Dr. Schauer | CRM and a conversational Kanzlei stack · Kanzlei MCP

What was built

A team facing MCP layer so the firm's staff move CRM and operational data through ChatGPT instead of clicking through the Pipedrive interface. Only supervisors work in the CRM UI. Everyone else gets the same data conversationally, under the same permissions.

Area	Description
Tool surface	Built on Metawork's MetaPipe, with Pipedrive roles as the single enforcement layer, so access rules are not reimplemented in a second place
Inbound sync	A nightly Pipedrive sync with webhook lead creation, a freshness based conflict resolver and person deduplication. Server syncs between several Pipedrive accounts based on lead source and destination.
Alerting	A grouped nightly digest that separates updated from stopped, plus heartbeat and failure alerts into the shared alert hub
Meeting notes	A meeting store and webhook receiver that ingests transcripts, applies a confidentiality rule per user, and pushes summaries back into Pipedrive
Playbooks	Company's knowledge base and worklfows are available through playbooks. AI knows exactly what each process is, whom to pass information to or assign tasks and how to handle automated worklows flawlessly.

What the team can do in chat

Ask who the open deals belong to, what stage a client sits in, which leads arrived overnight and where a qualification form stalled. Read and write person, organisation and deal records. Pull the notes from a client meeting and get the summary written back as a Pipedrive note with the follow up activities assigned. Reach accounting data through the DATEV bridge instead of exporting files by hand.

Every tool call runs as the user, so a junior sees a junior's data and a supervisor sees everything. That is the whole point of putting Pipedrive roles at the centre rather than building a parallel permission model.

Inbound lead flow

Qualification requests arrive through a web form and land in Pipedrive as leads. The nightly sync reconciles the full dataset against one of the several CRM accounts. The tricky part was staleness: Pipedrive update timestamps move for reasons that have nothing to do with the fields we care about, so the resolver uses a freshness cursor, compares multi enum sets properly and mirrors deal person and organisation links. Person deduplication merged the fully identical name and email duplicates and self heals the sync links afterwards.

Operations

The sync jobs and the webhook receiver report into the alert hub with heartbeats and failure alerts, so a silent night is a detected night. Builds, tests and deploys run through a CI/CD pipeline, and the meeting store for transcripts and summaries is encrypted.

Tech stack

• MCP layer: Metawork's core, exposed to Claude and ChatGPT
• CRM: Pipedrive, roles as the single source of truth
• Sources: CRMs, DATEV, web forms, meeting transcripts
• Accounting: DATEV on premise
• Identity: Entra ID and on premise AD, with a Personio driven identity sync concept
• Delivery: CI/CD pipeline, containerised runtime, Metawork alert hub for heartbeat and failure alerting

Roadmap

The meeting note pipeline is being finished end to end, including the Pipedrive write back of summary notes and assigned activities. After that, the DATEV bridge moves fully behind Kanzlei MCP so the team reaches accounting data in the same conversation as the CRM.