Agents
AI agents on Topaz
Topaz brings AI to DeFi in two ways: the Topaz Agent — a live strategist that reasons about your positions and builds transactions — and the Agent Skill — a static package that teaches any AI agent the full Topaz protocol.
The Topaz Agent
The Topaz Agent is a live, agentic service that uses multi-step AI reasoning with domain-specific playbooks to analyze pools, recommend liquidity positions, optimize gauge votes, plan bribe campaigns, and build wallet-ready transactions. It is non-custodial — the agent builds calldata, your wallet signs.
You can interact with the agent through its web interface or programmatically via REST API, MCP (27 tools), Agent-to-Agent protocol, or ERC-8183 on-chain job submission.
- ✓Users who want AI-powered DeFi strategy — LP recommendations, vote optimization, reward claiming, and portfolio analysis.
- ✓Protocols planning liquidity campaigns with budget scheduling and APR projections.
- ✓AI agents that need to connect programmatically via MCP or A2A for autonomous DeFi operations on Topaz.
See the Topaz Agent page for capabilities, integration protocols, and the full API reference.
The Agent Skill
The Topaz Agent Skill is a runtime-neutral package — markdown instructions, protocol references, worked examples, evals, and TypeScript helpers — that gives an AI agent the structured context it needs to act on the Topaz protocol. It is a set of files that any agent can read and follow, not a hosted service.
The canonical source lives at github.com/topazdex/agent-skill. The website mirrors /skill.md and /skill.json dynamically from that repo, so an agent can load the latest version from either source.
Agent vs. Skill — when to use which
| Use case | Use |
|---|---|
| Chat about DeFi strategy, get LP/vote recommendations | Topaz Agent |
| Build a trading bot, dashboard copilot, or automation | Topaz Agent API |
| Connect your own AI agent to Topaz via MCP or A2A | Topaz Agent MCP |
| Teach Claude Code / Cursor / Hermes the Topaz protocol | Agent Skill |
| Build an agent that understands Topaz from scratch | Agent Skill |
| Add Topaz context to project instructions or agent memory | Agent Skill |
Skill loading options
The skill is designed to plug into whatever loading mechanism your agent supports. Pick the one that matches.
- ✓Read from URL — best for browser agents, chat agents, or anything with a web-fetch tool. Point at www.topazdex.com/skill.md and you're done.
- ✓Clone the full package — best for coding agents and CLI agents. Includes references, examples, scripts, and TypeScript helpers.
- ✓Copy into a native skill system — for agents with their own skill, rule, or memory concept, drop the package in and point at the repo for support files.
The Runtimes & Install page walks through each of these per runtime with the exact commands and prompt templates.
Safety, by default
Both the Topaz Agent and the agent skill are designed so agents quote, explain, build, and simulate by default — and only broadcast transactions when the user explicitly asks and signing credentials are configured. Every output is labeled so the user always knows whether anything actually moved.
The full safety model — including output labels and when each is appropriate — is on the Safety Model page.
Continue reading
Topaz Agent →
Live strategist — capabilities, integration protocols, and API reference.
Runtimes & Install →
Per-runtime install paths, prompts, and the auto-detect installer.
Capabilities →
Example prompts grouped by workflow — quote, swap, vote, claim, lock.
Safety Model →
The four output labels, when broadcasting is OK, and how to enforce read-only.