TL;DR: AI agents that operate across DeFi protocols need to bridge tokens between chains. Across gives them two integration paths: a Skills CLI that drops bridging capabilities into coding agents like Claude Code, Codex, and Cursor, and an MCP Server that gives any MCP-compatible client live access to Across documentation, supported chains, and fee data.
The first generation of AI agents mostly talked. They summarized, they suggested, they drafted emails. The next generation is going to do things. Specifically, onchain things. And the moment an agent tries to execute a DeFi strategy that spans more than one network, it discovers the same problem every human bridger already knows: getting tokens from chain A to chain B is way harder than it should be.
Agents can't click through a bridge UI and wait. They need an API that takes a request and returns a result. Preferably in one call.
Why AI Agents Need Crosschain Bridges
AI agents operating autonomously across DeFi need crosschain bridges because protocols, liquidity, and yield opportunities are distributed across dozens of networks. An agent can't act on an Arbitrum opportunity if its assets are sitting on Ethereum.
Most bridges were designed for humans staring at browser windows. Multi-step flows. Manual claim transactions. Challenge periods where you wait and hope. Ambiguous error states that a human can puzzle through but an LLM will retry in a loop until it hits a rate limit.
Agents need something simpler. Send tokens here, receive tokens there, confirm it worked. One input, one output, move on to the next action.
Why Across Fits AI Agents
Across's Swap API reduces crosschain bridging to a single REST call that returns ready-to-execute calldata. That's the exact interface pattern AI agents handle well: structured input, deterministic output, no branching logic.
Three reasons this matters that aren't obvious until you've actually tried to make an agent bridge tokens.
LLMs can reason about it. The Swap API's interface is small enough that a language model can construct the request correctly. Origin chain, destination chain, input token, output token, amount, depositor. No decision tree about which settlement mechanism to use, which route is cheapest, or whether to split across multiple paths. The API makes that decision. The agent just calls it.
The security model is passive. An autonomous agent can't evaluate whether a bridge is trustworthy. It can't read an audit report and make a judgment call. It relies entirely on the infrastructure being safe by default. Across has processed $35B+ with zero exploits across its entire history. The agent doesn't appreciate the feature. The agent's developer will sleep better knowing about it.
The docs are machine-readable. Across publishes llms.txt endpoints and an AGENTS.md template specifically for agents, plus an MCP server that exposes docs, chain data, and live fee queries as structured tools. When a coding agent needs to figure out how to add crosschain bridging, the information is in a format it can actually use.
Two Ways to Integrate
Two officially supported paths. They solve different problems. Use both together or pick whichever fits your workflow.
Skills CLI
The Skills CLI installs Across bridging capabilities directly into AI coding environments with a single command.
npx skills add https://github.com/across-protocol/skills --yes
Compatible with Claude Code, Codex, Cursor, and Openclaw.
After installation, your coding assistant can access bridge functions through natural language. Ask it to fetch a quote, construct a bridge transaction, or wire up a crosschain swap flow. The agent gets structured tool definitions that map directly to the Swap API: chain discovery, quote generation, approval and bridge transaction helpers, and status tracking.
Best for: developers building agents interactively with AI coding tools, or production agent architectures where the agent itself needs to execute bridging as part of a larger workflow.
MCP Server
The Across MCP Server at https://mcp.across.to/mcp gives any MCP-compatible client live access to Across's documentation, supported chains, and bridge fees.
MCP (Model Context Protocol) is becoming the standard way agent frameworks connect to external tools. The Across MCP Server works with Claude Desktop, Claude Code, Cursor, VS Code Copilot, Windsurf, and Codex.
Worth understanding what this server does and doesn't do. It exposes seven tools:
search_across_docs: full-text search across Across documentationget_page: retrieve a complete doc page by pathget_api_reference: browse REST endpoints and their parametersget_supported_chains: list mainnet and testnet chains with IDsget_bridge_fees: query live bridge feesget_code_examples: get SDK and integration code samplesrecrawl_docs: force a documentation refresh
The MCP server is a discovery and context layer. It lets your agent look things up, reason about integration patterns, and fetch live fee data. Actual bridging execution still happens through the Swap API, and the MCP server helps the agent understand how to make those calls correctly. Pair it with the Skills CLI if you want execution too.
Best for: giving coding agents and chat clients up-to-date Across context without baking it into a prompt.
What Agents Can Do with Across
With Across integrated, agents can bridge tokens, fetch live quotes, execute crosschain swaps, and chain these actions into larger onchain workflows.
What this looks like in practice:
Rebalancing. Agent spots yield on Base, bridges USDC from Arbitrum, deposits into the position.
Crosschain arbitrage. Agent monitors price gaps across chains and bridges to capture them before they close.
Treasury ops. Agent distributes protocol funds across chains based on where capital is most productive.
User-facing actions. "Bridge 500 USDC to Optimism" as a natural language command to a chatbot that actually executes it.
Fair warning: these are possibilities, not production case studies. Nobody's published a paper on their autonomous bridge agent yet. The infrastructure is live, the integration paths work, and the early builders are experimenting. What comes out of that experimentation is the part nobody can predict, which is also the part worth paying attention to.
Frequently Asked Questions
Can AI agents bridge tokens automatically using Across?
Yes. The Across Swap API accepts a standard REST request and returns ready-to-execute calldata for crosschain swaps. Agents call the API programmatically, submit the resulting transactions, and confirm the fill via the /deposit/status endpoint. Integration is available through the Skills CLI (which gives agents bridging capabilities) and the MCP Server (which gives agents live Across context).
Which AI tools work with Across?
The Skills CLI (npx skills add https://github.com/across-protocol/skills --yes) supports Claude Code, Codex, Cursor, and Openclaw. The MCP Server at https://mcp.across.to/mcp works with Claude Desktop, Claude Code, Cursor, VS Code Copilot, Windsurf, and Codex.
Is Across safe for autonomous agent transactions?
Across has zero exploits across its entire operating history and over $35B in bridged volume. The intent-based architecture means relayers front their own capital to fill orders, so agent-initiated transactions receive the same security guarantees as human-initiated ones. Settlement bundles are secured by UMA's Optimistic Oracle.
Developer docs: docs.across.to | AI Agents guide: docs.across.to/ai-agents | Bridge: app.across.to