IMPLICATOR
On Friday morning, Ryan Wiggins, Mercury's VP of Product, made the bank terminal sound less like science fiction than a back-office shortcut. The San Francisco fintech had just released a command-line interface that lets developers work with Mercury accounts from a shell, while its hosted MCP server gives Claude and ChatGPT read-only account context after OAuth login. Wiggins' point was narrower than the agent-banking pitch: early users want balances, transaction cleanup, reports and invoices first.
Three days earlier, the Office of the Comptroller of the Currency, the Treasury bureau that charters national banks, gave proposed Mercury Bank, N.A. preliminary conditional approval. That timing matters because Mercury's bet is not that an AI should run the bank account. It is that the dashboard is becoming only one surface for it. A founder, finance lead or agent can query, classify and prepare work against the account before a human opens the old banking app.
Key Takeaways
- Mercury now pairs its banking API with a terminal-native CLI and a read-only MCP server for AI tools.
- Ryan Wiggins says early users mostly build reports, clean transactions, and prepare finance work before payments enter the flow.
- Chase and Bank of America run large internal AI programs but do not expose retail bank accounts through first-party MCP endpoints.
- Ramp and Slash show the wider shift: accounts are becoming permissioned work surfaces for AI tools.
AI-generated summary, reviewed by an editor. More on our AI guidelines.
The terminal gets the first account
Mercury's public pitch is direct. The CLI supports "cards, transactions, accounts, and more" from the terminal and can make payments. The hosted MCP is narrower by design: read-only access, OAuth login and no account data for an LLM until the customer authenticates.
At Mercury, Wiggins said, the CLI began as a tool the company wanted for itself. After customers tried it, the requests moved beyond account lookup. "We heard from a lot of people, hey, can we go do this? Can we go do more?" Wiggins said. MCP gave Mercury, he said, "a structured way" to put its API into agent workflows.
The small detail is the tell. Mercury's install command is not a bank branch slogan. It is curl -fsSL https://cli.mercury.com/install.sh | sh.
The CLI fits an older Mercury habit. In a company blog post on its public demo, Wiggins and CEO Immad Akhund questioned why finance-tool buyers rarely see the product before signup; Mercury says candidates, regulators and policymakers later used that demo to inspect the product directly.
The same page advertises "scoped API tokens," IP allow-listing and separation between read and write actions.
Balance comes before payments
The most useful part of Wiggins' answer was what he did not sell. The market talks about agent payments. He talked about the ATM.
"What is it used for most today?" Wiggins said. "It's kind of really checking your balance." The same pattern is showing up in Mercury's early CLI use, he said. Customers want transaction history, burn-rate calculators, investor memos and category cleanup. "What people really love doing and what we hear so much about is the reports they're building," he said.
His example was not a robot paying rent. It was a pre-seed founder turning balances, spending and revenue into a fundraising packet without opening a spreadsheet. Wiggins cited "a baker in Arizona," music producers in Nashville, AI founders and finance teams using Claude Code. That is the customer profile to watch.
Ramp shows how fast the habit can spread once it lands inside finance work. Its MCP and CLI count went from 275 businesses in January to 2,500-plus; 62 percent of usage involved completion work such as coding transactions, attaching receipts and approving requests. Ramp says 50,000-plus businesses use the broader platform. Slash, at 5,000-plus businesses, lets agents prepare cards, payments, invoices and transaction lookups before a human signs off.
So yes, the new banks are booming, but in a specific corner of the market. Not checking accounts for everyone. Operating desks for people who already treat finance as a workflow. The first AI wave inside banks answered questions, summarized documents and helped employees. The newer financial institutions are building around scoped permissions, audit trails, approval walls and machine-readable tools. Mercury's difference is the charter path: if the San Francisco company completes it, the endpoint would sit nearer to the regulated bank account, not only the software layer around it.
The first agent banker is a bookkeeper.
Get the next shift in AI finance
Strategic AI news from San Francisco. No hype, no "AI will change everything" throat clearing. Just what moved, who won, and why it matters. Daily at 6am PST.
No spam. Unsubscribe anytime.
Chase and BofA are not there yet
The contrast with the largest U.S. banks is sharp. JPMorgan Chase said LLM Suite went "from zero to 200,000 onboarded users within eight months," and J.P. Morgan Payments is working with Mirakl Nexus on agentic commerce. Bank of America said "over 90% of employees use Erica for Employees" in its 2025 disclosure.
None of that is a retail bank account exposed as a first-party MCP endpoint. As of May 1, Chase and Bank of America still show bank apps, chat assistants and merchant-payment infrastructure, not a permissioned tool layer that Claude Code or Cursor can call for a small-business owner.
Capital One is the partial exception. Its Chat Concierge can "take specific actions" for auto dealers and shoppers, but it does not yet make checking accounts programmable for outside agents. Agentic banking is the customer's agent meeting the bank at an authenticated endpoint.
The bank still wants the human
The trust model decides whether this becomes a category or a demo. Wiggins pushed back on the idea that APIs are unsafe. "Our modern lives live on APIs," he said. The control layer starts with the customer. "You can choose the permissions," Wiggins said. "You can whitelist the IP address." The MCP layer cannot initiate payments.
When asked what happens if an agent fires off a $5,000 payment to the wrong recipient at 2 a.m., Wiggins compared it to an employee making the same mistake. "People authorize access," he said. "We believe in giving people the tools to actually run the finances the way they want to, but it's also a human choice." Payment-capable API access, he added, requires setup, administration and intent.
That answer will satisfy builders sooner than regulators. The OCC's April 28 approval letter says Mercury still needs final authorization before opening, FDIC deposit-insurance approval remains under review, and the bank cannot begin banking until pre-opening requirements are met. A charter would not primarily mean Mercury opening storefronts. The letter describes a fully online national bank with its main office in Salt Lake City and no physical branch presence. The bigger change would be control: Mercury could move closer to owning the regulated bank account itself, rather than offering banking services through Choice Financial Group and Column N.A.
Wiggins' longer answer was less about autonomy than ownership. "We're not going to give away our human agency over to things like that," he said. "We believe in building tools for humans, not just like for agents." Agents, in Mercury's view, are "extensions of the humans." He also called becoming a national bank part of Mercury's three-to-five-year path.
The account is still there. The dashboard is what starts to disappear.
Frequently Asked Questions
What did Mercury launch?
Mercury added a command-line interface for account, transaction, payment, invoice, card, and webhook work. It also offers a hosted MCP server that lets AI tools read account data after OAuth login.
Can Mercury MCP move money?
No. Mercury says its hosted MCP server is scoped to read-only access. Write actions such as payments and invoices sit on the API and CLI side, where permissions, tokens, and IP controls matter.
Why does a banking CLI matter?
A CLI lets founders, finance teams, and agents work with banking data from scripts, terminals, and coding tools instead of a dashboard. That changes where finance work starts.
How do Chase and Bank of America compare?
Both banks have large AI programs. JPMorgan Chase has LLM Suite and agentic commerce work; Bank of America has Erica and employee AI tools. Neither exposes a retail account through a public first-party MCP endpoint.
What is the open risk?
The hard question is liability. If an agent prepares or triggers a payment the customer did not intend, banks, users, and regulators still need a clearer rule for consent and responsibility.
AI-generated summary, reviewed by an editor. More on our AI guidelines.
Marcus Schuler
Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: [email protected]
