Lending and risk teams are already experimenting with AI agents against their internal tools — and in most integrations we've seen described, that means one of two failure modes: agents locked out entirely, so the productivity case never materialises, or a full API key handed over with no per-agent boundary, so a leaked credential or a misbehaving prompt has the same blast radius as a compromised human login.
Today we're launching MCP for AI agents — a third way, built from the ground up so an agent can never hold a wider door than the account behind it.
One connection, five evidence families
AssureLocker now exposes five evidence families over the Model Context Protocol: AssureFirst (the neutral duplicate-financing check), Risk Signals (pre-credit deal indicators), Export Provenance (trade-document and eCoO/e-invoice status), Business Identity (GSTIN/Udyam/MCA lookups against a verified entity credential) and Identity & KYC (consent-gated verification requests). These are the same checks a credit analyst already trusts today — now callable by an agent under a key your own team scopes.
Scope that narrows, never exceeds
The core design decision is simple to state and, we think, important to get right: an agent key's usable scope is the intersection of what you chose at mint time and what your account currently holds — never a union, never wider than either side. Narrow your account's access and every agent key cut from it narrows instantly. A tool call outside a key's scope and a tool call for something that was never built return the identical response — scope is never discoverable by probing.
Nothing sensitive lingers in agent context either. Full evidence Packs are never returned inline, only signed links that expire after three minutes. Consent-bound personal data crosses only if your organisation explicitly opts in to a lighter disclosure mode, and even then, only the specific fields the holder already chose to share.
Indicators in, decisions with you
Every tool response is evidence-grade — a band, a status, an indicator — never a credit decision, an approval or a decline. Each one links back to the authenticated platform surface your team already works in. The agent orchestrates; your people decide; the holder consents.
See it before you build against it
The full seven-step journey — mint a key, discover the tools, check for duplicate financing, get a deal signal, orchestrate a KYC request, see what the agent receives, and revoke — is walked through on synthetic data in a guided, no-login simulation.
Honest status
The access-control model and all five families are live today, and we've published a raw JSON-RPC protocol trace proving the wire contract — scope-filtered discovery, deny-by-default access, correct protocol-vs-tool error handling. What we haven't done yet: named-client compatibility testing (we'll say a specific AI client works only once we've recorded a test on that exact surface) and the independent penetration test, which is scheduled but not yet performed.
Explore the developer overview, or book an MCP workflow session with the team.