Docket and Dossier in your AI assistant.

Server address: https://mcp.adaptivesystems.dev/mcp

1. Claude (claude.ai, desktop or mobile): Settings, Connectors, Add custom connector, paste the address, Connect.
2. A consent page opens on our server: a quick human check, then Allow. Untick the write permission if you only want to read.
3. You're back in Claude, with the tools listed below.

We've tested it with Claude. It uses the standard MCP authorization (OAuth with PKCE, client metadata documents and dynamic registration), so other assistants that support remote MCP servers with OAuth should connect the same way, for example Claude Code: claude mcp add --transport http adaptive-systems https://mcp.adaptivesystems.dev/mcp.

Things to ask: “read the tampered sample invoice and tell me why it was stopped”, “check a Dossier pack with the edited amount forgery and explain the finding”, “what's my audit trail?”.

sandbox_statusno document

Your remaining allowance, the sandbox limits, the forgeries you can ask for and your recent calls.

docket_extract1 document

Read and check an invoice: a clean sample, a tampered sample or your own file (PDF, PNG, JPEG or WEBP, up to 5 MB and 15 pages). Returns every field with where it was read and whether code confirmed it, the checks, and the risk decision with its reasons.

docket_risk_checkno document

The risk gate's decision for a document already read: decision, score, reason codes, gates.

docket_approve_supplierno document

Record a person's decision in the sandbox: approve the supplier and its bank account. Code re-checks it, no model involved. Needs the write permission.

dossier_check_pack3 documents

Build a synthetic applicant (bank statement, utility bill, passport) with one forgery you choose, or none, and run it through Dossier: outcome, findings, the consistency matrix and the second opinion.

explainno document

Why a document came out as it did; without an id, your own audit trail.

No tool takes free text that reaches a model: only ids, fixed choices, a file name and a file. The only text a model reads is the document itself, which is treated as data, never as instructions.

Per person: 6 documents an hour and 20 a day, 30 tool calls a minute; for everyone together, 100 documents a day. A new person passes a human check on the consent page, and one network can add only a few new people a day. These limits are separate from the try-it pages, so connector traffic never uses up the website's demos.

Everything runs in a sandbox session of your own, erased with everything read from it 30 minutes after it opens. Documents are read through model providers that agree not to store or train on them, and they never reach our real data or statistics. Please don't upload a real passport or ID card, or anyone else's documents.

For each call we keep an audit row: which tool, the size of what was sent (never the content), whether it worked, how long it took and the id it produced. See the privacy page.

Docket reads invoices and stops fraud and duplicates before payment. Dossier checks applicant documents for edits and inconsistencies. Both work the same way in your own systems, behind your login; talk to us about a connector for your own data.