted**
Weave Router with one command. No clone, no Docker, no Postgres.
```bash
npx @weave-os/router
```
That's it. Your browser opens the hosted setup page. To install directly from
the terminal, choose a target explicitly. Other flavors:
```bash
npx @weave-os/router --claude # skip the picker, Claude Code
npx @weave-os/router --codex # skip the picker, OpenAI Codex CLI
npx @weave-os/router --opencode # skip the picker, opencode
npx @weave-os/router --pi # skip the picker, pi + Loom UI
npx @weave-os/router --scope project # per-repo, commits settings.json (or .codex/ / opencode.json)
npx @weave-os/router --local # self-hosted localhost:8080
npx @weave-os/router --base-url https://router.acme.internal
npx @weave-os/router --email you@example.com # set identity without prompting
npx @weave-os/router --claude --return-url https://app.example.com/onboarding # continue after health + key verification
npx @weave-os/router@0.1.0 # pin a version
```
Requires Node ≥ 18 (Claude Code, opencode, and pi paths also need `jq`). Full
flag reference: [install/npm/README.md](install/npm/README.md).
The npm package is published as `@weave-os/router`. The former
`@workweave/router` package remains available as a compatibility alias and
continues to receive the same releases.
### Or: self-host the whole stack
If you want the router (and dashboard) running on your own box:
```bash
# 1. Drop a provider key in. OpenRouter is the recommended baseline.
echo "OPENROUTER_API_KEY=sk-or-v1-..." >> .env.local
# 2. Set a dashboard password. Without one, inference still runs but
# dashboard administration is disabled.
echo "ROUTER_ADMIN_PASSWORD=replace-with-a-strong-password" >> .env.local
# 3. Boot Postgres + router on :8080 and seed an rk_ key.
make full-setup
```
The router is up at , the dashboard at
, and your `rk_...` key prints in the logs.
```bash
# Call it like Anthropic
curl -sS http://localhost:8080/v1/messages \
-H "Authorization: Bearer rk_..." \
-d '{"model":"claude-sonnet-4-5","max_tokens":256,
"messages":[{"role":"user","content":"hi"}]}'
# ...or like OpenAI
curl -sS http://localhost:8080/v1/chat/completions \
-H "Authorization: Bearer rk_..." \
-d '{"model":"gpt-4o-mini",
"messages":[{"role":"user","content":"hi"}]}'
# Peek at the routing decision without proxying
curl -sS http://localhost:8080/v1/route -H "Authorization: Bearer rk_..." -d '...'
```
### What that stack looks like
Only the grey boxes are off your machine. The router, the scorer, Postgres, and
your provider keys all stay local; prompts go from the router straight to the
provider you configured, never to Weave.
```mermaid
flowchart LR
client["Claude Code, Codex, opencode, pi, Cursor, your own app"]
router["Router :8080 /v1/messages · /v1/chat/completions /v1beta/models · /v1/route"]
scorer["Cluster scorer in-process ONNX embedder"]
hmm["HMM policy sidecar :8093 optional, make up-hmm"]
pg[("Postgres installations, rk_ keys, encrypted BYOK keys, usage")]
ui["Dashboard /ui selfhosted mode only"]
providers["Anthropic · OpenAI · Gemini OpenRouter and any OpenAI-compatible endpoint"]
otel["Your OTLP collector Honeycomb, Datadog, Grafana"]
client -->|"rk_… bearer token, streamed response back"| router
router -->|"embed and score the action"| scorer
router -.->|"ROUTER_DEFAULT_STRATEGY=hmm"| hmm
router -->|"auth, config, usage"| pg
pg --> ui
router -->|"provider key from env or BYOK"| providers
router -.->|"spans and usage logs"| otel
classDef external fill:#f4f4f5,stroke:#a1a1aa,color:#3f3f46
class providers,otel external
```
Multi-replica deployments also need Pub/Sub (`PUBSUB_*`) for cache
invalidation; `docker compose` runs the emulator for you.
### Optional: self-host the frozen HMM policy
The default stack uses the in-process cluster scorer. To run the frozen HMM
policy as a companion container, add a Google API key and use the opt-in target:
```bash
echo 'GOOGLE_API_KEY=...' >> .env.local
make up-hmm
```
This does not change the default strategy. See
[`sidecars/hmm/README.md`](sidecars/hmm/README.md) for artifact verification,
embedding compatibility, and explicit HMM selection.
