Agentic Wallet Guardian

Moteur de décision auto-hébergé qui se place entre les agents IA et l'exécution blockchain, renvoyant ALLOW/WARN/BLOCK avant toute signature de transaction.

GitHub
1
Essayer ce MCPSponsorisé

Documentation

Agentic Wallet Guardian

agentic-wallet-guardian-v3 MCP server

📄 Read the white paper

A self-hosted decision engine that sits between an AI agent and blockchain execution. Agents submit a proposed action, Guardian returns an explainable ALLOW / WARN / BLOCK before anything gets signed or broadcast.

POST /decision   ->   ALLOW / WARN / BLOCK  (with a reasoned explanation)

It runs on your own infrastructure, using your own policy rules and your own reputation data - see Why self-hosted for why that matters and how this differs from calling a hosted security API directly.

Why self-hosted

There are good hosted alternatives for agent-transaction security (GoPlus's AgentGuard, Blockaid, Chainalysis/TRM for compliance). If you just want a risk score and don't care who sees the query, calling one of those directly is less work than running this. Guardian exists for the cases where that tradeoff doesn't work for you:

  • Nothing about which wallets, contracts, or amounts your agents touch leaves your infrastructure. Threat-intel and contract allow/deny checks are local JSON files you populate yourself (see data/threat_lists/README.md), not a lookup call to a third party. A hosted API inherently sees every address and amount you ask it about.
  • Your policy rules live in your code, not a vendor's dashboard. Spending caps, reputation gates, and which action types require confirmation are plain Python in guardian/policy/, reviewable and changeable without waiting on anyone else's product roadmap.
  • No per-call fees or rate limits imposed by someone else - only the ones you configure for your own users (GUARDIAN_RATE_LIMIT_PER_MINUTE).
  • No vendor lock-in. Every external data source (RPC endpoint, Blockscout instance, DexScreener) is swappable behind a small provider interface - see Architecture.

The honest tradeoff going the other way: you also take on running it, keeping your local threat lists current, and you don't get a hosted vendor's chain coverage or dedicated threat-research team for free. This is the right choice for teams that specifically need data sovereignty or deep policy customization - not a strict upgrade over every hosted option.

Architecture

                AI Agent
                    |
                    v
             Action Intent
   { agent_id, wallet, chain, action_type,
     target, amount, metadata }
                    |
                    v
        ┌───────────────────────────────┐
        │   Guardian Decision Engine    │
        ├───────────────────────────────┤
        │  1. Hard Rules                │  <- chain support, sanity checks
        │  2. Wallet Intelligence       │  <- mock | real RPC (web3.py)
        │  3. Token Intelligence        │  <- mock | real DexScreener | real GoPlus
        │  4. Contract Intelligence     │  <- local lists, then mock | real Blockscout | real GoPlus
        │  5. Simulation                │  <- mock | real eth_call dry-run (see below)
        │  6. Threat Intelligence       │  <- local JSON allow/deny lists
        │  7. Anomaly Detection         │  <- vs. this agent's own history (see below)
        │  8. Policy Engine             │  <- spending caps, reputation gates
        │  9. Risk Fusion               │  <- signals -> single 0-100 score
        │ 10. Reputation Adjustment     │
        │ 11. Explanation               │  <- evidence -> human-readable reasons
        └───────────────────────────────┘
                    |
                    v
          ALLOW / WARN / BLOCK
                    |
                    v
          Blockchain Execution

Every data source in steps 2-4 is a small provider interface with a mock implementation (zero config, zero network calls) and a real one, selected per-source by environment variable - see .env.example. Switching from demo mode to a real deployment is a config change, not a code change.

