Agentic crypto accounting is here: no more waitlist

Agentic crypto accounting is now self-serve on Tokenbooks. The Model Context Protocol server that once required an early-preview waitlist now onboards instantly: connect Claude, ChatGPT, or your own agent, approve the OAuth scopes, and it is working your books in a few minutes, free to start, no call required.

Five months ago we shipped an early preview of the Tokenbooks MCP server and onboarded teams one at a time, reviewing each connection request before granting access. That waitlist is gone. Removing it is the easy headline, though. The real change is what "working your books" now covers, and that is what the rest of this post is about. (Engineering-led teams: the same server works from Cursor, Codex, or an agent you build yourself; more on that below.)

A Tuesday morning close, start to finish

Here's a controller's actual morning, compressed into seven exchanges with an agent connected to Tokenbooks.

1. Start with what needs attention, not a blank dashboard.

"What needs my attention before I close September?"

The agent doesn't guess. It calls the same issue-summary the Tokenbooks UI uses and comes back with exact counts: how many are uncategorized, how many have processing errors, how many involve a token already flagged as spam, plus a sample of each so you can see what you're dealing with before you dig in. No dashboard to click through first.

2. Chase the cross-chain thread that never has a clean answer.

"Reconcile our USDC inflows across Base and Arbitrum for Q3, and flag anything that doesn't have a matching leg on the other side."

This is the question that normally eats an afternoon across two block explorers and a spreadsheet. The agent walks both chains' transaction history, matches bridge legs by amount and timing, and answers with every transaction hash, every wallet, and every unmatched leg named; cited, not summarized. The same query works for any chain pair Tokenbooks syncs, not just this one.

3. Turn a recurring judgment call into a rule, safely.

"Every incoming transfer from our payroll wallet should book as a compensation expense. Set that up, but show me what it would have changed for the last 90 days before you save it."

The agent drafts the rule and runs it in preview mode first. No write happens until you see the exact transactions it would touch and approve the shape of the change. Say yes, and it saves the rule and applies it going forward; recurring classification work that used to happen by hand, one transaction at a time, is now a rule you wrote once by describing it in plain English.

4. Drill into a position the way an analyst would.

"Show me our stETH position: cost basis, current quantity, and unrealized gain, broken out by wallet."

Real accounting depth, not a price-chart summary: lot-level cost basis, wallet-by-wallet quantities, and unrealized gain computed off the same fair market value resolution the ledger uses everywhere else. The agent answers in the same accounting vocabulary a controller already thinks in.

5. Sync, then generate the artifact your auditor asked for.

"Run an incremental sync on the treasury wallets, wait for it to finish, then generate a balance sheet as of quarter close."

The agent triggers the sync, polls until it completes, and kicks off the report export: the same PDF, XLSX, or JSON your team already generates from the UI, just requested in one sentence instead of four clicks and a wait.

6. Reconcile bank activity against on-chain activity, in the same thread.

"Do our USD transfers to the payroll wallet this month match what's showing up on-chain, and is anything missing on either side?"

Bank statements come into Tokenbooks the same way on-chain activity does: as source transactions in the same ledger. The agent queries both sides with the same tools it used for the cross-chain question above, and answers with the same rigor: cited, not summarized, wherever the money actually moved.

7. Clean up the noise before it distorts anything.

"Are there any spam tokens showing up in our transaction history?"

A couple identified, flagged, and excluded from the books: the kind of hygiene task that's easy to keep deferring until it quietly skews a report.

Seven questions, and zero context-switching between explorers, bank portals, spreadsheets, and the UI. Every answer traceable back to a real transaction, and every write (the rule, the classification) held for your sign-off before it touched the ledger.

None of those seven is a standalone feature. Triage tells you where to look; reconciliation and position drill-down tell you what happened, on-chain or off; rules and reporting are how you act on it and prove you did. Treated separately, that's seven tabs and seven mental context switches per question. Treated as one conversation with one agent that already holds the full ledger in view, it's a single thread you can follow from "what needs attention" to "here's the filed report" without losing your place.

Why agentic crypto accounting isn't a chatbot bolted onto a dashboard

Most "AI accounting" products in 2026 are a chat window that can see your screen. It's pattern-matching on what's rendered, which means it can hallucinate a number that looks plausible and is wrong, and you'd have no way to know from inside the chat.

Tokenbooks built the inverse: a real ledger and a real audit trail first, with an agent surface on top of it, not stapled to the front of it. Mechanically, the agent calls defined tools, such as "search transactions" or "draft a rule," rather than reading pixels off a screen, and each tool call runs against the same database and the same permission checks a REST API request from the Tokenbooks UI would hit. The agent has no side channel into your data that a human user of the product doesn't also have.

Four things make that concrete rather than a marketing line, and each one is a real constraint enforced in the platform, not just a promise made in prose.

No key access, ever. Tokenbooks reads public on-chain data by wallet address. We never ask for, store, or use private keys, seed phrases, or signing access. The agent inherits that same constraint: it cannot move a single token on-chain, in any scenario, with any permission grant.

Every tool call is an audited operation. There's no model-generated query and no scraped screenshot standing in for real data. Every answer comes from the same ledger a controller would query by hand, and every write goes through the same review path that write already used.

Nothing posts without sign-off. Read tools answer immediately. Write and destructive operations are separately scoped and always prompt your client for explicit approval before they run, so a draft classification never becomes a posted journal entry on its own.

Workspace-scoped, revocable. Tokens carry exactly the permissions your team grants: read-only, sync, reports, or full write. Revoke a token and the agent's access ends immediately, the same way revoking a teammate's access does.

The agent accelerates the close. It doesn't get to run it unsupervised.

How agentic crypto accounting fits the close your team already runs

The MCP surface mirrors the same workflow the Tokenbooks UI already supports, at every stage of the month, on-chain and bank activity alike.

Before close, it's the reconciliation drill-downs from the walkthrough above: cross-chain matching, bank-to-wallet matching, position drill-down, and the issue-summary triage that tells you exactly what needs a decision before you start. Mid-close, it's classification: the agent proposes a rule or a reclassification, you approve the exact set of transactions it would touch, and it applies once. Post-close, it's reporting and the audit trail: export a balance sheet, a P&L, or a cost-basis ledger, and every action the agent took during the month is already sitting in the same log your auditors read.

None of that requires a separate login or a second source of truth. The agent works from the exact ledger the UI renders, so a number the agent gives you in chat and a number your controller pulls up in the app are, by construction, the same number.

Picture a typical week. Monday morning, you ask what changed since Friday and get a short list of new uncategorized transactions instead of opening the app to look. Midweek, one of them turns out to be a new counterparty worth a rule, so you describe the rule once and it handles every later transaction from that counterparty without you touching each one by hand. Friday, close day, you ask for the balance sheet and the P&L back to back, and both come from the same sync that already ran Monday's triage. The agent didn't replace any of those three moments; it just removed the separate places you used to have to go to make them happen.

The setup cost that used to justify a waitlist is gone too. There's no integration project, no data export to review, and no separate credentials to manage. Your workspace, your wallets, your bank statements, and your existing permission model are already there; connecting an agent is closer to adding a teammate than standing up a new system.

Try it today

1. Connect your client from the AI-Native Accounting page: the endpoint, the OAuth flow, and the full tool list are all there.
2. Start on the free plan. Sync a wallet, ask one of the questions above, and see what comes back before you talk to anyone.
3. Questions or an edge case you want us to cover? Book a 30-minute walkthrough and we'll work through your actual books together.

Connect your agent and try agentic crypto accounting now.