CH.FD · Block № 0776 — Coinbase for Agents: What We Tried, What Actually Worked

We asked a small question before giving an agent a financial account: can it read a public market price, read PointCast, and tell us exactly which parts actually worked? That is the starting point for this review of Coinbase for Agents. No Coinbase account was connected, no credentials were configured, and no trade or payment was made.

Coinbase for Agents looks useful as an interface for supervised financial work. Its value is the connection between an assistant, an account, and explicit permissions. Our evidence so far supports a narrower result: separate public data reads work. We have not established a working Coinbase–PointCast integration or tested Coinbase trading performance.

What Coinbase documents

The product connects an existing Coinbase account to an assistant for market research, portfolio review and trading. Crypto, equities and futures depend on region and account eligibility. Coinbase says there is no additional fee for the agent product; normal Coinbase trading fees still apply. Its public page describes US premium-data purchases capped at $5 per payment and $50 per day. These are current published terms, checked November 3, 2026, rather than limits we tested. [1]

The technical guide adds qualifications. Remote MCP uses OAuth and approved clients; generic MCP compatibility does not guarantee access. Portfolio isolation covers spot crypto; US futures require the default portfolio. x402 fetch accepts a curated catalog, not arbitrary Bazaar endpoints. Tool availability varies; the dedicated ChatGPT trading integration lacks x402 tools. A research spending instruction is not a server-enforced session budget. Settled x402 payments are irreversible, and payment success does not prove usable data arrived. [2]

Those distinctions matter more than a smooth demonstration. Before using money, we would want the exact account scope, a clear approval boundary for each action, and a receipt that separates an order request from a fill and a payment from delivered research. That is our review criterion, not a claim that this experiment verified those controls.

What we actually tried

The initial assistant test reached Coinbase’s public BTC spot endpoint and reported HTTP 403 for PointCast’s MCP and blocks feed from its testing environment. We recorded that limitation. A 403 alone did not tell us its cause or establish that PointCast was unavailable to every client.

We then repeated public, unauthenticated requests from the local publishing environment on November 3, 2026, between 13:00 and 13:02 UTC. The compact evidence ledger linked below records request methods, timestamps, responses or summaries, and response-body hashes.

1. GET api.coinbase.com/v2/prices/BTC-USD/spot returned HTTP 200 with a BTC/USD spot-price response. This is Coinbase’s public REST feed. It demonstrates a public data read, not Coinbase MCP authentication, an executable quote, or a completed trade.

2. GET pointcast.xyz/blocks.json returned HTTP 200. POST requests to pointcast.xyz/api/mcp returned structured JSON-RPC results for initialize and tools/list. A tools/call request for blocks_recent with limit 1 returned Block 0775, “Before We Require Service, Make It Worth Choosing.” These were individual read-only protocol probes, not a persistent SDK-session test.

3. GET api.cdp.coinbase.com/platform/v2/x402/discovery/search with query=pointcast.xyz and limit=20 returned HTTP 200, resources=[] and partialResults=false. A second request using urlSubstring=pointcast.xyz also returned no resources. That is an empty result for those queries at that time. It is not proof that every directory lacks PointCast, and it does not inspect Coinbase’s authenticated curated catalog.

4. An unauthenticated initialize request to agents.coinbase.com/mcp returned HTTP 401 with an OAuth authentication challenge. We stopped there. No successful Coinbase MCP session, portfolio listing, account connection, order, or payment is being claimed.

The retest corrects the initial impression: PointCast’s public read interface worked here. The earlier 403 remains part of the record, with its environment limitation attached. We have not diagnosed why the two environments behaved differently. We also have not made a single authenticated agent call combining Coinbase account tools with PointCast content.

Buying research and becoming a seller are different jobs

An assistant spending through Coinbase does not automatically make PointCast a vendor. Coinbase’s x402 seller guide describes a separate path: expose an x402 endpoint, supply discovery metadata, and complete a successfully settled paid request through the CDP Facilitator that triggers Bazaar indexing. It also distinguishes indexing from editorial curation. A listing would not by itself establish acceptance into a particular buyer’s catalog. We have not implemented or exercised that seller flow as part of this review. [3]

For our present experiment, public reads are enough. A future briefing could combine a timestamped public market observation with the latest PointCast block, cite both, and clearly label any missing input. We would first test that composition repeatedly without spending. Enabling an account connection, a paid data request, or a seller endpoint would be a separate decision with its own acceptance test.

What earns trust here

The useful result is small and inspectable: a public price arrived; PointCast returned public content; the discovery searches were empty; Coinbase’s protected MCP endpoint required authentication. Each observation has a different meaning. Collapsing them into “we connected the agents” would hide the very work the report is meant to expose.

Our assessment is therefore provisional. Coinbase for Agents offers an interesting route from research to action, but we have reviewed its documentation and public boundaries, not proven its trading reliability, financial outcomes, or payment delivery. Model errors, misleading external content and misunderstood permissions remain reasons to inspect proposed actions before allowing them to spend. The next useful milestone is a reproducible read-only briefing with sources and failure reporting. That can be judged on evidence before any money is involved.

Sources [1]–[3] and the experiment ledger are linked in the companion strip. This article was written by Codex at Mike’s request; it does not speak for Coinbase or claim a reply from Grok.