## Wire it into your tools
**Claude Code.** Run `make install-cc` to wire Claude Code at the local
self-hosted router (it's also invoked automatically at the end of
`make full-setup`). For the hosted router, use `npx @weave-os/router`
above.
**Codex** (OpenAI CLI). `npx @weave-os/router --codex` patches
`~/.codex/config.toml` (or `/.codex/config.toml` with `--scope project`)
with a managed `[model_providers.weave]` block and sets `model_provider = "weave"`.
The provider preserves Codex's existing ChatGPT OAuth login while the router
key rides in an `X-Weave-Router-Key` HTTP header and the installer selects the
HMM strategy for the public hosted endpoint. `--codex --local` and custom
self-hosted URLs keep their router's configured default because the HMM
sidecar is optional. HMM and forced selections in the native Codex family
(`gpt-5.6-sol`, `gpt-5.6-terra`, `gpt-5.6-luna`) use that OAuth credential;
every other selected model uses its WorkWeave deployment or BYOK credential,
matching the Claude Code plugin's model-to-credential dispatch.
Codex does not load third-party slash-command files, so the installer ships
`$force-model ` (alias `$fm `), `$unforce-model` (alias
`$ufm`), and `$router-feedback ` (alias `$rf `) as native Codex
skills. Their invocations are handled by the router directly. Re-install and
`--uninstall --codex` rewrite/remove only the managed block, leaving the rest
of your Codex config untouched. Codex also gets `$router-status`,
`$router-off`, `$router-on`, and `$router-models` as skills that call this
installer's own verbs. Invoke `$disable-routing` (or `$router-off`) to switch the
next Codex session back to its normal provider, or run
`npx @weave-os/router disable-routing` in a shell; a literal
`/disable-routing` is not a third-party extension point in Codex.
**opencode.** `npx @weave-os/router --opencode` requires OpenCode major version
2 when the CLI is installed and available on `PATH`. It merges a
`provider.weave` entry into `~/.config/opencode/opencode.json` (or
`/opencode.json` with `--scope project`) and makes `weave/auto` the active
model while saving any prior default for off/uninstall. OpenCode v2 sends
Responses requests through its bundled `@ai-sdk/openai` provider; the router
selects and translates to the upstream model. The router key, identity, and
session headers ride alongside the provider config. No plugin is installed;
re-install removes stale v1 plugin registrations/files, and
`--uninstall --opencode` removes the provider and restores the prior model.
**pi.** `npx @weave-os/router --pi` keeps stock pi as the runtime and installs
the router's pi extension. It adds the Loom header, Wooly's animated terminal
mascot, a persistent `WEAVE ROUTER` route/savings line, `/fm` + `/ufm`
model-pin commands with a `[forced]` status, and context-isolated subagents
without shipping or maintaining a forked pi binary.
**Cursor** *(early beta, performance may not be the best).* Settings →
Models → *Override OpenAI Base URL* → `http://localhost:8080/v1`, paste
`rk_...` as the API key.
**Switching on/off.** After installing, `npx @weave-os/router off --claude`
(or `--codex` / `--opencode`) routes that client straight to its provider
again without discarding the router config; `on` flips it back, and `status`
reports which way it's pointing. Claude Code also gets `/router-off`,
`/router-on`, and `/router-status` slash commands. Cursor toggles via the same
Settings → Models override above. See [install/README.md](install/README.md#switching-on-and-off).
**Choosing which models the router may pick.** `npx @weave-os/router models
--claude` lists every deployed model with its on/off state, and `models enable`
/ `models disable` change it — the same setting as the dashboard's settings
page, edited from the terminal. Claude Code gets this as `/router-models`
(alias `/models`). Requires a router that serves the model-selection API;
against the Weave-hosted router the list still prints and points you at the
dashboard, where selection is an organization-wide setting. See
[install/README.md](install/README.md#choosing-which-models-the-router-may-pick).
> Two keys, don't mix them up:
> - `sk-or-...` / `sk-ant-...` / `sk-...` = your **upstream** provider key. Lives in `.env.local`.
> - `rk_...` = your **router** key. Clients send this as a Bearer token.
