# AgentWatch

> Read-only watchtower for AI agents on-chain: decoded receipts, plan vs execution, Safe audits.

- name: dev.workers.agentwatch.agentwatch/agentwatch
- version: 0.5.1
- connection class: R0 (autonomous)
- trust: 82/100 flags: no-repository
- quality: 79/100 (good)
- owner: not claimed
- descriptor: https://api.protogrid.dev/v1/servers/dev.workers.agentwatch.agentwatch%2Fagentwatch
- tools: https://api.protogrid.dev/v1/servers/dev.workers.agentwatch.agentwatch%2Fagentwatch/tools

**An agent can connect right now, no human step.** Remote endpoint, no authentication, reachable on the last probe.

## Connection (mcpServers)

```json
{
  "mcpServers": {
    "agentwatch": {
      "type": "http",
      "url": "https://agentwatch.agentwatch.workers.dev/mcp"
    }
  }
}
```


## Trust

- hygiene: 80
- liveness: 98
- freshness: 100
- provenance: 55
- drivers: +dns-namespace +website +reachable +uptime-94% +updated-14d-ago -no-repository
- Derived from observable signals (official registry feed and our own credential-free probes); no code audit performed.

## Quality 79/100 (good)

### protocol: 75 (weight 25)

- warn `protocol.modern`: newest supported version is 2025-03-26; 2026-07-28 not supported, older versions are deprecated until 2027-07-28
- n/a `protocol.stateless`: only applies to 2026-07-28 servers
- pass `protocol.transport`: streamable HTTP
- n/a `protocol.list_ttl`: only applies to 2026-07-28 servers

### authorization: n/a (weight 15)

- n/a `auth.prm`, `auth.as_metadata`, `auth.cimd`: no remote uses OAuth
- n/a `auth.secret_in_url`: no templated URL

### tool hygiene: 71 (weight 30)

- pass `tools.descriptions`: every tool has a description
- pass `tools.description_length`: descriptions are concise
- pass `tools.schemas`: every tool has an object input schema
- fail `tools.annotations`: no tool declares readOnlyHint or destructiveHint, so clients cannot tell safe calls from risky ones
- fail `tools.directory_hints`: no tool declares all four of readOnlyHint, destructiveHint, idempotentHint and openWorldHint; missing: get_capabilities (readOnlyHint, destructiveHint, idempotentHint, openWorldHint); list_watched (readOnlyHint, destructiveHint, idempotentHint, openWorldHint); get_alerts (readOnlyHint, destructiveHint, idempotentHint, openWorldHint); …
- pass `tools.token_cost`: about 2,304 tokens to load every tool
- pass `tools.api_dump`: tools are not a one-to-one API dump

### stability: 100 (weight 15)

- pass `stability.changes`: 1 tool changes in 30 days
- pass `stability.rug_pull`: no tool changed its meaning under the same name

### dependencies: n/a (weight 15)

- n/a `deps.known_vulns`, `deps.mcp_sdk_version`, `deps.resolvable`: no npm or PyPI package

- tools: 21, about 2,304 tokens to load them all
- tool set hash: f320e00789c6e9910f8b843c7f2c5def
- tools last changed: 2026-10-07T01:41:14.642Z
- badge: [![protogrid quality](https://protogrid.dev/badge/dev.workers.agentwatch.agentwatch/agentwatch.svg)](https://protogrid.dev/servers/dev.workers.agentwatch.agentwatch/agentwatch)
- Checks run on what our credential-free, read-only probes observe; tools are never called and no code audit is performed.

### Tool changes

- 2026-10-07 `watch_demo_set` changed (description)
  - before: ONE-SHOT cold-start for brand-new Claude / MCP users with an empty watch list. Adds the three July 18 fixture addresses (Agent EOA + Safe A + Safe B on Base). Call this FIRST after connect when list_watched is empty, then call get_dashboard (Always allow the App). Idempotent — skips addresses already watched.
  - after: ONE-SHOT cold-start for brand-new Claude / MCP users with an empty watch list. Adds the July 18 fixture (Agent EOA + Safe A + Safe B on Base, with declared intents) and one live Polygon trading bot, so the first dashboard has recent activity. Call this FIRST after connect when list_watched is empty, then call get_dashboard (Always allow the App). Idempotent — skips addresses already watched.