Repository layout

guardian/
    config.py          GuardianConfig - the one place that reads os.environ
    core/               ActionIntent, Signal, Decision, EvaluationContext
                            (zero external dependencies - no pydantic/FastAPI)
    decision/           DecisionEngine (orchestrator), RiskFusionEngine, hard rules
    reasoning/          explanation + confidence builders
    intelligence/
        wallet/           analyzer.py + providers.py (mock | RpcWalletDataProvider)
        token/            analyzer.py + providers.py (mock | DexScreenerTokenDataProvider | GoPlusTokenDataProvider)
        contract/         analyzer.py + providers.py (mock | BlockscoutContractDataProvider | GoPlusContractDataProvider)
        simulation/       pre-execution dry-run (mock | real eth_call) + tx_builder.py (real calldata for transfer/approve)
        goplus_client.py  shared GoPlus Token Security API client (used by both contract + token)
        threat/           blocklist.py (local AddressList) + intelligence.py
    policy/             PolicyEngine + policy templates (spending caps, reputation gates)
    reputation/         AgentReputation (score derived from decision history)
    memory/             storage.py (protocol) + InMemoryStorage + sqlite_storage.py
api/
    main.py             FastAPI app: /decision, /health, /capabilities, /agents/{id}/history, /demo/{scenario}
    security.py         API-key auth dependency + rate-limit middleware
    schemas.py          pydantic request/response models (API boundary only)
mcp_server.py           MCP stdio server - same DecisionEngine, no HTTP required
data/threat_lists/      local, operator-maintained allow/deny lists (empty by default - see its README)
scripts/
    refresh_ofac_list.py   fetch OFAC's public SDN list into the local threat list
tests/                  101 tests covering the engine, policy, reputation, and every provider

guardian/* is intentionally dependency-free (standard library only, except where a real provider needs httpx or web3), so the decision core can be unit-tested, embedded in another service, or ported to a different web framework without dragging FastAPI along. Only api/ touches pydantic/FastAPI.

Honesty about the current state

This is real, runnable, tested decision infrastructure with real (not mock) data sources available for every signal source - but "available" isn't the same as "flip a switch and trust it blindly." Specifics:

