deBridge Agent Skills

Agent Skills for the deBridge cross-chain DeFi ecosystem. Skills teach AI agents how to bridge tokens, swap, sign transactions, query balances, and monitor orders across 20+ EVM chains and Solana.

Skills are the guidance layer; the deBridge MCP server and SDK are the execution layer.

Install

npx skills add debridge-finance/debridge-skills

This installs all deBridge skills into your agent's skill directory. Works with Claude Code, Cursor, Windsurf, Copilot, Cline, and 35+ other agents.

Skills

Skill	Description
common	Shared prerequisite — environment detection, MCP setup, chain config, RPC discovery
bridge	Cross-chain bridge via DLN: quote, preflight, sign, execute, monitor
swap	Same-chain and cross-chain token swaps
signing	Transaction signing — routes to OWS, ethers/viem, cast, MetaMask, or Privy
wallets	Wallet setup: OWS, EOA, Foundry keystore, Privy embedded
analytics	Token prices, balances, TVL, DEX pools, on-chain data via third-party MCPs

Connect the MCP Server

Skills tell the agent what to do; the MCP server exposes the on-chain tools. Install both.

Claude Code

claude mcp add --transport http debridge https://agents.debridge.com/mcp

Cursor / Windsurf / Claude Desktop

Add to your MCP config file:

{
  "mcpServers": {
    "debridge": {
      "type": "streamable-http",
      "url": "https://agents.debridge.com/mcp"
    }
  }
}

Other Agents

See setup/ for per-agent setup guides covering 35+ environments, including Copilot, Cline, Codex, Gemini CLI, VS Code, JetBrains, and more.

Remote Skill Discovery

Agents can also discover skills remotely via progressive disclosure:

https://agents.debridge.com/llms.txt          # llmstxt.org catalog
https://agents.debridge.com/skills/index.json  # structured index

How Skills Work

Each skill directory contains a SKILL.md entry point and sibling reference .md files. Skills follow four patterns:

1. Shared Prerequisite — common runs first, detecting environment, MCP connectivity, and available signers.
2. Router — signing and analytics detect the available tool (OWS, ethers, cast, etc.) and route to the matching reference file.
3. Sequential Pipeline — bridge follows a strict order: quote -> preflight -> sign -> execute -> monitor.
4. MCP Probe with Fallback — every skill that calls MCP tools provides an alternative path when MCP is unavailable.

bridge/SKILL.md -> preflight.md -> ../signing/SKILL.md -> execute -> monitoring.md
                                        |
                              ows-signing.md | sdk-signer.md | foundry-cast.md | metamask.md | privy-mcp.md

Repository Structure

debridge-skills/
├── skills/
│   ├── common/               # Shared prereq: env detection, MCP setup, chain config, RPC discovery
│   ├── bridge/               # Sequenced: quote -> preflight -> sign -> execute -> monitor
│   ├── swap/                 # Same-chain and cross-chain swap flows
│   ├── signing/              # Router: wallet/signer strategy selection
│   ├── wallets/              # Wallet setup: OWS, EOA, Foundry keystore, embedded
│   ├── analytics/            # Balances, prices, TVL, DEX pools, on-chain data
│   └── index.json            # Generated: skill discovery catalog
├── setup/
│   ├── agents/               # Per-agent setup guides (35+ agents)
│   └── wallets/              # Per-wallet setup guides (7 wallets)
├── scripts/
│   ├── validate.ts           # Schema + structural validation
│   ├── build-index.ts        # Generate skills/index.json
│   └── generate-llmstxt.ts   # Generate llms.txt
├── evals/                    # LLM evaluation harness
├── plugin.json               # Plugin descriptor
└── llms.txt                  # Generated: llmstxt.org skill catalog

Development

Prerequisites

npm install

Commands

npm run validate              # Schema + structural checks (228 rules)
npm run validate:verbose      # Validate with passing rules shown
npm run build                 # Full pipeline: validate -> index.json -> llms.txt
npm run build:index           # Generate skills/index.json
npm run build:llmstxt         # Generate llms.txt
npm run check                 # CI: verify generated files are up to date

Creating a New Skill

1. Create a directory under skills/:

   mkdir -p skills/{skill-name}

2. Write SKILL.md with YAML frontmatter (name, description, license, metadata) and a markdown body. The description field is the trigger mechanism — include action verbs, token/chain names, and explicit trigger phrases.

3. Add sibling reference .md files for environment-specific instructions. One concern per file, under 300 lines each. Every reference must be linked from SKILL.md.

4. Validate and build:

   npm run build

SKILL.md Anatomy

---
name: debridge-bridge                    # External ID: debridge- prefix required
description: >
  Execute cross-chain bridge transactions via deBridge DLN...
  Triggers: "bridge", "cross-chain transfer", "move tokens".
license: MIT
metadata:
  author: deBridge
  version: "1.0.0"
---

The body starts with a PREREQUISITE line referencing ../common/SKILL.md, followed by a Quick Reference table, MCP availability check, numbered workflow steps, common errors, and a references list.

Validation Rules

The validator enforces structural, content, and cross-referencing rules:

• Structural: frontmatter fields present, name has debridge- prefix, body under 500 lines, references under 300 lines, no orphan references, no subdirectories.
• Content: no hardcoded addresses/keys/paths, string-canonical amounts only, MCP fallback sections present, signing delegated to signing/SKILL.md, preflight referenced before execute.
• Cross-refs: all references resolve, no circular dependencies, common is prerequisite for every skill.

Writing Guidelines

• Only include knowledge agents don't already have: MCP tool names, SDK signatures, parameter formats, sequencing constraints, chain-specific quirks.
• Use string-canonical amounts everywhere ("100.5", never wei or BigNumber).
• Reference MCP tools as mcp__debridge__<tool_name> alongside SDK equivalents.
• Put "when to use" context in the description frontmatter, not in the body.
• Mark dangerous operations: ⚠️ CAUTION: This executes a real transaction.

Commit and PR Rules

• One skill per PR.
• Run npm run build before committing — both validate and generate must succeed.
• Conventional commits: feat(bridge): add Solana destination support.
• Bump metadata.version in SKILL.md frontmatter on every skill change.
• Never edit llms.txt or skills/index.json directly — regenerate with npm run build.

Evals

Skills are tested with an LLM eval harness in evals/. Each eval case is a prompt + assertions verifying that an agent using the skill produces correct behavior.

npm run eval -- --skill debridge-bridge   # Run evals for a skill
npm run eval -- --all                     # Run all evals

Every skill should have eval cases covering: happy path, MCP-unavailable fallback, error recovery, environment routing, and cross-skill sequencing.

Related

• deBridge MCP Server — live cross-chain infrastructure tools
• Agent Skills Specification — the standard these skills follow
• llms.txt Standard — the format used for llms.txt
• skills.sh — skill discovery platform
• Setup Guides — per-agent and per-wallet setup instructions

License

MIT