convez.ai
MCP ServerSegurançaAvançado
agent-receipts

Agent Receipts

agent-receipts/obsigna

Assinatura criptográfica e auditoria para ações de agentes IA

Ver no GitHub
BomPluginScore · de 0 a 100

MCP server que cria trilhas de auditoria criptograficamente assinadas para ações de agentes de IA. Implementa um protocolo aberto com suporte para múltiplas linguagens (Go, TypeScript, Python) e permite rastreamento tamper-evident de todas as chamadas de ferramentas.

Instalação

$npx convez-ai agent-receipts/obsigna

1 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

Estrelas
22
Forks
3
Acompanhando
0
Saúde
87%
Último commit
há 3 dias
Versão
sdk-py-v0.16.0
Issues abertas
38
Contribuidores
2

Casos de uso

  • Auditoria de ações executadas por agentes IA em produção
  • Conformidade regulatória para sistemas com agentes autônomos
  • Verificação e rastreamento de chamadas de ferramentas em workflows

Veredito

Solução robusta e bem-testada para governança de agentes IA, atendendo necessidade real de auditoria e compliance. Excelente para equipes que precisam de trilhas de auditoria verificáveis e imutáveis em sistemas com agentes autônomos.

README

A tradução para o português deste README ainda está na fila.

Agent Receipts

Cryptographically signed audit trails for AI agent actions

Go Tests TS Tests Python Tests Daemon MCP Proxy Hook License: Apache-2.0

Protocol site & specagentreceipts.ai
Tooling docsobsigna.dev
Daemon setup & migration guideobsigna.dev/getting-started/daemon-setup/
API referenceGo · TypeScript · Python
BlogYour AI Agent Just Sent an Email · Every MCP Tool Call My AI Makes Now Gets a Signed Receipt
Gosdk/go · mcp-proxy · dashboard
npm@obsigna/sdk-ts
PyPIobsigna

What is this?

Agent Receipts is an open protocol for producing cryptographically signed, tamper-evident records of AI agent actions. It defines the receipt format, signing scheme, chain structure, and taxonomy of action types. Anyone can implement it — in any language, in any runtime.

Obsigna is the reference toolset that implements the protocol:

ToolWhat it does
obsigna-mcpMCP stdio proxy — signs every tool call, adds policy hooks
obsigna-daemonOut-of-process signing daemon — holds the key, owns the audit chain
obsigna-hookPostToolUse hook for Claude Code and other runtimes
obsignaCLI for browsing and verifying receipt databases
sdk/go, @obsigna/sdk-ts, obsigna (Python)SDKs for embedding receipt creation in your own code

Start here

Both paths below require the daemon — it holds the signing key and owns the audit chain. Install it first:

brew install agent-receipts/tap/obsigna
obsigna daemon start

Fastest path — PostToolUse hook (Claude Code): one config snippet and every tool call gets a signed receipt automatically:

{
  "hooks": {
    "PostToolUse": [{ "matcher": "", "hooks": [{ "type": "command", "command": "obsigna-hook" }] }]
  }
}

Add that to ~/.claude/settings.json, then inspect the audit trail:

obsigna list
obsigna show <seq>
obsigna verify

More control — MCP proxy: wraps any MCP server and adds policy hooks and risk scoring on top of signed receipts:

Project layout

ProjectDescription
spec/Protocol specification, JSON schemas, governance
sdk/go/Go SDK
sdk/ts/TypeScript SDK
sdk/py/Python SDK
daemon/Signing daemon — out-of-process key custody, shared audit chain
mcp-proxy/MCP proxy with receipt signing, policy engine, intent tracking
hook/PostToolUse hook binary for Claude Code and other runtimes
cross-sdk-tests/Cross-language verification tests
docs/adr/Architecture Decision Records
dashboardLocal web UI for browsing and verifying receipt databases
openclawAgent Receipts plugin for OpenClaw

SDK quick start

Choose your trust model. The snippets below keep the signing key inside the agent process — a deliberate deployment model where the agent host is trusted and tamper-evidence is aimed at downstream parties. In this model, anyone with code execution in the agent can forge receipts. To defend against a compromised agent, use the daemon-mediated path, where a separate daemon owns the key and your app only sends events over a socket. See the Trust Model page for the full spectrum (in-process → daemon-isolated → HSM/KMS).

Go

go get github.com/agent-receipts/ar/sdk/go
import "github.com/agent-receipts/ar/sdk/go/receipt"

keys, _ := receipt.GenerateKeyPair()
unsigned := receipt.Create(receipt.CreateInput{
    Issuer:    receipt.Issuer{ID: "did:agent:my-agent"},
    Principal: receipt.Principal{ID: "did:user:alice"},
    Action:    receipt.Action{Type: "filesystem.file.read", RiskLevel: receipt.RiskLow},
    Outcome:   receipt.Outcome{Status: receipt.StatusSuccess},
    Chain:     receipt.Chain{Sequence: 1, ChainID: "chain_1"},
})
signed, _ := receipt.Sign(unsigned, keys.PrivateKey, "did:agent:my-agent#key-1")

TypeScript

npm install @obsigna/sdk-ts
import {
  createReceipt,
  generateKeyPair,
  signReceipt,
} from "@obsigna/sdk-ts";

const keys = generateKeyPair();
const unsigned = createReceipt({
  issuer: { id: "did:agent:my-agent" },
  principal: { id: "did:user:alice" },
  action: { type: "filesystem.file.read", risk_level: "low" },
  outcome: { status: "success" },
  chain: { sequence: 1, previous_receipt_hash: null, chain_id: "chain_1" },
});
const signed = signReceipt(unsigned, keys.privateKey, "did:agent:my-agent#key-1");

Python

pip install obsigna
from obsigna import (
    create_receipt, generate_key_pair, sign_receipt,
    CreateReceiptInput, Issuer, Principal, Outcome, Chain,
)
from obsigna.receipt.create import ActionInput

keys = generate_key_pair()
unsigned = create_receipt(CreateReceiptInput(
    issuer=Issuer(id="did:agent:my-agent"),
    principal=Principal(id="did:user:alice"),
    action=ActionInput(type="filesystem.file.read", risk_level="low"),
    outcome=Outcome(status="success"),
    chain=Chain(sequence=1, previous_receipt_hash=None, chain_id="chain_1"),
))
signed = sign_receipt(unsigned, keys.private_key, "did:agent:my-agent#key-1")

See the Python SDK README for the full quick start and daemon delivery.

Contributing

See CONTRIBUTING.md for development setup and PR guidelines.

Security

See SECURITY.md to report vulnerabilities. The threat model documents trust boundaries, in-scope and out-of-scope threats, and the mitigation roadmap.

License

Apache License 2.0 -- see LICENSE. The protocol specification in spec/ is licensed under MIT.

Ver o original no GitHub

Bom para

Relacionados