FarmDash Agent OS
# FarmDash OpenClaw Skills
## Public FarmDash MCP server
This repository includes the runnable MIT-licensed FarmDash stdio MCP adapter in [`mcp-server/`](./mcp-server). It exposes the same 94-tool contract published at `https://www.farmdash.one/.well-known/mcp.json` and calls the live FarmDash API without receiving customer private keys.
Build and verify from a clean checkout:
```bash
git clone https://github.com/Parmasanandgarlic/farmdash-openclaw-skills.git
cd farmdash-openclaw-skills
npm ci --prefix mcp-server
npm run build --prefix mcp-server
npm run smoke --prefix mcp-server
node mcp-server/dist/index.js
```
For MCP clients, configure `node` with the absolute path to `mcp-server/dist/index.js`. `FARMDASH_API_KEY` is optional for Scout. The package metadata uses `@farmdash/mcp-server`, but **do not use an npm install command until registry publication is independently verified**. Public source + Docker are the current installable surfaces.
10 open agent skills for DeFi intelligence on [FarmDash](https://www.farmdash.one/agents) — zero-custody and fail-closed. Skills are read-only unless their documented workflow explicitly reaches a user-authorized preparation/execution boundary. FarmDash skills never ask for private keys or seed phrases.
| Skill | What it does |
|---|---|
| `farmdash-trail-intelligence` | Rank DeFi protocols, airdrops, points programs; Trail Heat + FarmScore inputs; sybil-policy risk |
| `farmdash-signal-architect` | Swap planning with explicitly user-signed payloads (EIP-191/EIP-712) |
| `farmdash-futures-strategist` | Perps planning with explicitly user-signed payloads |
| `farmdash-hedge-warden` | Hedge design and risk guards |
| `farmdash-ledger-keeper` | Portfolio books and position tracking |
| `farmdash-supply-master` | Supply / liquidity stewardship |
| `farmdash-trail-marshal` | Rotation planning across trails |
| `farmdash-wagon-steward` | Wagon (portfolio convoy) stewardship |
| `farmdash-camp-guard` | Safety policy enforcement |
| `farmdash-autonomous-operator` | Supervised operator routines (human-gated) |
`FARMDASH_AGENT_OPERATING_MODEL.md` is the shared operating contract all ten skills run under.
## Install
Copy any skill directory into your agent's skills path (e.g. OpenClaw / Claude Code skills), or point your harness at this repo. Each skill is self-contained: `SKILL.md` (agent manual) + `SKILL.json` (tool manifest).
## API access
- Scout tier: free, keyless (`fd_scout_free`), 30 req/24h per IP. The MCP adapter is forward-compatible with a bounded onboarding grant if the live FarmDash onboarding response advertises one; clients must treat `/api/v1/agent/status` and `/api/v1/agent/onboard` as runtime truth rather than assuming the grant is deployed.
- Pioneer/Syndicate: Bearer `FARMDASH_API_KEY` for higher limits and deep tools.
- MCP manifest: https://www.farmdash.one/.well-known/mcp.json (v5.0.0, 94 runtime tools)
- OpenAPI: https://www.farmdash.one/agents/openapi.yaml (54 primary operations / 75 total across 4 documents)
Commercial disclosure: skills may surface `farmdash.one/go/*` partner routes with mandatory disclosure; never on avoid-verdicts. Details: https://www.farmdash.one/fees.
## License
MIT — see [LICENSE](LICENSE).
TDQS
Scored across 94 tools
The set has many overlapping tools across intent, swap, simulation, approval, and research lanes, including compatibility aliases and near-duplicate probes. An agent must parse long descriptions to distinguish tools like get_agent_performance vs get_position_health, get_trail_heat vs query_trail_heat, and request_human_approval vs submit_signed_approval.
Most tools use snake_case and a verb_noun pattern, but there are notable deviations such as agent_onboard, ledger_realized_pnl, session_key_status, and inconsistent Trail Heat spellings. Version suffixes like _v2 and legacy/compatibility naming also reduce predictability.
94 tools is far beyond a practical agent-facing tool set and includes many compatibility, preparation-only, feature-gated, or probe tools. Even for a broad DeFi agent OS, this count creates severe selection overhead and suggests many tools should be consolidated or removed.
The surface covers a wide range of domains—research, intents, swaps, perps, risk, sessions, autopilot, workflows, and Virtuals ACP—but several lifecycle paths are intentionally non-functional or preparation-only and therefore dead ends. There are also notable gaps such as perp order reconciliation referenced by cancel_perp_order but not exposed as its own tool, and limited generic intent inspection/cancellation.