  • Wallet (RPC provider): is_contract and tx_count (nonce-based) are reliable with any JSON-RPC endpoint. Wallet age requires an archive-capable node and is off by default (GUARDIAN_RPC_ESTIMATE_AGE=false) - most free public RPC endpoints don't serve historical state, so this fails closed to "unknown" rather than guessing.
  • Contract (Blockscout provider): real verification-status lookups against a public Blockscout instance. Their exact response schema and rate limits can change - this is written to degrade to "unknown" on any unexpected response, never to fabricate an answer, but hasn't been load- tested against production traffic.
  • Token (DexScreener provider): real liquidity data, but matching a bare ticker symbol to an on-chain pair is inherently ambiguous (many unrelated tokens share a symbol, and scammers deliberately mint look-alikes). The provider picks the highest-liquidity pair on the requested chain and reports its own match confidence rather than presenting a guess as certain - for anything where that ambiguity matters, match by contract address instead of symbol.
  • Contract + Token (GoPlus provider): real contract-security (owner-can-drain, mintable, self-destruct, hidden owner) and trading-security (honeypot, buy/sell tax, blacklist, pausable transfers, holder concentration) from GoPlus's Token Security API - meaningfully more signal types than Blockscout/DexScreener give individually, since GoPlus's own static analysis covers both in one call. Two real limits: it only has data for contracts it's actually analyzed (mostly token contracts, not generic dApp/router contracts), and GoPlusTokenDataProvider needs a contract address - a bare symbol like "PEPE" can't be resolved and is honestly reported as unverifiable rather than guessed at.
  • Sanctioned-address list is real, populated data: 103 addresses (100 EVM + 3 Solana) from OFAC's SDN list, via 0xB10C/ofac-sanctioned-digital-currency-addresses - verified end-to-end (a known-sanctioned address correctly triggers BLOCK through the full pipeline) and verified to correctly reflect delistings, not just additions (Tornado Cash's addresses, removed from the SDN list in March 2025, are correctly absent). Re-run scripts/refresh_ofac_list.py periodically - sanctions change in both directions.
  • malicious_contracts.json / verified_contracts.json still ship empty on purpose (see data/threat_lists/README.md) - there's no single authoritative source for "malicious contract" the way OFAC's list is authoritative for sanctions, so populating these is a judgment call for whoever operates this instance, not something to seed by default with unverified entries.
  • Simulation is real, but conditional. RpcSimulationProvider (GUARDIAN_SIMULATION_PROVIDER=rpc) genuinely dry-runs a transaction via eth_call/eth_estimateGas against current chain state - a revert comes back with its actual reason, not a guess, and ERC-20 approve() amounts are decoded from real calldata instead of inferred. This activates when the caller supplies raw calldata via intent.metadata["data"], OR - new - when GUARDIAN_TX_BUILDER=rpc is also set and the intent is a plain transfer or approve (see next bullet). A swap intent with no transaction built yet still has nothing to dry-run - Guardian reports that honestly (simulation_not_attempted) rather than guessing.
  • Transaction building closes part of that gap, deliberately not all of it. RpcTransactionBuilder (GUARDIAN_TX_BUILDER=rpc) turns a semantic transfer/approve intent into real calldata - it fetches the token's actual decimals() via RPC rather than assuming 18 (a wrong assumption there would scale the amount by orders of magnitude), and deliberately has no hardcoded token-address registry: a bare symbol like "USDC" is refused rather than guessed at, since a wrong address here wouldn't just be a bad risk signal, it'd be an artifact that could end up in a real transaction. swap is built against Uniswap V2 Router02 only (one immutable, well-known contract - function selectors computed locally via Web3.keccak, not copied from memory) - real getAmountsOut() on-chain quote, caller-supplied max_slippage_bps required (never a default, same reasoning as decimals above). bridge is a genuinely open-ended L2/bridge routing problem in general - dozens of protocols, wildly different trust models - but this module handles one well-scoped slice of it: L1 -> L2 deposits through a destination chain's own official OP Stack bridge (currently: Base and Optimism - depositETHTo/depositERC20To on L1StandardBridge, both addresses independently cross-checked - Base against Etherscan's label plus basehub.org, Optimism against the official ethereum-optimism/superchain-registry plus a second independent dev-tool config - before being hardcoded). L2 -> L1 withdrawals are NOT built - that's a genuinely different, much slower proof/challenge-window flow, not a variant of the deposit call. Bridging to anywhere else, or via any non-canonical bridge, returns None rather than guessing.
  • Storage: InMemoryStorage (default, zero setup), SQLiteStorage (GUARDIAN_STORAGE_BACKEND=sqlite - persists across restarts, no external infra), or PostgresStorage (GUARDIAN_STORAGE_BACKEND=postgres + GUARDIAN_POSTGRES_DSN - the fit for multiple replicas behind a load balancer, where SQLite's single-writer model becomes the bottleneck; pip install -r requirements-postgres.txt). Tested against a real local Postgres instance, not mocked - see tests/test_postgres_storage.py. Redis is still open if you specifically want it; the two-method MemoryBackend interface is small enough to implement against anything.
  • API auth/rate-limiting are intentionally minimal - built for one self-hosted instance behind your own network boundary, not a multi-tenant gateway. Put a real API gateway in front if you need that.
  • Not security-audited. The policy engine and risk fusion logic have not been reviewed by anyone outside this repo. Treat BLOCK as a strong signal, not a guarantee, until that's happened.

Everything downstream of a Signal - fusion, policy, reputation, explanation, the API - does not need to change as any of the above gets hardened further. That boundary is the actual design contract here.

Quickstart

Zero-config demo mode (mock providers, in-memory storage, no auth):

pip install -r requirements.txt
uvicorn api.main:app --reload

Or with Docker:

docker compose up --build

Try the canned scenarios:

curl http://localhost:8000/demo/safe
curl http://localhost:8000/demo/unknown
curl http://localhost:8000/demo/malicious

Or submit your own intent:

curl -X POST http://localhost:8000/decision \
  -H "Content-Type: application/json" \
  -d '{
        "agent_id": "trading-agent-001",
        "wallet": "0x742d35Cc6634C0532925a3b844Bc454e4438f44e",
        "chain": "ethereum",
        "action_type": "swap",
        "from_token": "ETH",
        "to_token": "USDC",
        "amount": 5
      }'

Going from demo to a real self-hosted deployment

Copy .env.example to .env and adjust:

cp .env.example .env

At minimum for a real deployment: set GUARDIAN_API_KEY (auth is off by default), GUARDIAN_STORAGE_BACKEND=sqlite (persistence), and whichever GUARDIAN_*_PROVIDER variables you want pointed at real data instead of mock - see the comments in .env.example for every option, and RpcWalletDataProvider/BlockscoutContractDataProvider/ DexScreenerTokenDataProvider/GoPlusContractDataProvider/ GoPlusTokenDataProvider's docstrings for what each one actually gives you.

