GRIP is an open specification for contracting agent behavior. Where protocols like MCP and A2A describe what an agent can do, GRIP defines what it may do, what it owes, and what it can prove — the missing accountability layer for autonomous systems.
Capability protocols answer “what tools does this agent have?” That's necessary, but it's not enough to deploy autonomy responsibly. GRIP adds the governance vocabulary: a machine-readable contract that travels with every job.
It is deliberately open. Accountability shouldn't be a proprietary moat — it should be a shared standard that any runtime, model, or tool can speak.
What an agent may do — the bounded envelope of resources, scopes, and budgets it is permitted to act within, expressed explicitly rather than implied.
What an agent owes — commitments it must satisfy, such as returning settlement receipts within token limits or producing required evidence.
What an agent can show — content-addressed evidence that anyone can verify independently, long after the job has closed.
GRIP doesn't replace capability and agent-to-agent protocols — it wraps them with the contract they were missing.
The Model Context Protocol exposes tools and context to a model. GRIP governs whether, when, and under what limits those tools may be used.
Agent-to-agent protocols let agents talk. GRIP makes each of those interactions contractual, obligated, and provable.
An open, draft-1.0 specification. Implemented natively by the Foxbra Runtime and Fennec, and free for anyone to adopt.