Azure Devops MCP Onprem

Azure DevOps MCP Server

Governed access to on-prem Azure DevOps (including TFVC) for AI agents and assistants, with typed tools, chainable workflows, and a server-side write-safety layer.

Query work items, repositories, and pipelines in natural language — running locally, no cloud proxy, no telemetry.

Coverage · TFVC · Profiles · Write Safety · Setup

At a glance

Example questions

"Show me all active bugs assigned to me in this sprint" "What changed in changeset 12345?" "Review PR 123 and post the findings as comments" "List my latest shelvesets" "Trigger the nightly build on the release branch"

Coverage

One server across the whole development cycle — Plan, Code, Review, Build & Release, Test, Document — with your own templates exposed as MCP resources alongside the built-in tools and prompts. Every write passes a server-side governance layer — readonly mode, dry-run preview, rate limit, and audit log (see Write Safety); tools can also be scoped per role (see Restrict tools per role below).

Full per-tool parameter reference: Tool Reference ↓

Restrict tools per role

Set AZURE_DEVOPS_ENABLED_DOMAINS to a comma-separated list — disabled domains aren't registered, trimming the AI client's tool list and reducing tool-selection confusion. Default loads all 6. get_current_user is core and always registered.

Unknown domain names fail at startup — no silent typos. Startup log reports what loaded:

Enabled domains (3/6): work_items, tfvc, pipelines
Disabled domains: git, wiki, test_plans

TFVC support

The reason this server exists. Cloud Azure DevOps disabled new TFVC repos in February 2017, and Microsoft's official MCP server doesn't cover TFVC. If your team is still on Team Foundation Version Control, this is the only MCP server that exposes it natively to AI assistants.

10 dedicated TFVC tools:

• Shelvesets — tfvc_list_shelvesets, tfvc_get_shelveset (file changes + work item links), tfvc_get_shelveset_file (shelved, not-yet-checked-in file content — AI review before check-in, a workflow TFVC never had)
• Changesets — tfvc_list_changesets, tfvc_get_changeset (incl. linked work items), tfvc_get_changeset_changes
• Browse, files & diffs — tfvc_browse, tfvc_get_file (at any changeset version), tfvc_get_file_diff (changed hunks between two changesets)
• Work-item linkage — get_work_item_changesets (all TFVC changesets touching a work item, with file contents)

Filters accept @me where relevant. Requires Code (read & write) PAT scope.

Prompts & Resources

Prompts are reusable, advisory workflows surfaced as slash commands in the AI client. Each one instructs the model to gather evidence with the read tools and ground every claim in concrete IDs — producing the report never calls a write tool. One exception by design: review_pull_request can afterwards publish its findings as file-anchored PR comments, but only when you explicitly ask, with every comment confirmed before posting. Prompts load on the same domain axis as tools, so disabling a domain hides its prompts.

lessons_learned is split per backend (Git vs. TFVC) so each variant names its own read tools; both need work_items enabled to read the bug itself.

The table lists 8; a 9th prompt, risk_impact_analysis, is conditional — it appears only when a risk-impact.md template is present (see External resources below).

External resources

Point AZURE_DEVOPS_RESOURCE_DIR at a folder and every *.md file in it is exposed as an MCP resource at template:<filename> (e.g. release-checklist.md → template:release-checklist; the filename is URI-encoded, so spaces are safe). Use this to share team templates and checklists with the AI without baking them into the server.

One template is wired to a prompt: dropping a risk-impact.md file in that folder enables the conditional risk_impact_analysis prompt, which fills the template from work-item evidence — optionally weighing an actual pending change via its shelvesetName argument. No template file → the prompt simply doesn't appear.

Profile-based secrets

mcp.json configs sync to the cloud (Claude Desktop, VS Code Settings Sync), get pasted into tickets, end up in dotfile repos. Inlining AZURE_DEVOPS_PAT there is one git add . away from a public leak.

The convention: set AZURE_DEVOPS_PROFILE=name in mcp.json, keep credentials in a gitignored .env.<name> next to the binary. The server resolves the profile name to that file path; mcp.json stays free of secrets and is safe to commit.

