---
title: Claude Cowork MCP Connector — Cromus Workflow Intelligence
description: Cromus ships a hosted HTTP MCP connector for Anthropic's Claude Cowork. Exposes six deterministic Workflow Intelligence tools inside any Cowork session via the MCP server at mcp.cromus.ai — no local install required.
canonical: https://cromus.ai/cowork-connector
source_html: https://cromus.ai/cowork-connector
---

# Claude Cowork MCP Connector

> Six deterministic Workflow Intelligence tools — inside any Claude Cowork session. No local install.

Cromus ships a direct HTTP MCP connector for Anthropic's Claude Cowork. Add a custom connector pointing at `https://mcp.cromus.ai/mcp` with a Bearer key and the full Cromus Workflow Intelligence toolkit is available inside your Cowork session.

---

## The six tools

| Tool | What it does |
|------|-------------|
| `score_skill` | Score a SKILL.md file — AI-readiness, completeness, Croms™ waste estimate |
| `simulate_cost` | Project per-run and monthly cost across 5 cost-quality modes and 62+ models |
| `validate_skill` | Validate a SKILL.md file against the open spec (structural + semantic checks) |
| `validate_ethos` | Validate an ETHOS.md behavioral governance file |
| `validate_memory` | Validate a MEMORY.md portable memory spec |
| `reality_check_summary` | Summarize the reality-check loop results for a workflow |

All six tools are **deterministic** — no LLM calls, no external data requests beyond the verified model registry. The same input always returns the same output.

---

## Setup

1. Open Claude Cowork settings → MCP Connectors
2. Add a new connector with:
   - **URL:** `https://mcp.cromus.ai/mcp`
   - **Auth type:** Bearer
   - **Token:** your Cromus API key (available in your workspace settings)
3. Save and reload — the six tools appear in your Cowork tool list

That's it. No local server, no CLI install, no Docker.

---

## No-retention contract

The Cromus MCP server operates under a **no-retention contract**: inputs passed to any of the six tools are used only to compute the response and are not stored, logged for training, or retained after the request completes. This makes it suitable for use with proprietary SOPs and internal workflow documentation.

---

## Authentication: OAuth 2.1 + PKCE

The MCP endpoint supports OAuth 2.1 + PKCE for team deployments alongside Bearer token auth for individual API access. Both methods are described in the workspace settings.

---

## Related pages

- [OpenAI Codex CLI Plugin →](/codex-plugin)
- [Pre-Execution Governance →](/pre-execution-governance)
- [Workflow Intelligence as a Service →](/workflow-intelligence-as-a-service)
- [SKILL.md Validator →](/validator/skill)