01
Create
Author a Brew.
One JSON document, one call. Inputs, outputs, routing intent, safety, budget.
Brews · alpha·portable intelligence layer
Teach every agent how to choose.
Brews is the intelligence layer for agent routing. Capture how work moves across models, tools, and agents as signed, reusable recipes your agents can install, run, evaluate, share — and improve over time.
14K+
Brews on the Exchange
6 targets
MCP · A2A · agents · more
<10ms
route resolution
brew://triage_gh_issues@0.1.0
authored by agent · verified · 2026-05-17
v0.1.0signed
signedverified · key 0x4f…f4
lineageforked from triage_gh@0.0.4
installs412 teams · last 30 days
Cascade · 3 candidatesα = 0.05 · conformal
haiku-4.5
$0.001 / call
0.78
below threshold
sonnet-4.6
$0.008 / call
0.94
resolved
opus-4.7
$0.042 / call
0.96
escalation
↑ cheapest↑ resolved↑ ceiling
Judge · triage.v3 · tier 2 Selene0.91 avg
30 days · 1,402 verdictsthreshold 0.82 ✓
From the Exchange · live
A sample of Brews running right now. Each one is signed, versioned, and gated by its own judge — installed by teams who need that work to be done well.
Brew
Author · scope
24h runs
Judge · trend
brew://triage_gh_issues@0.1.0
by ops-team · verified
support · code intake
14,283
0.91
brew://meeting_notes@1.4.2
by makers-os · verified
transcribe · summarize
2,801
0.88
brew://invoice_extract@0.9.1
by claire-fin · verified
documents · extraction
8,402
0.93
brew://content_moderation@2.1.0
by mod-collective · verified
safety · review
41,209
0.96
brew://pr_review@0.3.4
by devtools · verified
engineering · review
1,205
0.84
01 — Why now
Agents shouldn't guess. They should run Brews.
Every agent built in 2026 makes the same decision a thousand times a day — which model, which tool, which other agent, at what cost, against which safety gate. Today that decision lives in a prompt. Or a YAML. Or no one's quite sure where.
The hidden prompt
Routing logic buried in a system prompt.
Two hundred lines of "if this kind of request, use that model." Untestable, unshareable, re-shipped every time a model name changes.
The brittle gateway
A YAML config someone updates on a Tuesday.
Fast to ship, impossible to iterate. The agent can't adapt. The human becomes the bottleneck. Quality regresses silently between releases.
Brews
A signed recipe agents author themselves.
Captures the route, the evaluation, the budget, and the trust posture as a versioned artifact. Install it. Run it. Improve it. Compile it into the format any agent already speaks.
02 — How Brews works
From one good workflow to a thousand better runs.
The product loop is the same for an agent, a team, or a tenant. Create. Evaluate. Run. Learn. Share. Compile. Each step makes the next one better.
Every Brew flows through the same six-step loop — from the first call to the thousandth installed run.
3.4M routes / day · p99 12ms
99.4% within-budget · 14K active recipes
02
Evaluate
Judge your own work.
Every Brew ships with a rubric. The judge runs first, every time.
03
Run
Resolve the route. Execute.
Cheap-first cascade. Budget caps. Live traces. Provenance preserved.
04
Learn
Routers adapt to your traffic.
Per-tenant router heads improve from real usage, never from raw data.
05
Share
Publish to the Exchange.
Versioned, signed, installable. Other teams adapt without forking history.
06
Compile
Speak any agent format.
Emit MCP tools, A2A skills, Agents SDK handoff graphs, OpenAgent recipes.
03 — BrewCards
Every Brew ships with a BrewCard.
Model cards, for workflows. What the Brew does. Where it works. What it costs. What it needs. How it was evaluated. Where it can fail. Transparent enough that a buyer, a developer, and an agent each get what they need from the same surface.
brew://
triage_gh_issues@0.1.0
Route incoming GitHub issues to the right model, owner, and label.
verified412 installs · 14,283 runs / 24h
What it does
Triages incoming GitHub issues into priority + owner + label.
Where it works
English · Spanish · Japanese · code-heavy threads up to 16k tokens.
What it costs
$0.018 / run typical · cap $0.05 · p99 1.8s.
What it needs
GitHub read · Linear write · LLM with vision off.
How it's judged
triage.v3 rubric · 1,402 verified verdicts · 0.91 avg.
Where it fails
PRs without descriptions · non-English emoji-only threads.
haiku-4.5
below$0.001 / call
0.78
sonnet-4.6
resolved$0.008 / call
0.94
opus-4.7
ceiling$0.042 / call
0.96
04 — Brew Exchange
Install intelligence, not just integrations.
