What's the difference between an agent skill and an MCP server?

A skill is the packaged capability itself — the tool definition, instructions, and code for a job. An MCP server is one standard way to deliver tools to a model: it runs as its own process and exposes capabilities the agent discovers at runtime. You can serve a skill over MCP, but the skill is the capability and MCP is the transport.

Is an agent skill the same as an SDK?

No. An SDK is the developer-facing code library you build with; a human writes the integration. A skill is that capability packaged so an agent can load and call it on its own, bundling the tool definition and instructions alongside the code. Skills are often built on top of SDKs like GOAT or Solana Agent Kit.

Where do I get agent skills, and what is a skill repo?

A skill repo (or plugin library) is a maintained collection of ready-made capabilities builders pull from instead of writing each integration by hand. For onchain agents, common sources include GOAT SDK, Solana Agent Kit, and the plugin system in ElizaOS. Browse more in the directory.

How do skills compose into an agent?

You load the skills a job needs, the framework registers each as a callable tool, and the model chains them at runtime — read state, decide, act, report — without you hand-coding the sequence. Because skills are modular, the same swap skill can serve a trading agent, a treasury agent, or a payments agent.

Can I trust a third-party skill with my agent's wallet?

Treat it as third-party code running with your agent's authority. A skill that can sign transactions can move funds, so read the code, grant the minimum permissions it declares, and pair it with wallet-level guardrails like spend limits. Popularity is a signal, not a security review — claims are unproven until the evidence backs them.

← Back to wiki

What Are Agent Skills?

Last updated 2026-06-24

Agent skills are packaged, reusable capabilities (code, prompts, tool definitions, and workflows) that give an agent the ability to perform a specific kind of task. Skill repos are libraries of these capabilities that builders can compose.

Key takeaways

▸A skill is the *unit of capability* an agent loads to do one kind of task — bundling a tool definition, instructions, and the code behind it.
▸Skill ≠ [MCP](/wiki/what-are-mcps) server ≠ SDK: the SDK is the code you build with, the skill is the packaged capability, MCP is one standard way to expose it to the model.
▸Skill repos (plugin libraries like GOAT or Solana Agent Kit) let builders compose capabilities instead of writing every integration from scratch.
▸For onchain agents, skills are increasingly the unit of capability — and the thing a deployer scopes, since a skill can carry wallet-signing power.
▸A skill being popular is not the same as a skill being safe. Read the code before you give it keys — claims are claims until the [evidence](/sato-score) backs them.

An agent skill is a packaged, reusable capability you hand to an AI agent so it can do a specific job — execute a token swap, mint an NFT, summarize a governance proposal, post to Discord. A skill bundles everything the model needs for that job in one place: a tool or function definition, the instructions for using it, and the code that runs underneath. Drop the skill into a framework and the agent can suddenly do the thing.

The word gets used loosely, so this page draws the lines that matter. A skill is not the same as an MCP server, and it is not the same as an SDK — they sit at different layers of the stack. We cover what a skill actually is, where skill repos come from, how skills compose, and the real examples builders reach for when wiring up an onchain agent.

Why It Matters

Skills turn agent development from bespoke engineering into composition. An onchain agent doesn't need a custom-built swap integration if a maintained skill library already provides one. For the agent economy, skills are the unit of capability, and increasingly the unit of monetization.

How It Works

▸A skill bundles everything needed for a capability: tool/function definitions, instructions, and supporting code.
▸Agent frameworks load skills at build time or runtime, exposing them to the model as callable tools.
▸Onchain skill libraries (like GOAT or Solana Agent Kit) wrap protocol interactions (swaps, staking, minting, transfers) into ready-to-use actions.
▸Well-designed skills declare their permission requirements so deployers can scope risk.

Key Components

•Tool definitions (what the agent can call)
•Instructions and prompt context
•Implementation code (SDK wrappers, API calls)
•Permission requirements
•Documentation and examples

Skill vs. MCP server vs. SDK

These three get blurred constantly, so be precise. They live at different layers and answer different questions.

▸SDK — the *code you build with*. A software development kit is a developer-facing library: functions, types, and helpers you write against. GOAT and Solana Agent Kit ship as SDKs. An SDK by itself doesn't make an agent do anything; a human writes the integration.
▸Skill — the *packaged capability* an agent loads. A skill wraps a job (often built on top of an SDK) into a self-contained unit the model can pick up and call: the tool definition, the instructions for when to use it, and the code behind it. Anthropic formalized this with Agent Skills — folders of instructions and scripts an agent loads on demand. In crypto frameworks the same idea shows up as *plugins* or *actions*.
▸MCP server — one *standard way to deliver* skills to a model. The Model Context Protocol is an open wire protocol: an MCP server runs as its own process and exposes tools the agent discovers at runtime, model-agnostic and out-of-process.

