Skip to content
Guide intermediate

Jama Connect Ships an MCP Server: Spec-Driven Development Comes to Regulated Engineering

Published May 5, 2026 · by Pondero Editorial

The short version

Jama Software released an MCP server in Jama Connect 9.35, letting engineers query specs from Claude, Cursor, Copilot, Codex, and Visual Studio while keeping permissions and audit trails intact.

Table of Contents

Jama Connect Ships an MCP Server: Spec-Driven Development Comes to Regulated Engineering

Most readers will never buy Jama Connect. The reason this release matters anyway is the architectural pattern it proves out: put an MCP server in front of the system that owns the data, and let that system's existing permission model do the access control, instead of scraping or syncing the data into the agent's reach. Jama is the first engineering-management vendor to ship this (Jama Connect 9.35, announced 2026-05-04), and it picked the hardest possible proving ground, regulated engineering, where getting access control wrong is a failed audit, not a bug. If the pattern holds there, it holds for your CRM and your ticket tracker. That is the takeaway to carry off this page even if you never touch Jama.

What Changed

Version 9.35 added a built-in MCP server. It exposes the platform's specifications, items, and full traceability graph to any MCP-compatible client: Claude, Cursor, GitHub Copilot, Codex, Visual Studio. Anthropic, OpenAI, Microsoft, and Cursor have all standardized on MCP as the wire protocol, so this is one endpoint reaching every major coding assistant rather than one integration per tool.

Jama CTO Jim Davidson's framing is that this is "the only product enabling Spec Driven Development via MCP for multidisciplinary engineering teams." Strip the marketing and the technical claim underneath is the interesting part. Most MCP servers shipping in May 2026 are point adapters: one database, one ticket system, one vector store. Jama is exposing an entire requirements-and-traceability graph, including the cross-discipline relationships and version lineage that regulated programs use to prove conformance. A point adapter answers "what is in this table." A graph server answers "which downstream tests does changing this requirement invalidate," which is the question that actually has value and the one a flat adapter cannot reach.

Per the announcement, the server is engineered to scale to 10 million items and 100 million instances of those items. That ceiling targets aerospace, defense, automotive, medtech, semiconductor, and energy programs where one product line generates millions of linked artifacts.

The load-bearing design choice is what the release does not change: the permission model. Existing roles, workflows, and audit policies pass through to the AI session unchanged. A contractor whose Jama account cannot read an export-controlled spec cannot read it through Claude either, because the MCP server enforces at the same layer the human UI does. That is the entire point of the pattern, and it is why scraping the data into a vector store, the common alternative, fails the same audit: a copy outside the permission boundary is a copy with no permission boundary.

Why It Matters

For the regulated teams Jama actually sells to, this resolves a specific deadlock: "can our engineers use AI assistants without breaking compliance?" had no good answer, because every prior option moved data out from under the controls. The MCP server keeps it under them. AI sessions inherit Jama's roles and the traceability lands in the system auditors already trust. Under DO-178C, ISO 26262, or IEC 62304, that is the gap between a conditional yes and a hard no.

For the much larger group of ops and program leads at non-regulated companies, the value is the template, not the product. The recurring unsolved problem is how an AI assistant should touch a source-of-truth system (CRM, ticket tracker, product catalog) without either copying the data somewhere it should not live or hand-syncing it stale. The server-in-front-of-the-owner pattern answers both: no copy, no sync, the owner enforces. Expect Salesforce, ServiceNow, and Atlassian to ship their own versions inside a year, and expect the early in-house builds to repeat the mistake Jama avoided, which is enforcing permissions in the agent layer instead of the data layer.

There is a third, smaller effect worth one line: MCP gains a high-stakes reference customer. The protocol hit 1.0 stable in April with the MCP Apps extension formalized as SEP-1865, and a regulated-industry vendor committing at this scale is the validation cautious enterprise buyers were waiting for.

How to Use It

Jama Connect 9.35 is available today for existing customers, and the MCP server is part of the release rather than a separate SKU per the launch announcement. Jama does not publish public pricing; Jama Connect is quote-based and sold for regulated engineering programs, so contact Jama for a quote if you need specifics.

To connect Claude Desktop, Cursor, or any other MCP-aware client, you point it at your Jama Connect tenant's MCP endpoint. The exact configuration syntax varies by client, but every major MCP client uses a JSON config file with the same general shape:

{
  "mcpServers": {
    "jama-connect": {
      "url": "https://your-tenant.jamacloud.com/mcp",
      "auth": {
        "type": "bearer",
        "token": "<JAMA_API_TOKEN>"
      }
    }
  }
}

Confirm the exact endpoint path and auth method against Jama's 9.35 release notes before you deploy. The evaluation that actually proves the thing is one query: spin up a sandbox tenant, log in as a restricted role, point Cursor at it, and ask the agent to read a spec that role is not permitted to see. The correct outcome is a refusal that matches the UI. If the agent returns the restricted spec, the permission model is not passing through and nothing else about the integration matters. Test the deny path, not the happy path; the deny path is the whole reason this architecture exists.

If you are an ops team eyeing this for non-engineering work, do not buy Jama. It is priced for regulated engineering programs. Take the pattern instead and apply it the day your CRM or ticket vendor ships an MCP server: prefer the server that enforces at the data layer, and run the same restricted-role deny test before you trust it.

This post is part of Pondero's daily coverage of AI tool updates. See all MCP guides