Build with a coding agent

A coding agent can scaffold and modify a Lucid service, but generated code must still pass the same build and payment tests as hand-written code.

Install the versioned Lucid Agents skill when your coding client supports portable Agent Skills. It adds a read-only channel inspector and routes the agent to focused SDK references.

Choose the channel first

Tell the coding agent whether it is modifying a Stable npm project or this repository's Next workspace. Include the pinned package versions and adapter. Do not ask it to “use the latest Lucid”; that can combine incompatible public and repository APIs.

Give the agent repository context

Keep the generated AGENTS.md file. It describes the template's public APIs, environment variables, and test commands. Ask the agent to:

1. Preserve the runtime/extension/adapter boundary.
2. Add Zod input and output schemas before implementing business logic.
3. State prices as USD decimal strings.
4. Keep private keys and facilitator credentials server-side.
5. Run typechecking and the unpaid 402 test before reporting completion.
6. Read package-local AGENTS.md instructions and existing tests before changing an API.
7. State protocol/version boundaries instead of claiming generic A2A, AP2, MPP, or ERC-8004 compatibility.

Example request

Add a capability named summarize. It accepts non-empty text and returns a
summary plus usage. Charge $0.02 USDC per invocation over x402. Preserve the
existing Stable package versions and adapter. Add tests that invalid input is
400, an unpaid request is 402 without fulfillment, and a same-key retry does
not run the handler twice. Do not expose secrets or add a second paywall.

Require evidence, not a prose claim

Ask the agent to return:

• changed files and the public behavior each one owns;
• exact commands run and whether each passed;
• the unpaid challenge status/header and paid testnet evidence, when funding is in scope;
• any snippet that remains illustrative rather than compiled;
• release-channel/protocol assumptions and unverified provider dependencies;
• production gaps such as in-memory stores, test keys, or missing readiness.

A useful completion gate is:

bunx tsc --noEmit
bun test
bun run build

In this monorepo, also run the package/example smoke tests and documentation checks required by the root AGENTS.md. Never let the coding agent spend real funds, register an on-chain identity, deploy, publish, or rotate credentials unless that external side effect was explicitly authorized.

Review high-risk changes

Human review should focus on recipient/network changes, wallet authority, price and policy changes, settlement timing, idempotency, task access tokens, discovery URLs, storage durability, log redaction, and adapter middleware order. Generated code is not safer merely because it type-checks.

The open-source workflow does not require the Hosted Lucid Platform. Hosted agent creation is a separate product surface; see Hosted agent creation only if that service is available to your account.

Continue with Add Lucid to an existing app, install the Lucid Agents skill, and use the production checklist.