vcp
hassard0/vcp
VCP — Verifiable Capability Protocol: zero-trust capability execution for AI agents. A stricter sibling of MCP where models plan and a signed gateway enforces. Signed content-addressed capabilities, proof-bound single-use grants, plan/apply, taint-aware policy.
Instalação
npx convez-ai hassard0/vcp1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
README
A tradução para o português deste README ainda está na fila.
VCP — Verifiable Capability Protocol
Zero-trust capability execution for AI agents. A model may propose a tool call, but it can never authorize one. Authorization comes from a signed, content-addressed manifest, a mandatory policy decision, explicit consent, and a single-use, proof-bound grant minted by an enforcing Gateway.
VCP is a stricter sibling of the Model Context Protocol (MCP). MCP's breakthrough is ecosystem simplicity — and that same easy composability is its security weakness. The MCP spec itself states it cannot enforce many security principles at the protocol level, and that authorization is optional.
VCP flips that: security, provenance, policy, and determinism are protocol requirements, not implementation advice.
Model proposes plan
→ Gateway validates manifests & plan
→ Policy authorizes a bounded grant
→ Provider executes within the grant (sandboxed)
→ Gateway validates the signed attestation
→ Model receives a tainted result (never authority)
Why VCP exists — the MCP failure modes it eliminates
| MCP failure mode | VCP control |
|---|---|
| Tool poisoning — hidden instructions in tool descriptions | Descriptions are never authority. The Planner gets a Gateway-compiled affordance from a signed manifest, never raw Provider text. |
| Rug pulls — tool definitions mutate after approval | Identity is the contract hash. Any change ⇒ a new capability id ⇒ rejected until re-approved. A silent mutation becomes a visible diff. |
| Over-trusted local servers — STDIO runs with the host's privileges | VCP-Local sandbox: signed launcher, no inherited env, filesystem/network allowlists, secret broker. Ambient authority only in dev. |
| Token passthrough / confused deputy | The unit of authority is a single-use, proof-bound grant bound to capability + arguments + plan + scope + budget + a holder key. No reusable bearer token to pass through. |
| SSRF, session hijacking, replay | Stateless VCP-HTTP: one request = one decision, guarded metadata discovery, single-use grants, no implicit sessions. |
| Stateful-session ambiguity | No implicit protocol sessions. State is an explicit, typed, expiring handle. |
| Prompt injection via resources | Taint labels: authority never flows from untrusted_* data, even when the model is tricked into proposing a bad plan. |
| Cross-server shadowing / confused deputy (multi-provider) | Per-provider scoped credentials via token exchange + an on-behalf-of delegation chain in every grant and audit event; one user approval covers the whole cross-service action (§26). |
The shape of the idea
Capabilities are signed.
Descriptions are not authority.
Models plan; gateways enforce.
Every side effect needs a bounded grant.
Every write has plan/apply semantics.
Every output is tainted until policy says otherwise.
Every state handle is explicit.
Every sensitive call is replayable or explicitly non-deterministic.
Every manifest change is a new identity.
Every production provider is sandboxed, authenticated, and auditable.
Repository layout
SPECIFICATION.md The normative v0.1 spec (RFC-2119). Start here.
schemas/ Normative JSON Schemas for every envelope (manifest, grant, plan,
policy request/response, invocation, attestation, audit, discovery).
rfcs/ Open RFCs — the deferred/large ideas, open for discussion.
docs/design/ Design rationale behind v0.1.
CHANGELOG.md Dated protocol revisions (Keep a Changelog).
GOVERNANCE.md How VCP is governed and how the RFC process works.
SECURITY.md Threat model + responsible disclosure.
Conformance ladder
| Level | What it adds |
|---|---|
| VCP-L0 | MCP-compatible bridge: wraps MCP servers, signs observed schemas, adds policy + audit, marks trust legacy. |
| VCP-L1 | Signed, content-addressed manifests; strict schema validation; no hidden metadata changes. |
| VCP-L2 | Mandatory auth; per-call proof-bound grants; sandboxing; network/file/secret isolation; policy interface. |
| VCP-L3 | Plan/apply; dry-run for writes; idempotency keys; replay logs; result attestations; snapshot refs. |
| VCP-L4 | Transparency registry; reproducible-build provenance; formal policy verification; DLP/data-flow proofs. |
Each level has a normative security test suite (12 attack scenarios; see SPECIFICATION.md §18).
Reference implementations
Reference SDKs and gateways live in
vcp-servers — a lightweight
client/SDK + MCP bridge and a heavy enforcing gateway, in TypeScript, Python,
Go, and Rust, all driven by shared, language-agnostic conformance vectors.
Ecosystem
VCP is designed to compose with existing building blocks rather than reinvent them:
cani— a local-first Policy Decision Point; a conformant Policy Authority for the §6 decision interface.mcp-ledger— append-only audit + budget enforcement; a conformant audit and budget substrate (grants carry a budget; the ledger enforces it).prosecco-ai-standards— the AI-interoperability standards directory where VCP is listed.- Sigstore / SLSA inform manifest signing and the L4 transparency registry.
- OPA / Cedar satisfy the §6 policy decision shape.
- OpenTelemetry is the observability substrate for §20.
Status & contributing
VCP is a Draft RFC — it may change incompatibly until it reaches Stable.
Discussion happens in this repo's Discussions; normative changes go through the
RFC process. See CONTRIBUTING.md and
GOVERNANCE.md. Found a security-relevant gap? See
SECURITY.md.
License
The prose specification (SPECIFICATION.md, docs/, rfcs/) is licensed
CC BY 4.0; the schemas/ directory and any code are Apache-2.0. See
LICENSE-SPEC and LICENSE.