← neuravant.ai
Runtime governance
This page describes what is deployed and running today, and what is in design-partner preview.
We label the difference, because the difference is the product.
MCP governance gateway live in production
Agents that speak the Model Context Protocol route their tool traffic through the gateway at
app.neuravant.ai/mcp/{upstream}. Every tools/call is evaluated against your
tenant's policy before it is forwarded to your MCP server — pre-action interception, not post-hoc logging.
- Tool policy — allow/deny lists with glob patterns per tenant.
- Egress control — an allowlist of hosts your agent's tools may reach, with a network-guard floor that is always on.
- Scope checks — tools that touch credentials or other sensitive scopes must be explicitly authorised, or the call is flagged as credential overreach.
- Velocity limits — a per-minute call ceiling per agent.
- Observe-first default — new policies default to
flag so you see what would be blocked before you turn on block.
- Kill switch — database-backed; when active, gated calls are refused immediately. Also enforced on LLM traffic routed through our proxy.
- Telemetry — every violation raises an alert and lands in the same telemetry stream your monitoring dashboard reads. A quiet-agent watchdog alerts when an agent goes silent, and a model watcher alerts when the model behind an agent changes.
Onboarding is currently white-glove: we register your upstream MCP server and set the initial policy with
you during a working session, then you manage policy via the admin API. There is no self-serve screen for
this yet. Start with
team@neuravant.ai or book via
certification.
Signed audit chain — neuravant-enforce design-partner preview
The Continuous Agent Attestation library embeds an enforcement gate in-process with your agent: policy match,
verdict, HMAC-SHA256 signature with per-tenant HKDF-derived keys, and a hash-linked decision chain. Mutating or
deleting a record breaks every subsequent signature. The methodology is public:
Continuous Agent Attestation.
Honest status: the library is v0.1.x, not yet on PyPI, and the hosted gate API and this page's former
dashboard are not deployed. Design partners run the library in-process, with durable chain storage and
periodic anchoring configured as part of the engagement. If you want to be one of those partners,
write to us.
What we will not show you
No synthetic uptime bars, no fabricated latency tables, no dashboards wired to endpoints that do not exist.
Live component health is at /status.html. Measured performance figures for the
in-process gate are on the methodology page, labelled with exactly
what was measured and when.