MCP (no HTTP required)

For agent frameworks that speak MCP (LangChain, CrewAI, Claude Desktop, etc.), mcp_server.py exposes the same decision engine as two tools (evaluate_action, get_agent_history) over stdio - install requirements-mcp.txt in its own virtual environment (see the comment at the top of that file for why it can't share an environment with requirements.txt) and point your MCP client at python mcp_server.py.

Running the tests

pip install -r requirements.txt -r requirements-chain.txt
pytest -q

requirements-chain.txt (web3) is only needed for the RPC-provider tests; the rest of the suite runs with just requirements.txt. The guardian/* core has no external dependencies beyond that, so it's also runnable with:

PYTHONPATH=. python3 -m unittest discover -s tests -v

CI (.github/workflows/ci.yml) runs the full suite on every push/PR against Python 3.11 and 3.12.

Signed, verifiable decisions (OAA)

Every decision this service returns — from a single policy check up through the full pipeline — is signed as an OAA (Open Agent Attestation) token: an Ed25519-signed JWT wrapping the decision, the action, and the reason.

Anyone holding the public key can verify a decision offline, without calling back to whatever instance of Guardian issued it — useful for an auditor, a downstream service, or just a record you want to trust later without trusting the server that produced it.

python examples/example_oaa_attestation.py
python examples/example_full_pipeline.py   # capability -> intent -> engine -> OAA

The reference OAA implementation is ~150 lines (oaa.py/attestation.py upstream) and is shared, unmodified, across this project and agent-guardrail — same signing format, same verification path, no per-project fork.

Using Guardian in front of MetaMask Agent Wallet

MetaMask Agent Wallet's Guard Mode / Beast Mode apply the same static spend limits and allowlists to every agent. Guardian is a second, independent check in front of it: does this specific action look right for this agent, right now — before the mm CLI is ever invoked.

skills/guardian-check/ is a standard Agent Skill — the same open format MetaMask itself uses for mm (npx skills add MetaMask/agent-skills). Install it alongside MetaMask's own skill in any Agent-Skills-compatible runtime (Claude Code, Cursor, Codex, OpenClaw), and the agent will call a running Guardian instance for an ALLOW/WARN/BLOCK decision before running any mm command that moves funds — mm send, mm swap, mm bridge, mm perps, mm predict trade, mm earn, mm aave, mm pay.

Guardian never holds keys and never executes anything — mm remains the only thing that signs or broadcasts. This is a decision gate the agent is instructed to consult first, not a modification to MetaMask's own pipeline (there's no public hook for that today).

uvicorn api.main:app --reload   # run Guardian locally
export GUARDIAN_API_URL="http://localhost:8000"
python skills/guardian-check/scripts/check.py \
  --agent-id my-agent --wallet 0x... --chain ethereum \
  --action-type transfer --target 0x... --amount 50

