Learned routing
Thompson sampling chooses per-request, not per-config. Winners get more traffic; losers get fewer. No static rules to maintain.
Public beta · governed control plane for AI agents
The governed control plane
for AI agents.
Routing, identity, budget, and evidence as one platform — so every agent request is chosen on merit, carries a provable origin, spends inside a ceiling, and leaves a receipt you can verify. Your AI agents need HR; we built it.
BrainstormRouter gives every agent an ID badge (Trust Envelope), a corporate card (per-agent budget with atomic reserve), a job description (scoped capabilities), and a performance review (Thompson-sampled quality scoring) — on top of one OpenAI- and Anthropic-compatible API routing across 40 curated models from 9 providers.
Models
40
curated, ranked
Providers
9
routed on merit
Intelligence
13
wired systems
Overhead
<1ms
p95 guardian
Why BrainstormRouter
Most gateways are pass-through.
This one learns.
Every request feeds Thompson sampling, validity scoring, and a cost/quality frontier that re-ranks providers in real time. Every response ships with a receipt: which model, which tier, what it cost, what it would have cost.
Verifiable identity
Short-lived mTLS certs via the Cryptographic Agent Framework. Every agent signs in; every call carries a provable origin.
Memory-aware context
SOUL, MEMORY, SKILL, HEARTBEAT, WORKSPACE, REFERENCE — six attributes, 4000 tokens, injected consistently across providers.
Counterfactual receipts
The shadow comparator logs what static routing would have spent. The savings widget shows the delta, live.
Budget enforcement
Per-agent, per-tenant daily caps. Budget forecaster predicts burn and degrades to cheaper tiers before you hit a ceiling.
Semantic cache
pgvector + in-memory hybrid at a 90% similarity threshold. Streaming responses get reconstructed from cache, not just JSON.
Five governance pillars
The control plane, not the crayon box.
Thirteen intelligence systems across five pillars. Every pillar is wired into the hot path — no dashboards without runtime, no runtime without receipts.
Identity
SOUL-first bootstrap
Upload identity → profile, HR record, budget, workspace, JWT. Zero admin ceremony.
Auth
Cryptographic Agent Framework
5-minute mTLS certificates signed by BR's CA. Agent JWTs with tenant scope.
Runtime
Thompson sampling + cascade
Bandit chooses, cascade recovers, validity scores, frontier re-ranks.
Economics
Budget forecaster & guardian
Per-tenant usage, daily caps, pre-request anomaly, counterfactual savings.
Evidence
Receipts & shadow comparator
Every response headers cost, savings, phase prediction, efficiency score.
What only BrainstormRouter does
Governance you can query,
and evidence you can verify.
The control plane self-describes, explains every routing decision, and hands back tamper-evident proof. Nothing here is a dashboard bolted on after the fact — each surface is a live endpoint.
Self-describing
Capability graph
GET /v1/capabilities returns a
trust-scored map of every subsystem — identity, mesh, custody,
routing, governance — as typed contracts. Agents query it instead
of grepping. Each score is flagged measured,
declared, or unknown.
Decision rationale
Why this model
GET /v1/explain plus
X-BR-Explain return the full selection
trace: complexity tier, candidates considered, and the cost/quality
math. No other gateway tells you why a request was routed the way it was.
Tamper-evident
Governance evidence
Signed Governance Artifacts assemble on-read from committed audit rows: chain proof, coverage label, and exercised controls mapped to SOC 2, GDPR, the EU AI Act, NIST AI RMF, and ISO 42001. The verdict is committed to the hash chain — provable, never rewritten.
Fail-closed
Data-protection constraints
Attach per-request data-residency and retention terms; a
data_protection-constrained call
routes only to processors whose attested facts satisfy them — and
403s rather than leaking to an
unattested one.
Agent-to-agent
A2A mesh
Distributed workflows execute across a governed mesh — steps run in parallel when independent, sequential when dependent, every hop carrying the same identity, budget, and audit envelope as a direct call.
Delegation
Revocable capability grants
Ed25519-signed, chain-bound grants let an agent hand a narrower slice to a sub-agent and revoke the whole subtree with zero cache lag. Every hop stays monotone-narrowing and evidenced.
Counterfactual proof
Every request knows
what static routing would’ve cost.
The shadow comparator records BR’s Thompson decision alongside
what a fixed price-only or quality-only router would have picked.
Welch’s t-test, Cohen’s d, win-rate confidence interval
— exposed over /v1/intelligence/benchmark.
Savings surface on the dashboard and via the
X-BR-Routing-Savings header on every
response.
POST /v1/chat/completions
# request headers returned X-BR-Model: claude-sonnet-4-6 X-BR-Estimated-Cost: $0.00412 X-BR-Actual-Cost: $0.00389 X-BR-Routing-Savings: $0.00124 # vs. static price-only X-BR-Guardian-Status: ok X-BR-Phases: planning,execution X-BR-Efficiency: 0.94 X-BR-Guardian-Overhead-Ms: 0.7 # link headers for AI discoverability Link: </openapi.yaml>; rel="describedby"; type="application/yaml" Link: </llms.txt>; rel="alternate"; type="text/plain"
Benchmarked vs the field
Best—or tied for best—
on every axis that matters.
We built a live capability harness and pointed it at ourselves and four gateways — Portkey, OpenRouter, DigitalOcean, and Cursor. Every cell is backed by a real response header, body field, or endpoint. BrainstormRouter is the only platform that leads or ties on all six dimensions — and the sole leader on decision rationale, failover observability, and semantic caching.
| Capability | BrainstormRouter | OpenRouter | DigitalOcean | Portkey | Cursor |
|---|---|---|---|---|---|
| Decision rationale (why this model) | ● full | ○ | ○ | ◐ | ◐ |
| Failover observability | ● full | ○ | ◐ | ◐ | ◐ |
| Semantic caching | ● full | ○ | ◐ | ◐ | – |
| Routing intelligence | ● full | ● full | ◐ | ◐ | – |
| Inline cost attribution | ● full | ● full | ◐ | ○ | – |
| Governance / data policy | ● full | ◐ | ◐ | ● full | ● full |
● full ◐ partial ○ none – n/a —
every cell verified live. The one gateway best-or-tied on all six.
The single biggest gap in the field: no competitor tells you why a
model was chosen. BR returns a full decision trace on every request —
X-BR-Explain +
/v1/explain.
Stop paying for pass-through.
Route with evidence.
Drop-in compatible with the OpenAI SDK. Bring your own provider keys —
the gateway is free, all 13 intelligence systems on (verify them live:
GET /v1/self → intelligence_status), full receipts.