Operate Sailor with a coding agent

Sailor is designed to be operated through a coding agent — Claude Code, Cursor, Codex, or any AI coding assistant. You describe the strategy you want; the assistant deploys the SMA, authors and tests the permissions that bound it, signs the mandate, and runs the agent. This is the intended way to use Sailor, and it's what the scaffold is built for.

How it works

When you run sailor init, the scaffold includes an AGENTS.md operator guide and a set of on-demand skills under .agents/skills/, following the open Agent Skills standard. Assistants that scan skills load each one only when relevant; assistants that don't follow a routing table in AGENTS.md to the same plain-markdown files. It works in Claude Code, Cursor, Copilot, and Codex.

A representative slice of the scaffolded skills (the full set of seventeen — including one per shared template — is on the Skills page):

Skill	Loaded when the assistant is…
sail-onboarding	setting up a new project, or resuming a partial one
sail-project-info	answering anything about project / account / mandate / chain / environment state
sail-servers	starting, stopping, or health-checking the dashboard or signing station
sail-templates + sail-template-*	reusing a shared permission template (one skill per template)
sail-token-resolve / sail-swap-quote	resolving tokens/liquidity and computing swap slippage floors
sail-transactions	building dispatches or any EVM transaction for the agent
sail-mandates	designing, authoring, testing, deploying, or authorizing custom permission contracts
sail-automation	running the agent unattended (GitHub Actions, self-hosted, Docker, or local daemon)
sail-extend	adding notifications or a custom dashboard once the agent is live

The flow

mkdir my-agent && cd my-agent && npm i @sail.money/sailor && npx sailor init && npm install

Open the folder in your assistant and say start. From there the assistant:

1. Deploys your SMA and creates your agent wallet.

2. Defines your strategy — it treats src/ as a blank slate and asks what you want; it does not assume the example code is your strategy.

3. Builds, tests, and signs your mandate — authoring permission contracts in the Foundry workspace, running forge test and sailor mandate simulate, and only then authorizing on-chain.

4. Runs your agent — locally with sailor run, or on a schedule via the bundled GitHub Actions workflow.

5. Extends — notifications, a custom dashboard.

What the agent is told to do (and not do)

The scaffold's AGENTS.md encodes hard invariants the assistant follows. The ones worth knowing as the operator:

• Owner signing is browser-only. The assistant never puts your owner key in the terminal — owner approvals happen in the signing station UI in your browser.

• Setup asks before spending gas; a running agent does not. Once the mandate is signed, the mandate is the authorization — the agent transacts autonomously within it. You are not asked to confirm each dispatch.

• Never authorize a permission before it passes tests. The assistant must see forge test and sailor mandate simulate pass against samples derived from your strategy before attaching a permission.

• Use the SDK's signing helpers. It signs dispatches with buildDispatchSignature and detects the kernel's dispatch model with detectKernelCapabilities — never hand-rolled EIP-712, never a hardcoded model.

• Example permissions are not a menu. The assistant treats shipped templates as examples to adapt and verify, not audited, drop-in production contracts.

• approve() needs explicit coverage. An ERC-20 approve is not covered by a swap/supply/deposit permission — it's bounded either per-call (separate single dispatches) or as an atomic batch via an IBatchPermission. See Build & register a mandate.

Pointing your agent at these docs

Coding agents read these docs on your behalf. The simplest path: give your agent the ready-made prompt on For AI agents, which points it at the docs' llms-full.txt (the full docs as one machine-readable file) and tells it what to do next.

A typical instruction: "Read the Sail docs' llms.txt, then set up a Sail SMA on Base that lets the agent only swap USDC↔WETH on Uniswap V3 up to 500 USDC per trade." The agent pulls the relevant pages, scaffolds, authors a bounded permission, simulates it, and runs.

The custody guarantee holds regardless of which assistant you use: capital stays in your Safe, the mandate is enforced on-chain on every dispatch, and you can pause or revoke at any time.