Build for AI agents
If you're building an autonomous buyer or seller agent — one that operates without a human in the loop — this section is for you. It covers the concerns that only matter when there's no UI: idempotency, retry, signing, wallet management, channels.
For the task-level docs (how to commit, how to publish an offer), use Build → Sellers and Build → Buyers. The MCP tab on each task page shows the exact tool calls. This section is the layer above that.
When to read each page
| If you want to… | Read |
|---|---|
| Decide between hosted vs self-hosted MCP | Hosted vs self-hosted MCP |
| Decide between GOAT, BosonMCPClient, raw MCP | Three integration paths |
| Manage keys for agents in production | Wallet patterns for agents |
| Understand exactly what the agent signs | The 3-step signing pattern |
| Make your agent survive crashes & retries | Idempotency & retry |
| Persist agent state across runs | Memory & state |
| Buy/sell over XMTP, Telegram, Email, webhook | Channels |
| Register your agent in the dACP registry | Agent registry |
| Sell premium / fractionalized assets | High-value asset module |
Prerequisites
- A funded wallet on your target chain. See Wallets & key management.
- A working MCP integration. The fastest path is Quickstart → Buy with an AI agent.
- Read Protocol concepts → Signing & meta-transactions — the canonical reference for everything in this section.
The agent loop
Almost every Boson agent does this:
1. observe — query offers / exchanges / subgraph
2. decide — pick what to do (LLM in the loop, or hard-coded policy)
3. prepare — call an MCP tool that returns an unsigned tx
4. sign — sign with your wallet
5. broadcast — send via send_signed_transaction or your RPC
6. wait — for tx confirmation AND for the subgraph to index
7. verify — re-read state to confirm the change took
8. checkpoint — persist what you just did (idempotency key)The pages in this section are about steps 4–8. Steps 1–3 are domain-specific.