The clean way to hold it: you might write a capability with an SDK, package it as a skill, and serve it over MCP. Same capability, three different layers.

Skill repos and how skills compose

A skill repo (or plugin library) is a maintained collection of these capabilities that builders pull from instead of writing every integration by hand. Think of it as a parts bin: a swap action here, a staking action there, a social-posting action over there.

Composition is the whole point. Skills are designed to stack:

Pick the skills the job needs — a treasury agent might load a swap skill, a balance-read skill, and a scheduling skill.
The framework registers them as tools the model can call, each with its own definition and instructions.
The model chains them at runtime — read balances, decide, sign a swap, report — without you hand-coding the sequence.

Because skills are modular, the same swap skill can serve a trading agent, a treasury agent, and a payments agent. That reuse is why skill repos matter: capability becomes a shared library, not a one-off build. Browse what's available across the directory before writing your own.

Onchain skills in practice: GOAT, Solana Agent Kit, ElizaOS

Three of the most-reached-for onchain skill sources, and what they package:

▸[GOAT SDK](/resources/goat-sdk) — a framework-agnostic plugin library from Crossmint that wraps onchain actions (swaps, transfers, token deploys, and more) across many chains, designed to plug into different agent frameworks. Open source on GitHub.
▸[Solana Agent Kit](/resources/solana-agent-kit) — a Solana-native toolkit from SendAI that bundles protocol actions (token operations, staking, trading, NFT actions) into skills an agent can call. Open source on GitHub.
▸[ElizaOS](/resources/elizaos) — an agent framework with a plugin system where capabilities ship as plugins (its flavor of skills), including onchain ones, that you compose into a character. Open source on GitHub.

Notice the pattern: each exposes protocol interactions as ready-to-use actions so the model calls swap instead of you wiring an RPC client and an ABI by hand. The capability is shown by the open code, not just described in a README.

Skills, permissions, and the trust gap

A skill is also an attack surface. The moment a skill can sign a transaction, it carries the power to move funds — so a skill is exactly the kind of thing a deployer should scope tightly.

Two failure modes show up most:

▸Over-permissive skills. A skill that requests broad signing power widens the agent's blast radius. Well-designed skills declare the permissions they need so you can grant the minimum — pair this with the controls in how agents use wallets (spend limits, scoped signers).
▸Unmaintained or untrusted skills. Skills break silently when a protocol upgrades, and a third-party skill is third-party code running with your agent's authority. Popularity is not a security review.

The rule that holds across the stack: a skill being widely used is a signal, not proof. Read the code, check that it's open and active, and treat any "safe" label as unproven until the receipts are there — which is exactly what the Sato Score grades (transparency and liveness, not a safety or returns guarantee).

Examples

▸GOAT SDK: cross-chain onchain action skills for any agent framework.
▸Solana Agent Kit: Solana-native protocol skills.
▸Custom skill: 'summarize new governance proposals from protocol X and post to Discord.'

Risks & Limitations

⚠Unmaintained skills break silently when protocols upgrade.
⚠Over-permissive skills expand an agent's blast radius.
⚠Quality varies widely, review code before giving a skill signing power.

Frequently Asked Questions

What's the difference between an agent skill and an MCP server?: A skill is the packaged capability itself — the tool definition, instructions, and code for a job. An MCP server is one standard way to *deliver* tools to a model: it runs as its own process and exposes capabilities the agent discovers at runtime. You can serve a skill over MCP, but the skill is the capability and MCP is the transport.
Is an agent skill the same as an SDK?: No. An SDK is the developer-facing code library you build with; a human writes the integration. A skill is that capability packaged so an agent can load and call it on its own, bundling the tool definition and instructions alongside the code. Skills are often built on top of SDKs like GOAT or Solana Agent Kit.
Where do I get agent skills, and what is a skill repo?: A skill repo (or plugin library) is a maintained collection of ready-made capabilities builders pull from instead of writing each integration by hand. For onchain agents, common sources include GOAT SDK, Solana Agent Kit, and the plugin system in ElizaOS. Browse more in the directory.
How do skills compose into an agent?: You load the skills a job needs, the framework registers each as a callable tool, and the model chains them at runtime — read state, decide, act, report — without you hand-coding the sequence. Because skills are modular, the same swap skill can serve a trading agent, a treasury agent, or a payments agent.
Can I trust a third-party skill with my agent's wallet?: Treat it as third-party code running with your agent's authority. A skill that can sign transactions can move funds, so read the code, grant the minimum permissions it declares, and pair it with wallet-level guardrails like spend limits. Popularity is a signal, not a security review — claims are unproven until the evidence backs them.