What was built
The core system is an MCP server that sales reps talk to in chat. It exposes OSI's sales process as a set of tools, skills and playbooks.
The tool surface covers the work reps do dozens of times a day:
- validate email copy and deliverability through ZeroBounce
- gate outreach against OSI's team data. No rep can email a blocked account or quote an unapproved vendor by accident.
- Pure function helpers handle the fiddly stuff: set the right send date and time based on cadence and time zone of the lead, format phones for VoIP app, fetch the right playbook for the job at hand.
Enrichment runs through ZoomInfo and one for Apollo, each behind a single API key with filters applied before anything reaches the rep. Every rep gets vetted leads without having to pass a 30+ roles into a qualification filter. On demand enrichment ensures credit spendage is in check.
The email composition skill is the piece reps use most. It drafts subject and body, then writes both into HubSpot contact custom fields across seven AI email slots, so the sequence picks up the copy without anyone pasting anything or having to send emails manually.
The interesting part of this project was the skills work. OSI's strongest rep had a personal set of prompts and rules that worked extremely well and did not exist anywhere else. We extracted that set, stripped out everything personal, and turned it into general sales skills every rep can run. One MCP, best practices available to the entire team without people figuring out how they work. It just works.
Governance is deliberate because OSI owns this code. We set up CODEOWNERS, branch protection, GitHub teams and a CONTRIBUTING guide, plus CI quality gates running five checks on every pull request.
Tech stack
- MCP server: Metabrain core, context tracking and skills toolsets enabled
- CRM: HubSpot, including contact custom fields for AI generated email copy
- Enrichment: ZoomInfo and Apollo
- Validation: ZeroBounce plus OSI's internal validation script
- Repo and CI: GitHub
Recently touched
- 2026-10-11 Build automated non-intrusive feedback loop for OSI Sales MCP Done
- 2026-10-10 Role fit: never surface Retired / former people as viable prospects Review
- 2026-10-10 Feedback: bump core to v0.41.0 and classify validator rejections with rule codes Review
- 2026-09-25 OSI Sales in company ChatGPT + offer MCP layer for OSI Sales Dashboard Done
- 2026-09-24 Migrate osi_start onto core register_start_tool: bump metamcp-core v0.4.0 → v0.17.0, keep the rep tutorial, name and tool surface Done
Other projects in this case study
HubSpot Optimization
OSI's HubSpot Optimization project is a large-scale initiative delivering a complete lead enrichment and order processing platform. The system combines multiple data sources (Apollo, Zerobounce,…
Read case studyWebsite Refresh
OSI Global ran on WordPress. We rebuilt their site on Astro — a faster, static site OSI owns outright, with content as Git files editable by their own team or directly through Claude and ChatGPT, a…
Read case studyWhat we built
- Performance score
- Largest Contentful Paint
- First Contentful Paint
- Total Blocking Time
- Cumulative Layout Shift
- Speed Index