.env.product-a (gitignored):

AZURE_DEVOPS_ORG_URL=https://tfs-1.example.com/tfs/ProductACollection
AZURE_DEVOPS_PROJECT=Product A
AZURE_DEVOPS_PAT=<pat-for-product-a>
# Optional per-profile domain restriction
AZURE_DEVOPS_ENABLED_DOMAINS=work_items,tfvc,pipelines

mcp.json (commitable):

{
  "mcpServers": {
    "ado-product-a": {
      "command": "node",
      "args": ["/path/to/dist/index.js"],
      "env": { "AZURE_DEVOPS_PROFILE": "product-a" }
    }
  }
}

Multi-instance

Once profiles are in place, running multiple ADO instances side-by-side is just adding entries. Each one loads its own .env.<profile> — own PAT, own project, own domain restriction. Per-process state means audit logs, rate limit counters, and @me identity caches never cross between tenants.

{
  "mcpServers": {
    "ado-product-a": {
      "command": "node",
      "args": ["/path/to/dist/index.js"],
      "env": { "AZURE_DEVOPS_PROFILE": "product-a" }
    },
    "ado-product-b": {
      "command": "node",
      "args": ["/path/to/dist/index.js"],
      "env": { "AZURE_DEVOPS_PROFILE": "product-b" }
    }
  }
}

Tool names auto-prefix per server — mcp__ado-product-a__list_repositories vs mcp__ado-product-b__list_repositories.

Env file precedence

Variables set directly in mcp.json's env block always win over file contents.

Each instance's startup log line env file: ... confirms which file was loaded — handy for debugging "which profile did this tool actually call?".

Write Safety

Six layers. The LLM cannot bypass the server-side ones — they short-circuit before any API call fires.

Audit privacy: add AZURE_DEVOPS_AUDIT_REDACT=1 to keep numeric IDs and field shape but drop all string values (titles, comments, branch names). Useful when work-item content carries classified data.

Plus baseline hardening: WIQL injection sanitization, scrubbed errors (no internal paths/URLs/stack traces in client output), bounded pagination (1-1000).

Privacy & data flow

The server runs entirely locally. ADO API calls go straight from your machine to your Azure DevOps Server. No telemetry, no phone-home, no cloud proxy, no shared analytics.

External destinations are limited to:

1. Your Azure DevOps Server — the URL in your .env.
2. Your AI assistant (Claude, GitHub Copilot, Cursor) — the AI client reads tool outputs as conversation context per its own privacy policy. The MCP server itself never talks to these services.

Every network call is visible in src/ — they all route through azure-devops-node-api pointed at your configured URL.

Cloud (Azure DevOps Services)

Technically works against dev.azure.com, but this server isn't positioned for cloud:

• TFVC doesn't exist on cloud — disabled for new orgs since February 2017.
• PAT-only auth — many cloud tenants require Microsoft Entra ID, which this server doesn't yet support.
• Microsoft ships @azure-devops/mcp for cloud — officially maintained, Entra ID, broader cloud-specific coverage.

Use this server against cloud only if you specifically need @me, profile-based multi-tenant config, or a tool the official server lacks.

Setup

Pick one path:

• Quick Start — run from npm, no clone. ~2 minutes.
• Enterprise Setup — clone, build, pin a commit. For air-gapped or audited environments.

Prerequisites

• Node.js ≥ 18
• Azure DevOps Server 2022.2 (tested; older versions with REST API 7.x likely work but untested)
• PAT with the scopes below. Grant only what you need — omitted scopes make the affected tools fail at call time, but the server still starts.

Create the PAT at https://<your-tfs>/_usersSettings/tokens. Set an expiration ≤ 90 days and rotate regularly.

Quick Start (npm)

No public npm access? Skip to Enterprise Setup — it builds from source and can use an internal npm mirror.

1. Credential file — create ~/.azure-devops-mcp.env (Linux/macOS) or C:\Users\you\.azure-devops-mcp.env (Windows):

AZURE_DEVOPS_ORG_URL=https://your-tfs-server/tfs/YourCollection
AZURE_DEVOPS_PROJECT=YourProjectName
AZURE_DEVOPS_PAT=your_pat_token
# AZURE_DEVOPS_SSL_IGNORE=true   # uncomment for self-signed certs

2. Register with your AI client. Shortest paths:

VS Code — one-click install

Claude Code — one command

claude mcp add azure-devops --env AZURE_DEVOPS_ENV_FILE=$HOME/.azure-devops-mcp.env -- npx -y @burcusg/azure-devops-mcp-onprem

Other clients (Claude Desktop / Cursor / Antigravity / Codex CLI)

JSON config — see Configure AI client below.

