Let the assistant draft while your team decides

Your team gets drafts filled from the page it is on and the documents it drops, and a member checks and signs each one unless an owner has let an agent act inside its on-chain mandate. Term sheets, written policies and subscription forms stop being re-keyed into controls, and a coupon nobody funded is found days before its payment date. The assistant drafts in the controls your people already use, standing checks prepare the next step, and an agent makes only the calls its on-chain mandate lists, with transfer and approval amounts capped until the mandate expires. The mandate does not cap the size of a mint or burn.

Before and after#

Before 3.2: someone reads a term sheet and types its coupon, dates and ISIN into the issuance form, then someone else checks it against the PDF. A compliance policy written in a memo is turned into token rules by hand, and a treasury shortfall surfaces when the payout fails.

With 3.2: A member asks from the page or drops the document, and the matching control opens with each value filled in and quoted from its page. Watchers post the next step to the Actions inbox of the people who own it. One watcher can be allowed to move funds on its own, and then only through an agent whose mandate the chain enforces.

Ask about the page you are on and get a draft to sign#

A member asks from the command palette about the asset, participant, organisation, trading venue, audit trail or filings on screen. The answer cites the read behind each figure, and the Console shows the value from that read. A draft opens the matching control already filled in, for the member to check and sign. Organisation owners and administrators manage model access: the assistant runs on the organisation's own OpenRouter key, or on the platform key where the operator set one and the organisation has none, and six switches turn its capabilities on or off for the whole organisation.

ENEnglishARArabic

, opens the full size screenshot

An Asset manager on a bond's page asks what changed last month. The assistant reads the bond's period summary and answers with the transfer volume, the net supply change and the active holders, each compared with the month before, and names the largest movers by handle. The member asks for a mint to a named investor, and the mint control opens filled in, waiting for the member's wallet verification.

The assistant calls models only through OpenRouter, which picks the model provider and region, and each request asks for zero-data-retention endpoints only. In the question and the records the assistant reads, holder names, emails and wallet labels are replaced with handles. Document pages and memos go as written, names included, until an owner or administrator chooses Mask, which keeps them in DALP. DALP cannot check what a provider keeps.

Turn documents into rules and drafts#

A token governor can write the asset's compliance policy in words on the Compliance tab and read it back as rules. A term sheet dropped on Issue asset fills the fields it states, and Draft from a document in the command palette reads a subscription agreement into the mint control, a bank statement into matches for parked bank credits, an identity document set into a verification, and a regulator or board letter into a new filing obligation. The assistant also drafts holder notices for a cash dividend or a distribution, and compares each term sheet on an asset's Documents tab with what the chain holds.

A Compliance officer who holds the Governance role on the bond selects Write the policy and enters “Only professional investors resident in Germany or Austria, with a 12-month lock-up.” “Read the policy” lists each clause, the rule it became and the words its values came from. A check against the current holders then tests those rules against the 50 largest holders and names each holder who would fail and why.

Open 3 rules opens the first add-rule sheet with its values filled in, and each rule is saved with the usual wallet verification before the next one opens.

The model picks a passage or answers unsure, and code reads every value from it, so a date it cannot parse stays empty with the page to check. Policies are read in English and German, and the term sheet fill and document drafts run only while the organisation sends documents to the model.

Let watchers prepare the next step#

Four watchers run standing checks on DALP data: a dividend, distribution or coupon due within fourteen days that the treasury cannot cover, a filing due with nothing submitted, verifications that run out while an asset's rules still need them, and a bank credit nobody could match. Each stays off until an organisation owner or administrator switches it on. At Suggest, a watcher posts what it finds to the Actions inbox of the people who own that work. At Prepare, it adds a draft or options. The filing and verification watchers stop at Suggest, and the bank-credit watcher stops at Prepare.

The payout watcher finds a coupon due in nine days whose treasury holds too little. It posts an item to the Asset managers and organisation owners that prepares the coupon's Fund schedule control with the shortfall. An Asset manager opens the draft, checks the amount and signs, and the item closes when the treasury covers the payout.

Stop all watchers drops every watcher to Observe at once, where it runs and logs only.

Let one watcher act, only inside an on-chain mandate#

Only the payout watcher can go further than Prepare, and only when the owner raises it. Act with veto window schedules the action and runs it unless a member vetoes it within four hours. Act runs the action and reports it. A raise to Prepare or higher needs the owner's minimum track record over the last 90 days: enough resolved items, and enough of them taken unchanged. Either act level also needs an agent principal. That is a participant whose smart wallet can make only the calls its mandate lists, with per-transaction and per-period limits on the transfers and approvals it makes for each asset, until the date the mandate names.

An Organisation owner selects Create agent, reviews the starter mandate for the bond's payout asset and confirms with wallet verification, then raises the payout watcher to Act with veto window. The next shortfall the watcher finds becomes a transfer of the payout asset to the treasury, scheduled four hours after the item is posted.

Any member who can see the item can select Veto transfer before then. When nobody does, the agent sends the transfer and the item reads “Done by” with the agent's name and the transaction hash.

The chain enforces the limit, and the agent's key cannot change its own mandate: only the organisation's own account can replace or revoke it. “Revoke agent” stops one agent for good, and Halt all agents revokes every live mandate and cancels the transactions the agents have queued. A transfer cannot be undone once the agent has run it, and whether the mandate covers it is checked again when it runs.

Connect Claude, ChatGPT or your own agents to DALP#

An organisation owner switches on the hosted Model Context Protocol (MCP) server, and members sign in from Claude, ChatGPT or another MCP app. The app reads what the member can read, and a change it prepares lands in the member's Actions inbox as a request from that app, where the member confirms it with wallet verification. Each call leaves an audit record that names the app, and an owner can revoke an app for every member. An app can also run a mint, burn, pause, unpause, block or unblock as one of the organisation's agents when it names the agent, with no member signature, and then only for the calls that agent's mandate lists. The mandate caps transfer and approval amounts, so it does not cap the size of a mint or burn.

A member connects ChatGPT and asks for a bond's holders, and the app shows the holders table as the member would see it. The member asks it to prepare a mint, the request appears in the Actions inbox, and the member opens the filled-in mint control and signs.

The DALP command-line tool also runs as an MCP server on the user's own machine, exposes every command and marks reads as read-only. What a connected app reads goes to that app's own provider, outside the assistant's Mask setting, and the command-line server runs writes under the signed-in user's API key without a prompt, so your own policy has someone review a command first.

Upgrading#

The assistant is on for every organisation and answers once a key exists: the organisation's own, or the platform key the operator sets. Documents are sent to the model by default, so an owner or administrator who wants them kept in DALP chooses Mask in the assistant settings. Setting the platform key also turns on Ask AI, which answers questions on the documentation site, for anyone who can reach it, billed to that key, unless the operator keeps it off or gives it a separate key. Watchers stay off, the hosted MCP server stays off until an owner switches it on, and no agent exists until an owner creates one. Decide on Ask AI describes the operator's choice.

Why it matters#

• Banks: the assistant reads only what the member may read, drafts go through the same controls and signatures, and “Mask” keeps documents and memos out of what the assistant sends to its model provider.
• Regulators: each agent action names the agent and its transaction, and an agent's permitted calls and its transfer and approval limits sit in a mandate on chain that its own key cannot change. The mandate does not cap the size of a mint or burn.
• Partners: Claude, ChatGPT and your own agents connect through MCP and the command-line tool under the member's own access.

Next steps#

• Connect an AI app →
• Set up the assistant →
• Understand agent mandates →