Guides · Skills

Agent Skills

SKILL.md skills that teach an agent the multi workflows. Where the MCP server gives an agent the tools, skills give it the know-how — how to combine them to analyze, trade, pair-trade, and swap, with the safety model built in.

Skills pair with the MCP server: the server exposes the tools(get_orderbook, execute_order, swap_quote…); a skill encodes the workflow (preview → confirm → execute, idempotency, which error means what). Each skill is self-contained and works whether the agent has the multi MCP tools or just SDK/HTTP access.

The skills

setup
multi-setup→

Connect to multi — install the MCP server / SDK, mint a scoped multi_sk_ key.

read-only
multi-market-analysis→

Analyze a market/token — price, orderbook, candles, routing, funding. Read-only.

execution
multi-trading→

Place & manage perp trades — smart-routed orders, open/close positions.

read-only
multi-portfolio→

Review the cross-venue portfolio — exposure, PnL, funding, liquidation risk. Read-only.

execution
multi-execution-algos→

Execute with algos — TWAP, scaled ladders, TP/SL brackets, batch orders.

execution
multi-pair-trading→

Market-neutral / relative-value trades — long one asset, short another.

execution
multi-swap→

Swap or bridge across chains (EVM + Solana), custodial or non-custodial.

The execution skills all defer to multi-setup when a key is missing, and multi-market-analysis is read-only — so an agent can reason about a market with no key, then ask for one only when it's time to act.

Install

Claude Code

Add as a plugin from the repo:

bash
/plugin marketplace add <path-to>/multi/skills
/plugin install multi@multi

# ...or copy a single skill into a project:
cp -r skills/skills/multi-trading .claude/skills/

Codex · OpenCode · Hermes · OpenClaw

Copy any skill directory (each skills/<name>/ is portable) into your harness's skills location, and make sure the multi MCP server is configured (see harness setup).

Safety, by design

Every skill encodes the same guardrails so autonomous use stays safe:

  • —Preview → confirm → execute — quote and restate the order before any real trade.
  • —Idempotency — execution carries an Idempotency-Key so retries never double-submit.
  • —Policy-aware — skills read the error codes (EXECUTION_DISABLED, ORDER_TOO_LARGE, DAILY_CAP…) and explain them instead of retrying blindly.
  • —Least privilege — multi-setup grants only the scopes an agent needs.

See the full safety model.