Roadmap

  1. Replace the mock wallet/token/contract analyzers with real data sources. Done - see Honesty about the current state for what "real" does and doesn't cover yet per source.
  2. Wire up real pre-execution simulation. Done for transfer/ approve end to end (GUARDIAN_SIMULATION_PROVIDER=rpc + GUARDIAN_TX_BUILDER=rpc - see Honesty about the current state). swap needs real DEX routing. Done against Uniswap V2 Router02 - real on-chain getAmountsOut() quote, explicit caller-supplied max_slippage_bps (never defaulted), no calldata built without a real quote. Fixed a real bug found while building this: simulation was dry-running against intent.target (the recipient/spender encoded inside ERC-20 calldata) instead of the actual contract being called (from_token) - meaning transfer/approve simulation silently "succeeded" against any EOA recipient regardless of whether the real call would have reverted. BuiltTransaction now carries an explicit to; see tests/test_tx_builder.py for the regression tests that would have caught it. bridge still open. Done for L1->L2 deposits to Base and Optimism via the official OP Stack L1StandardBridge (depositETHTo/depositERC20To) - other destinations, other bridge protocols, and L2->L1 withdrawals all remain open; see Honesty about the current state.
  3. Populate threat-intel / sanctions feeds; stop shipping empty sets. Done for sanctions (sanctioned_addresses.json - 103 real OFAC SDN addresses, refreshable via scripts/refresh_ofac_list.py). malicious_contracts.json / verified_contracts.json remain empty by design - no single authoritative source exists to seed them the way OFAC's list does for sanctions.
  4. Swap InMemoryStorage for a persistent backend. SQLiteStorage is available; a Postgres/Redis backend is still open for multi-replica deployments. PostgresStorage done - tested against a real local Postgres instance (tests/test_postgres_storage.py), same two-method MemoryBackend interface as the other backends. Redis remains open if specifically wanted.
  5. Add an MCP server wrapper. Done (mcp_server.py). A packaged Python/TypeScript SDK on top of the REST API is still open.
  6. Publish an OpenAPI spec and a hosted demo endpoint.
  7. Get the policy engine and risk fusion reviewed/audited before anyone relies on a BLOCK from this service in production - it's a security tool, so it needs the same scrutiny it applies to others.
  8. Add per-agent capability limits (delegation scoping). Done - guardian/policy/capabilities.py. Opt-in: an operator can grant a specific agent a scoped capability (allowed action types, allowed chains, per-action and daily spending caps, an expiry) with zero private-key material involved. Agents with no grant are unaffected - see examples/example_capability_limits.py. Real key management (session keys, account abstraction) remains deliberately out of scope - a categorically higher-stakes problem.
  9. Verify declared intent against decoded simulation results. Done for approve - guardian/decision/intent_verification.py catches the case where an agent declares one amount but the actual calldata it was handed encodes a meaningfully different (but still finite) one. This is distinct from the existing "unlimited approval" signal, which only catches near-uint256-max values - see examples/example_intent_verification.py.
  10. Flag actions that deviate from an agent's own historical pattern. Done - guardian/intelligence/anomaly/analyzer.py. Distinct from reputation (a single trust score) and policy (static, operator-set limits): this compares the current intent against this specific agent's own recorded history - new action type, new chain, or an amount that's a statistical outlier versus what this agent has done before, even if it's within policy limits and the agent's reputation is fine. Honestly reports "insufficient history" rather than guessing a baseline from fewer than 5 prior data points - see tests/test_anomaly_detection.py.
  11. Sit in front of a real agent wallet, not just accept intents from a generic API caller. Done for MetaMask Agent Wallet - skills/guardian-check/ is a standard Agent Skill an agent installs alongside MetaMask's own mm skill; the agent calls it before running any fund-moving mm command and only proceeds on ALLOW. Tested end-to-end against a live uvicorn instance (ALLOW/WARN/BLOCK/config-error all exercised for real, not just asserted) - see the "Using Guardian in front of MetaMask Agent Wallet" section above. No public hook exists (yet) to run inside MetaMask's own pipeline; this works at the agent-orchestration layer instead.

Related projects

Same author, same principle applied elsewhere:

  • agent-guardrail - a generic policy firewall for AI agent tool calls (not blockchain-specific). Published on PyPI, MIT, 46 tests.
  • x402-attest - cryptographically signed (Ed25519), independently verifiable attestations for agent-to-agent payment policy decisions. Early proof of concept.
  • open-agent-attestation - vendor-neutral open spec (JWT+EdDSA) for signing agent policy decisions, verifiable by anyone. x402-attest above uses a custom format; this is the generalized version. Draft v0.1.

License

MIT - see LICENSE.