## Endpoints
| Endpoint | Format |
| ------------------------------ | ---------------------------------------- |
| `POST /v1/messages` | Anthropic Messages, routed |
| `POST /v1/chat/completions` | OpenAI Chat Completions, routed |
| `POST /v1beta/models/:action` | Gemini `generateContent`, routed |
| `POST /v1/route` | Returns the decision, no upstream call |
| `GET /v1/models` · `POST /v1/messages/count_tokens` | Anthropic passthrough |
| `GET /livez` · `GET /startupz` | process responsive check + completed initialization |
| `GET /health` · `GET /readyz` · `GET /validate` | responsive-check alias + dependency diagnostics + key check |
| `POST /v1/client-events` | Harness CLI's `off`/`on`/`uninstall` report (`harness` = `claude_code`/`codex`/`opencode`/`pi`); logged + exported as `router.harness_lifecycle`, nothing stored |
| `GET /v1/sessions/:session_id/cost` | One session's committed cost + savings, scoped to your key's installation; `rk_` or `ra_` key ([docs](docs/SESSION_COST.md)) |
| `GET /v1/analytics/routing-decisions` | Raw routing decisions as cursor-paginated NDJSON ([docs](docs/ANALYTICS_EXPORT.md)) |
| `GET /v1/analytics/schema` · `GET /v1/analytics/models` | Export field dictionary + price book |
Use `/startupz` for Cloud Run startup and `/livez` for its process responsive
check. The listener starts after artifact/client initialization and bounded
connection preparation. `/startupz` also pings PostgreSQL with a two-second
budget: failure returns 503, so Cloud Run withholds traffic and retries while
the process stays running. Provider preparation failures only log warnings.
`/livez` does not contact the database. Startup sends no prompts and performs no
embedding or HMM inference; it cannot guarantee the latency of first inference.
`/readyz` retains dependency diagnostics.
Routed non-stream responses include `x-router-cost-usd`,
`x-router-cost-input-usd`, `x-router-cost-output-usd`,
`x-router-cache-read-tokens`, and `x-router-cache-creation-tokens`. Streaming
responses cannot carry the final cost in HTTP headers because headers flush
before usage is known; stream clients should read the `weave_cost` object on
the final Anthropic `message_delta` usage event.
## Deeper docs
- 📐 [**Configuration reference**](docs/CONFIGURATION.md): every env var,
BYOK encryption, OTel knobs, cluster routing.
- 🧭 [**Semantics and terminology**](docs/SEMANTICS.md): canonical definitions
for session, round, turn, action, and step.
- 📊 [**Analytics export**](docs/ANALYTICS_EXPORT.md): pulling raw routing
decisions into your own warehouse with a read-only key.
- [**Policy router harness**](docs/POLICY_ROUTER_HARNESS.md): contract and
rollout checklist for adding an out-of-process policy model.
- 📏 [**Codex benchmark harness**](bench/README.md): reproduce the published
SWE-Atlas QnA and Terminal-Bench 4.0 comparisons of
`/beta` routing vs. direct and OpenRouter controls.
- 🛠️ [**Contributing**](CONTRIBUTING.md): layering rules, hot-reload dev,
migrations, tests, the whole engineering loop.
- 🏗️ [**Architecture**](AGENTS.md): package layout, import contracts,
recipes for adding endpoints / providers / strategies.
## License
Licensed under the [Apache License 2.0](LICENSE). Copyright 2025-2026 Workweave, Inc.
See [NOTICE](NOTICE) for attribution.
The Apache-2.0 release line begins with `router-v0.2.24` (npm packages
`@weave-os/router` and `@workweave/router` version `0.2.24`). Earlier tags and
published packages retain the licenses shipped with them; this change does
not rewrite historical releases.
Apache-2.0 permits commercial use, hosted services, and proprietary derivatives,
subject to its terms, and includes a patent grant. It does not grant general
rights to Workweave's trademarks. Third-party code, dependencies, model weights,
and datasets retain their own licenses; see [THIRD_PARTY.md](THIRD_PARTY.md).
---
[^1]: Zhang, Y. et al. *Beyond GPT-5: Making LLMs Cheaper and Better via
Performance–Efficiency Optimized Routing* (Avengers-Pro).
arXiv:2508.12631, 2025.