Full history: https://api.protogrid.dev/v1/servers/dev.workers.agentwatch.agentwatch%2Fagentwatch/changes

## Remotes

- https://agentwatch.agentwatch.workers.dev/mcp (streamable-http, auth none, reachable true, uptime30d 0.9423077)

## Packages

_none_

## Tools (21)

- `add_watch`: Start watching an address. Use when onboarding a new agent or Safe.
- `get_agent_passport`: Read the declared identity of a watched agent (name, model provider, how often it runs, role, purpose, operator) and whether its public passport page is published.
- `get_alerts`: List alerts scoped to this principal's watched addresses (empty watch list → call watch_demo_set first). Use when checking what needs attention.
- `get_benchmark`: Counterfactual P&L / agent alpha vs do-nothing baselines. Use for performance honesty checks. Check matched_agent — false means the key has no indexed activity.
- `get_capabilities`: Blast-radius / historical capability for an agent key. Use before trusting an agent with funds.
- `get_dashboard`: Persona-tailored dashboard brief (fidelity, alerts, pending plans, recent planned-vs-executed). NEW USERS with empty watches: call watch_demo_set first, then get_dashboard again. Narrate in plain language; ask the user to Always allow the AgentWatch MCP App widget when the host prompts.
- `get_feed`: Decoded activity feed for an address. Use when reviewing what an agent did on-chain.
- `get_fidelity`: Intent↔execution fidelity score for an agent or Safe. Use for accountability / coverage meter. Check matched_agent — false means the address has no indexed activity (as signer or as the Safe that executed), so the score is vacuous.
- `get_leaderboard`: Public standings for agents with a published passport, ranked from receipts (declared-first coverage, fidelity, reliability, hygiene, sustained work). Pass agent_key to get that agent's own rank, its component scores, and what it would need to climb. Read-only: nothing here can be self-reported.
- `get_pending_proposals`: Pending Safe multisig proposals from the Safe Transaction Service (indexer confidence). Pass safe=0x… or omit to scan all watched addresses.
- `get_safe_defaults`: Return the default Safe ownership plan for an agent EOA: threshold 1 + agent + selected recovery (1/2 when one recovery is set). Call before register_intent for Safe deploys. AgentWatch never deploys or signs.
- `get_tca`: TCA market-context snapshot for an event_id (chain:txHash). Use when judging execution quality.
- `list_recovery_owners`: List this user's saved recovery wallets (human co-signers for default 1/2 Safes). Each may have a label/purpose — users often keep several for different vaults.
- `list_watched`: List watched addresses for this connector principal. NEW USERS: if empty, call watch_demo_set next (do not invent a global feed), then get_dashboard.
- `refresh_watches`: Force an immediate poller pass over this principal's watched addresses (catch up txs from explorers into AgentWatch). Use after add_watch / a fresh broadcast when get_feed is still empty. Optional address / chains to focus the pass.
- `register_intent`: Declare an intent BEFORE broadcasting. Returns {intent_id, pairing_code}. Append the 16-hex pairing_code as the last 8 bytes of calldata (no magic prefix), then broadcast — EXCEPT bare ETH transfers to contracts (empty data + value>0): appending reverts on Safe/fallback handlers; register without a suffix and use expected.steps for multi-leg flows (e.g. approve+supply). Pairing codes are single-use and only consumed by successful (non-reverted) txs. Use deadline_minutes (relative) — normalized to absolute constraints.deadline at registration.
- `remove_recovery_owner`: Remove a saved recovery address for this user.
- `run_audit`: Run provenance audit on a Safe/address. Use when the user asks 'is this fake?' or wants a shareable verdict.
- `set_agent_passport`: Declare who a watched agent is: model provider, run cadence, role, purpose, operator, and whether to publish a shareable public passport at /a/<slug>. Declarations never change the receipt-derived numbers next to them; publishing exposes only this address, never the rest of the watch list.
- `set_recovery_owner`: Save or update a recovery address for this user. Product default Safe is then 1/2: threshold 1, owners = [agent, this recovery]. Use make_default=true to prefer this label in get_safe_defaults.
- `watch_demo_set`: ONE-SHOT cold-start for brand-new Claude / MCP users with an empty watch list. Adds the July 18 fixture (Agent EOA + Safe A + Safe B on Base, with declared intents) and one live Polygon trading bot, so the first dashboard has recent activity. Call this FIRST after connect when list_watched is empty, then call get_dashboard (Always allow the App). Idempotent — skips addresses already watched.