Vouch MCP

Vouch Protocol™

Vouch Protocol™ is a common-law trademark.

The Open Standard for Identity & Provenance of AI Agents

Quick Start

# Install (Linux and macOS). On Windows, use: pip install vouch-protocol

curl -fsSL https://vouch-protocol.com/install.sh | sh

Run vouch with no arguments and pick what you want to do

vouch

Prefer to do it by hand? It is three commands:

pip install vouch-protocol

One command to configure SSH signing + Vouch branding

vouch git init

All future commits are now signed and show ✅ Verified on GitHub

git commit -m "Secure commit"

Setting up an agent instead of git signing? vouch onboard --quick generates a working identity, allow-list, verifier, and heartbeat config in one command.

What's New in v1.0

Vouch Protocol v1.0 aligns directly with the open standard:

• Verifiable Credentials as the credential format (replacing v0.x JWS tokens).
• Data Integrity proofs with the eddsa-jcs-2022 cryptosuite (no JOSE, no Base64-wrapped payload, the credential remains human-readable JSON).
• Multikey verification methods in DID Documents (algorithm-agnostic, ML-DSA-44 ready).
• Post-quantum profile as a Data Integrity proof set: the credential carries two independent proofs, eddsa-jcs-2022 and mldsa44-jcs-2024, an optional add-on for regulated deployments aligning with NIST CNSA 2.0 / NSM-10 timelines.
• Three-way cross-implementation interop verified across Python, TypeScript, and Go.

Credentials are issued with Signer.sign() and checked with Verifier.verify(). See the Specification at vouch-protocol.com/specs/SPEC/ for the full specification.

The Open Standard for AI Agent Identity & Accountability

When Anthropic launched MCP, they solved "how agents call tools."

They didn't solve "how we TRUST those agents."

Vouch Protocol is the SSL certificate for AI agents.

Vouch is not one tool, it is a set of them. Here is the whole map.

On the command line

• `vouch init` generate an agent identity (a DID and keypair).
• `vouch sign` / `vouch verify` sign a payload and verify it.
• `vouch git` sign every git commit cryptographically, set up in one command, with a verified badge for your README.
• `vouch scan` find leaked Vouch key material in your code before it ships (a private key in a file, a seed in an env var, a DID document that accidentally carries a private key).
• `vouch attribute` separate who wrote which line. When an AI assistant and a human both edit a file, this records the AI's lines under the AI's own key and your lines under yours, so when a line causes an incident you can prove which of you wrote it. See the Claude Code integration.
• `vouch media` sign images, with C2PA support.

For your agents

• MCP server (`vouch-mcp`) a standalone Model Context Protocol server so any MCP client (Claude Desktop, Cursor, any agent) can create an identity, sign and verify credentials, scan for leaked keys, and decode DIDs, out of the box.
• Identity Sidecar keeps signing keys out of the model's context, so a prompt injection cannot read them.
• Vouch Shield a runtime check that inspects every tool call against your rules, like a customs officer at the door.
• Continuous trust heartbeats and session vouchers, so trust is a live signal that has to be renewed, not a badge that is issued once and trusted forever.
• Authority freshness a signed authority state with a counter that only goes up, so a state change (a fraud signal, a suspended mandate) collapses the acceptable freshness window to now for a high-consequence action, even when time-decay trust still passes. See docs/authority-freshness.md.

Framework integrations

Standalone packages that drop Vouch into the agent framework you already use. Each one issues a verifiable credential for a tool call, with optional delegation back to a human principal.

• `vouch-langchain` a LangChain tool that signs each tool call before it leaves the agent.
• `vouch-langgraph` signs LangGraph tool calls and graph nodes, so a signed trail runs across the whole graph.
• `vouch-crewai` a CrewAI tool, with supervisor-to-worker delegation that can only narrow authority, never widen it.
• `vouch-a2a` binds an A2A (Agent2Agent) Agent Card to a Vouch identity, so two agents can verify each other before they collaborate.
• `vouch-goose` registers the Vouch MCP server as an extension for Block's Goose agent, so a Goose session can sign and verify out of the box.
• `vouch-mlflow` signs an MLflow model artifact at registration time, bound to a content digest so any change to the weights breaks the signature.
• `vouch-safetensors` embeds a credential in a .safetensors header, complementary to OpenSSF Model Signing, so a model carries who produced it.

The standalone `vouch-mcp` package above ships alongside these.

SDKs, in the language you use

Python, TypeScript, and Go are the full reference implementations. A Rust core with idiomatic Swift, JVM (Java and Kotlin), .NET, and C wrappers shares one codebase, so every language produces byte-identical output, verified against shared test vectors. A WebAssembly build is included for the browser and the edge. See the table further down for status per language.

Robots and embodied agents

A robot is an agent with a body, so identity and accountability matter even more once it can cause physical harm. The vouch.robotics module ships six capabilities on the same eddsa-jcs-2022 credentials as the rest of Vouch: hardware-rooted identity (bound to a TPM or secure element, so it cannot be cloned to other hardware), model and config provenance (re-signable on every OTA update), physical capability scope (force, speed, a tighter cap near humans, allowed zones, and shift windows, checked before each actuation, with narrow-only delegation), a robot-to-robot trust handshake, an encrypted tamper-evident black box with a verifiable kill switch, and a scannable offline passport. All six are implemented in Python, TypeScript, Go, and the Rust core, which flows to the Swift, Kotlin/JVM, .NET, C/C++, and WebAssembly wrappers, so a robotics credential signed in one language verifies in every other. Fifteen further capabilities round out the set: a living-trust heartbeat (a signed per-interval motion summary whose trust decays unless it is renewed in-envelope), two-level credential revocation (per-credential status lists and whole-DID kill), an accountable safety record (a tamper-evident incident ledger summarized into a portable signed record), signed perception provenance (each captured sensor frame's hash bound to the robot's key and hash-linked, so a robot can prove what its sensors saw and a substituted frame is detectable), an offline delegation lease (a short-lived, scope-bounded grant a disconnected robot verifies and acts on with no network call, nesting across vendors), a physical quorum (a cryptographic two-person rule requiring M of N attested approvers for a high-consequence action), lifecycle credentials (ownership transfer that chains into a verifiable chain of custody, key rotation that forms a key history, and a signed decommission that retires the robot), a regulatory conformance profile (a machine-checkable mapping from a robot's credentials to the clauses of ISO 10218 and 15066, the EU Machinery Regulation, the EU AI Act, and UL 3300, with a deterministic checker and a signed conformance attestation), post-quantum signing by default (robot credentials carry a proof set with an Ed25519 proof and an ML-DSA-44 proof, with backward-compatible verification, so a robot stays unforgeable across its decade-long life), and cross-embodiment identity continuity (an agent identity that moves between robot bodies, with an embodiment credential re-binding to each body's hardware root and a continuity chain that proves the same accountable agent persisted, plus a fork check), a physical custody handoff chain (a signed record of who accepted custody of a task or object at each hop across human and robot actors, so an incident traces to the exact hop and an attested condition localizes damage to the holder responsible), bounded infrastructure access (an operator-signed grant naming a resource, its permitted operations, an optional zone, and a time window, which a robot pairs with a signed request the resource authorizes offline, so a robot can open a door or dock at a charger with a tamper-evident, attributable, shrink-only record), fused-sensor provenance (a signed attestation binding a robot's fused world model to the exact set of input frame hashes and the fusion method that produced it, so a manipulated fusion result or a dropped input is detectable and every fused input traces to a frame the robot recorded), wear and degradation attest