Enterprise Setup (clone)

git clone https://github.com/burcusipahioglu/azure-devops-mcp-onprem.git
cd azure-devops-mcp-onprem
npm install
npm run build
cp .env.example .env       # copy .env.example .env on Windows
# fill in .env with your TFS details
npm start                  # smoke-test the connection — Ctrl+C to stop

Expected stderr on startup:

Azure DevOps MCP Server "CompanyOrg" running on stdio
env file: /path/to/.env
Enabled domains (6/6): work_items, git, tfvc, pipelines, wiki, test_plans
External resources loaded: 0
Authenticated as: Your Name ([email protected])

Then point your AI client at dist/index.js (see below). No env block needed — the server reads .env from the repo root.

Configure AI client

⚠ Never inline AZURE_DEVOPS_PAT / AZURE_DEVOPS_ORG_URL / AZURE_DEVOPS_PROJECT in client configs. Client configs sync to the cloud (Claude Desktop sync, VS Code Settings Sync) or get pasted into tickets. Use AZURE_DEVOPS_ENV_FILE (Quick Start) or .env in the repo (Enterprise Setup). For multiple TFS instances see Profile-based secrets.

Enterprise Setup config template (any client):

{
  "mcpServers": {
    "azure-devops": {
      "command": "node",
      "args": ["/absolute/path/to/azure-devops-mcp-onprem/dist/index.js"]
    }
  }
}

Quick Start config template (npm + credential file):

{
  "mcpServers": {
    "azure-devops": {
      "command": "npx",
      "args": ["-y", "@burcusg/azure-devops-mcp-onprem"],
      "env": {
        "AZURE_DEVOPS_ENV_FILE": "C:\\Users\\you\\.azure-devops-mcp.env"
      }
    }
  }
}

Restart the client. All 48 tools appear in the tool picker. Server name is auto-detected from AZURE_DEVOPS_ORG_URL (e.g. https://dev.azure.com/acme → acme); override with AZURE_DEVOPS_SERVER_NAME.

Tool Reference

Typed results (first wave): query_work_items, list_pull_requests, get_pull_request, get_build, list_builds, tfvc_get_changeset, tfvc_list_changesets, list_test_runs declare an MCP outputSchema and return structuredContent alongside the usual JSON text (text shape unchanged — existing clients see no difference). An agent can feed one tool's structuredContent straight into the next without parsing prose.

Dry-run: every write tool also accepts dryRun: true (not repeated in the tables below).

Work Items (9 tools)

Git (9 tools)

Git Advanced (4 tools)

TFVC (10 tools)

Pipelines (5 tools)

Core (1 tool, always registered)

Test Management (7 tools)

Wiki (3 tools)

@me token

Several filter parameters accept the magic token @me — the server resolves it to the authenticated PAT owner's identity via Azure DevOps ConnectionData (display name for author/owner filters, user id where the API requires one, e.g. reviewer). The identity is cached for the lifetime of the process.

Why: in multi-agent setups, sub-agents rarely know the human's display name. @me lets any agent filter to "my stuff" without needing identity context.

// "Show me my latest shelveset"
tfvc_list_shelvesets({ owner: "@me", top: 1 })

// "Which commits did I push to feature/login this week?"
list_commits({ repositoryId: "my-repo", branch: "feature/login", author: "@me", fromDate: "2026-04-10" })

WIQL queries (query_work_items) use Azure DevOps' native macros — @Me, @CurrentIteration, @Today — no server-side resolution needed. "My sprint items" is one query away: ... WHERE [System.IterationPath] = @CurrentIteration AND [System.AssignedTo] = @Me.

Need the identity explicitly? Call get_current_user.

Development

npm run dev      # watch mode
npm run build    # build once
npm start        # start the server

License

Released under the MIT License. Copyright (c) 2026 Burcu Sipahioglu Gokbulut.

Free to use, modify, and redistribute — commercial or personal — provided the copyright notice and license text are preserved.