Using Codex with a crypto workflow: planning, tools, and approval

Codex is most useful in a crypto workflow when preparation and authority are separate. It can help turn a goal into a file-backed plan, inspect project instructions, read a documented tool result, and report what remains unknown. A wallet and the application still define the transaction the user reviews and approves. The exact tools available depend on the host and its configuration, so check the environment before describing an action as possible.

Start with a transaction brief

Write the job in a small record before asking for tool use:

Goal: prepare a review for swapping SUI to USDC
Network: Sui mainnet
Wallet: [canonical wallet supplied by the user]
Asset identifiers: [full coin types, not symbols]
Amount and budget: [decimal amount and maximum]
Mode: read-only preparation; no quote, build, sign, or send
Evidence needed: source origin, asOf, route, wallet prompt, confirmation

This brief makes an ambiguous request observable. “Swap SUI” omits the wallet, coin types, amount, network, and spending limit. The mode line also creates an approval checkpoint: the assistant may prepare and explain while the user decides whether to continue.

Let files carry the durable rules

Read the project’s AGENTS.md, the relevant SKILL.md, and the local product references before calling a tool. The public participant kit README says the kit contains read tools for status, season, Swap history, Desk, and Partners plus a review-only action planner. Its AGENTS.md template says to preserve pending, confirmed, failed, partial, unavailable, and empty zero states and never infer Points, allocation, or eligibility from incomplete data.

That is more reliable than keeping a rule in a prompt. A project instruction can state the network, permitted origin, budget ceiling, evidence fields, and stop condition. A skill can explain which endpoint fields matter. The report should link the files it actually used and identify any conflict instead of silently choosing a convenient rule.

Know what each tool can do

The local kit documents six MCP tools. The first five make bounded public reads. suibrokers_prepare_action returns a checklist and app route; it does not quote, build, sign, send, activate a referral, boot a wallet, or mint an NFT. The kit also documents the Codex CLI registration form for its remote HTTP server and for a local stdio command. Those commands are configuration examples, not evidence that a particular installation is connected.

OpenAI’s Codex documentation lists the surrounding tool surfaces, including local shell, apply patch, MCP, and skills. The official MCP guide for ChatGPT and Codex is the source for host setup. Read the host’s returned tool schema and result fields rather than assuming that a tool with a familiar name can sign or submit a transaction.

Use one deliberate approval boundary

For a SUIBROKERS Swap review, the safe sequence is:

1. Read the Swap skill and the project instructions.
2. Validate the supplied wallet and full coin types.
3. Request a read-only history or preparation result with a bounded limit.
4. Report the source origin, endpoint, asOf, and state exactly as returned.
5. Open the application route and compare asset, amount, output, fee, gas, minimum received, slippage, and recipient.
6. Let the user review the final wallet transaction and sign explicitly if they choose to proceed.
7. Read the confirmed ledger result before treating activity as verified.

The assistant should stop when a coin type is ambiguous, the service is unavailable, a result is pending, or the app prompt differs from the brief. Do not paste a seed phrase or private key into a prompt, file, MCP configuration, or command. Do not turn an action plan into a claim that a transaction occurred.

Report the work, not a promise

A good Codex report records the files read, tool name, source timestamp, result state, unresolved fields, and the next human decision. For example, “the action planner returned a review-only Swap route; no quote or transaction was produced” is useful evidence. “The Swap is ready” hides the approval and execution steps.

Current Points are credited only after eligible activity is confirmed and verified by the service. A prompt, skill installation, read, simulation, or unconfirmed plan is not a ledger event. Keep that product rule in the report’s evidence section rather than implying that preparation has financial value.

Read MCP connections, AGENTS.md for crypto work, and how to read Swap terms when the workflow moves from planning to a specific review.