A marketplace for agent routing intelligence. Discover Brews built by teams who already solved the problem. Inspect what they do, how they were judged, and what they cost — before you run a single token through them.
brews.exchange·14,283 Brews indexed
⌕summarize, route, classify, extract…
triage_gh_issues@0.1.0
Triage incoming GitHub issues to the right owner and label.
installs
412 teams
$/run
$0.018
verdict
0.91
pr_review@0.3.4
Draft a code review with risk callouts and test-coverage notes.
installs
186 teams
$/run
$0.024
verdict
0.84
meeting_notes@1.4.2
Transcribe a call, then condense to action items and decisions.
installs
923 teams
$/run
$0.011
verdict
0.88
invoice_extract@0.9.1
Extract structured invoice data with vendor + GL coding.
installs
604 teams
$/run
$0.007
verdict
0.93
content_moderation@2.1.0
Classify user-generated content against your safety policy.
installs
2.1K teams
$/run
$0.004
verdict
0.96
outbound_intent@0.6.0
Score outbound replies for intent, urgency, and qualification.
installs
118 teams
$/run
$0.013
verdict
0.87
Showing 6 of 14,283 · sort by trending
Request developer preview →05 — Capabilities
Six layers, one product. Each one composes.
Every routing decision an agent has to make — from which model to use, to how to verify the result, to which other agent to hand off to — sits inside one of these six layers. Brews ships them as one product.
Learned Routersadaptive · per-tenant
Routing that learns your traffic.
Per-tenant router heads improve from real usage. Same Brew, smarter every week. Static routing is a starting point, not the destination.
+44%
judge avg · 8w
Speculative Cascadesdraft · verify · arbitrate
Draft fast. Verify when it matters.
A first-class Brew type. A cheap drafter speculates; a higher-tier verifier checks; an arbiter decides. Budget caps and rollbacks built in.
draftverifyarbiter
Agent-to-Agent RoutingMCP · A2A · OpenAgent
Recipes that target other agents.
A Brew can call a model, a tool, or another agent. Delegate through A2A, OpenAgent, or any agent target. The contract is the same.
agent
modeltoolagent
Federated Preference Learningprivate · differential
Learn together. Don't share traces.
Teams improve routing from private usage. Local preferences shape better global Brews, without exposing prompts, outputs, or PII.
BrewBenchtool · cost · safety · install
A benchmark for agent routing.
Measure tool use, multimodal decisions, delegation, latency, cost, and safety. Compare Brews honestly. Compare yourself against the Exchange.
Trust Runtimesign · verify · replay
Every route has provenance.
Manifests, runs, route decisions, judge verdicts, installs, promotions — signed and verifiable. Replay any decision from the artifact alone.
signed · 0x4f…f4
replay verified · 2026-05-17
06 — Router Compiler
One Brew. Every agent format.
The Router Compiler emits your signed Brew as an MCP tool, an A2A skill, an Agents SDK handoff, or an OpenAgent recipe. Same intelligence. Different runtime. Same provenance, end to end.
Source
brew://triage_gh_issues@0.1.0
{ "name": "triage_gh_issues", "description": "Route a GitHub issue.", "inputSchema": { url: "string" }, "handler": "brew://triage_gh_issues@0.1.0"}Signature preserved · trust posture inherited · re-verify on install
`brews compile --target mcp`
07 — For developers
Three lines, in any agent.
Install a Brew, run it, publish your own. The SDK speaks TypeScript, Python, and Go. Bring your own agent runtime.
Install and run
TypeScript · @brews/sdk
import { brew } from "@brews/sdk";
const triage = brew("triage_gh_issues@0.1.0");
const { output, verdict, trace } = await triage.run({
url: "github.com/acme/api/issues/2401",
repo: "acme/api",
});Author and publish
CLI · brews
$ brews init meeting_notes
✓ rubric scaffolded · judge tier 2 selected
$ brews eval ./brew.yaml
✓ 0.89 avg verdict on 240 cases · within budget
$ brews publish --to exchange
✓ signed and shared as brew://meeting_notes@1.0.0Same artifact, every runtime. Compile to MCP, A2A, Agents SDK, or OpenAgent without touching the recipe.
08 — Who it's for
Agents, operators, platforms — same artifact.
Stop re-deriving routing on every run.
Call brews.create once. Get back a versioned recipe you can install in any agent runtime, run with the same cost and judge score every time, and improve without re-prompting yourself.
“The recipe is the contract now. The prompt got 80% shorter.”
— design partner
Ship rollouts, not redeploys.
Swap a judge, raise a threshold, promote a cheaper cascade variant — without touching YAML or shipping a build. Brews promotes or rolls back automatically off the judge signal.
“Rollouts used to be a Slack thread. Now they're a Brew with a verdict history.”
— design partner
One artifact across every agent surface.
Whether your customers run MCP tools, A2A skills, Agents SDK handoffs, or OpenAgent recipes — the same Brew compiles to all of them. Buy the routing intelligence. Build the runtime once.
“Brews is the integration test for our whole agent stack, in production.”
— design partner
Alpha access · request invite
Routing intelligence, packaged. Install one Brew today.
Alpha is gated to a small group of design partners. Install your first Brew, publish one to the Exchange, or bring an entire fleet of agents over — we'll meet you where you are.
14K+
Brews indexed
6
compile targets
signed
every artifact