Coding agent skills for onchain development are packaged instructions that teach an agent such as Claude Code, Codex or Cursor how to write, test and ship blockchain code: which language features exist, which contract addresses to use on each chain, and which mistakes to avoid. The open Agent Skills specification defines a skill as a directory containing at minimum a SKILL.md file with YAML frontmatter followed by Markdown instructions (Agent Skills specification). Skills are one of three ways to give a coding agent onchain capability, alongside MCP tool servers and documentation servers. This page explains how the three differ, what onchain skills contain, how to install them, and what skills cannot enforce.

Last reviewed: Nov 2026. Skill names and install paths are quoted from each project's documentation.

What Is a Skill vs an MCP Server?

A skill is knowledge: written instructions and reference files the agent reads when a task calls for them. An MCP server is capability: tools the agent can call to do something, such as compile a program or build a transaction. A documentation MCP server sits between them, giving the agent search over reference docs without any power to act.

The Agent Skills specification describes progressive loading. An agent loads only each skill's name and description at startup, about 100 tokens per skill, loads the full SKILL.md body when the skill is activated, with under 5,000 tokens recommended, and pulls supporting files only as needed (Agent Skills specification). That design keeps many skills installed without crowding the agent's context.

MCP is the other half. The Model Context Protocol is an open-source standard for connecting AI applications to external systems (MCP documentation). Eco's documentation draws the same line for its own tools: the Eco docs MCP server searches documentation, while the Sauce MCP server runs development tools that compile, simulate and build transactions (Eco docs: Build with AI agents).

Type	Gives the agent	Can move funds?	Eco example	Sources
Skill	Instructions, language rules, chain and protocol references	No	sauce-script, sauce-chains, eco-routes-api	Eco docs: Sauce for AI agents
Docs MCP server	Documentation search	No	eco-docs at docs.eco.com/mcp	Eco docs
Tool MCP server	Compile, simulate, build; optionally sign	Only if wallet tools are enabled	Sauce MCP (sauce mcp)	Eco docs

What Onchain Skills Contain

Onchain skills usually cover four things an agent gets wrong without help: the syntax and limits of the target language, the token and contract addresses on each chain, the quirks of specific protocols, and complete worked examples. The Sauce SDK's skill set is a concrete example of that split.

Eco's documentation lists thirteen Sauce skills. sauce-overview is the starting point and maps the others; sauce-script covers the language, its program forms, contract calls, failures and every builtin; sauce-globals lists token symbols and protocol namespaces; sauce-chains lists each chain's tokens and deployments; sauce-protocols covers each protocol's names, ABI and pitfalls; and there are skills for swaps, distributed swaps on EVM and Solana, Solana accounts, the CLI, TypeScript use and examples (Eco docs: Sauce for AI agents). The docs state the skills follow the Agent Skills format.

Eco's documentation skills serve a different task. Installed with npx skills add pointed at docs.eco.com, they include eco-routes-api for quote requests, verification, funding and status; eco-smart-deposit-addresses; eco-solver; and eco-sauce, which points the agent to the Sauce MCP server and skills (Eco docs). Those skills help an agent build an application that calls an API. The Sauce skills help it write programs that run onchain.

Install Skills in Claude Code, Cursor and Codex

Skills are installed by copying their directories into the folder each agent reads. Claude Code reads project skills from .claude/skills, and Codex, Cursor, Copilot and Gemini CLI read from .agents/skills, according to Eco's Sauce documentation. Some MCP servers also serve their skills directly, so nothing needs to be installed in the project.

The sauce skills install command writes the Sauce skills to .claude/skills for Claude Code and to .agents/skills for Codex, Cursor, Copilot and Gemini CLI (Eco docs). The Sauce MCP server also exposes the same skills through its sauce_skills and sauce_skill tools and as sauce://skills/ resources, so an agent connected to the server can read them without any files in the project. A sauce_write_program prompt bundles the workflow: read the overview, language and CLI skills, look up names, compile, simulate, check the Pot and build the unsigned transaction.

For Eco's documentation, Claude Code users can add the docs MCP server with one command, claude mcp add --transport http eco-docs https://docs.eco.com/mcp, which provides documentation search and does not submit transfers or provide wallet credentials (Eco docs). Eco's Claude Code vs Cursor vs Codex comparison covers how the three agents differ for onchain work.

task: pay 1 USDC on Base if the balance allows
  |
  |-- skill (sauce-script, sauce-chains): how to write it, which USDC address
  |-- docs MCP (eco-docs):                 look up an API detail if needed
  |-- tool MCP (sauce_compile, sauce_simulate, sauce_tx): do it, unsigned
  v
person or wallet policy signs

What Skills Cannot Enforce

Skills shape what an agent writes; they do not constrain what it can do. A skill can tell an agent to add a minimum-output check or to never sign without approval, but the agent may skip the instruction, and text from elsewhere can override it. Limits need to live in tool permissions, wallet policy and the transaction itself.

Eco's security documentation spells out the gap for agents that run as the user: such an agent can read wallet files and password sources, sign with tools that apply no spending policy, and edit policy or ledger files (Eco docs: Security considerations). No skill changes that. The controls that do are a server started without wallet tools, signing in a separate wallet, and program conditions such as a required balance that make the transaction revert when they fail (Eco docs: Sauce introduction). See prompt injection attacks on agents that hold funds and the AI agent security checklist.

Where This Falls Short

Skills have three practical limits for onchain work: they go stale when chains and protocols change, they rely on the agent choosing to load them, and they add no enforcement. Teams should version skills alongside the SDK they describe and re-check addresses before mainnet use.

Staleness. Eco's docs note new documentation skills appear only after the corresponding docs deploy (Eco docs).
Activation. Agents load a skill's body only when they judge it relevant (Agent Skills specification). A vague description means the skill may never load.
No enforcement. Instructions are advice to the model, not a control.

Frequently Asked Questions

These are the common questions about coding agent skills for onchain development.

What fields does a SKILL.md need?

A name of up to 64 characters and a description of up to 1,024 characters (Agent Skills specification).

Do I need both skills and an MCP server?

For writing and testing programs, yes: skills teach the language and the server compiles and simulates. The Sauce server can serve its skills directly (Eco docs).

Can a skill sign transactions?

No. Skills contain instructions and files; signing requires a tool with wallet access.

Where do Codex and Cursor read skills from?

From .agents/skills, per Eco's install documentation (Eco docs).

Sources and Methodology

The skill format is quoted from the Agent Skills specification and the skill lists from Eco's documentation, fetched on November 12, 2026. No skills were installed for this page.

Coding Agent Skills for Onchain Development