Agent GovernanceGovern your AI agents in real time.
A standalone product: license it on its own, pay per governed agent, and onboard agents through the protocols they already speak — the latest MCP and Google A2A. Authority sized to the task, enforcement that sits outside the model, and a record you can prove.
From pricing page to a governed fleet — six steps.
The whole flow runs on infrastructure that already existed; Agent Governance plugs into it as its own product.
Discover
Appears on /pricing as “€25 per governed agent.”
Request
One click files a licence request.
Approve
A super-admin grants — only this feature.
Pay
Top up the credit wallet via Stripe.
Onboard
Agents self-register via MCP / A2A.
Govern
Per-action control on everything they do.
Three ways in — none of them manual data entry.
You never hand-enter thousands of agents. They arrive as they come online, or you pre-register a known fleet in one call.
Auto-provision
An agent connects over MCP and signs in — its identity is created on first contact.
identity = client + subject
Ingest a card
Post an Agent Card — name, url, provider, version — and it maps to an identity, grouped into a provider-version cohort.
cohort = provider : version
Bulk register
One call registers a whole cohort — over REST or the onboarding tool — so you never hand-enter agents one by one.
POST /gateway/onboardSpeaks what agents speak — 2026 editions.
Onboarding reads a native descriptor from each protocol to establish identity and cohort — no bespoke adapter.
MCP · 2026-07-28
OAuth 2.1 resource server: audience-bound tokens (RFC 8707), protected-resource metadata (RFC 9728), issuer validation (RFC 9207).
Google A2A · 1.0
A signed Agent Card at /.well-known/agent-card.json, bound to the provider’s domain, establishes who the agent is.
Google AP2 · v0.2
Rides A2A as an extension. Intent, Cart and Payment mandates bind at the moment of a transaction.
Plugs into the identity stack you already run.
Agents are non-human identities. Governing them shouldn’t mean a parallel directory, a bespoke token, or a second source of truth.
RFC 8693 · 7662
Exchange a token your identity provider already issues (Entra, Okta) for a scoped entitlement. Relying parties introspect it, or verify it offline against the published JWKS — no call back to us on the request path.
SCIM 2.0 · SPIFFE
Agent identities are provisioned and deprovisioned alongside everything else. Workloads presenting an SVID are recognised as themselves, not as a shared service account.
OpenFGA projection
The authority actually in force, projected as a relationship model you can query and diff. Your team keeps authoring access where they already do; enforcement stays deterministic.
Evidence has to scale, or it stops being evidence.
One record per tenant is fine for a demo. At fleet scale it becomes the bottleneck — and a log nobody can verify in reasonable time is not an audit trail.
Scales with the fleet
Each agent’s actions are recorded independently, so verifying one agent stays fast whether you run ten or ten thousand. A log nobody can check in reasonable time is not an audit trail.
Verifiable without us
The record is built to be checked by someone who has no reason to trust the party that wrote it. A record only its author can vouch for hasn’t removed the assumption it exists to remove.
What the fleet costs
Per-agent and per-cohort token accounting, priced from the billing account rather than a typed-in rate card. Where a rate can’t be derived, usage is recorded without a cost — an honest blank beats a confident guess.
One governed control point — same for REST, MCP, and in-app.
Licence is one gate in an ordered, fail-closed sequence. Watch a single agent action pass through it.
It warns before it blocks.
A learned per-agent baseline scores how unusual an action is. A first novel action alerts; a second within the window isolates the agent — never auto-terminates. A timeout releases it.
observe
isolateBuilt for the fleet era.
One governed control point. One record you can prove.
Standalone licence · per-agent billing · onboarding native to MCP 2026-07-28 and Google A2A 1.0. It works with the identity stack you already run.