Skip to main content
Glama

mcp-ado-browser

Azure DevOps for MCP — via your browser, not a PAT.

An MCP (stdio) server that gives read-only access to Azure DevOps using only your existing browser session — no PAT, no Azure CLI, no official ADO MCP, no credential provider. The only source of authentication is the session of a real browser, driven by Playwright on an isolated, dedicated profile.

It is org-wide by default: only the organization is required, and it browses every project, repo and feed you can access.

Requests reuse whatever credential your signed-in browser session already carries. Where Azure DevOps authenticates the web app with a bearer token rather than cookies, the server loads the app shell once, observes the Authorization: Bearer header on the app's own API traffic, and replays that token for every host (dev.azure.com, feeds, pkgs, almsearch). You get JSON back — never DOM scraping for data the REST API provides. No PAT is ever created, and no token is written to disk.

Why these choices (restricted-environment friendly)

Concern

Decision

No Playwright browser download

playwright-core + channel: 'chrome'/'msedge' uses an already-installed browser; nothing is downloaded.

SQLite without a native build

node:sqlite (built into Node ≥ 22.5) — zero compilation.

One package, one binary

The MCP server and the authenticate mechanism ship in the same package and the same npx binary.

No hardcoded values

org/project/ids come from flags/env or discovery; api-versions live only in src/ado/versions.ts.

Related MCP server: Azure DevOps MCP Server

How it works

flowchart TD
    A["authenticate<br/>(visible, chromeless window)"] -->|"sign in once · MFA"| P[("browser session<br/>persisted on an isolated profile")]
    P -. restored .-> W["headless work session"]
    W -->|"loads the app shell once"| T{{"session access token<br/>(observed, in memory only)"}}

    MC["MCP client<br/>(Claude / Cursor / …)"] -->|"tools/call"| SRV["mcp-ado-browser<br/>(stdio MCP server)"]
    SRV --> W
    T --> ADO["dev.azure.com · feeds · pkgs · almsearch<br/>(your real session)"]
    ADO -->|"JSON"| W
    W --> SRV
    SRV <-->|"TTL + Rev freshness"| DB[("SQLite cache")]
  1. Authentication is your browser, not a PAT. authenticate opens a real, visible browser window on a dedicated, isolated profile (never your daily browser). You sign in normally (MFA included). The tool detects success by polling an authenticated endpoint, then snapshots the browser session to disk. No PAT is ever created.

  2. Work runs headless. Subsequent runs launch the same profile headless and restore that snapshot — no window, no re-login until the session expires. The snapshot is needed because the cookie that keeps the Azure DevOps app shell loaded is a session cookie, which Chrome drops when it exits.

  3. Requests reuse the web app's own credential. On the tenants this was tested against, the Azure DevOps web app authenticates its API with Authorization: Bearer (MSAL) and a cookie-only request gets 401 on every host — the sign-in URL even carries protocol=cookieless. (This is observed behaviour; we found no Microsoft announcement of such a change, so treat it as something that varies rather than a rule.) The server therefore loads the app shell once, observes the token on the app's own API calls, and replays it. One token covers dev.azure.com, feeds, pkgs and almsearch alike.

    The token is held in memory only and refreshed automatically on a 401. It is deliberately never written to disk: a bearer token reachable from JS is more exposed than an httpOnly cookie — Microsoft's own MSAL guidance states browser storage is only safe absent XSS, and that its cache encryption "reduce[s] the persistence of auth artifacts, not to provide additional security". This server does not widen that exposure.

  4. 401 and 403 are kept distinct. 401 means the session is dead → AUTH_REQUIRED, re-run authenticate. 403 means you are signed in but lack permission on that resource (a private feed, say) → a 403 error saying so. Re-authenticating cannot fix a permission problem.

  5. Responses are cached in a local SQLite DB (node:sqlite) with a configurable TTL. On a stale hit, a cheap freshness check (System.Rev for work items) avoids re-downloading unchanged data.

  6. When the session dies, tools fail fast with AUTH_REQUIRED — just re-run authenticate and continue.

Getting started

Prerequisites: Node ≥ 22.5 and Google Chrome (or Microsoft Edge) installed. You do not need a PAT, the Azure CLI, or any admin setup.

Setup is two steps:

  1. Register the server in your MCP client (one config entry — see your client below).

  2. Sign in once — just ask your assistant: “authenticate to Azure DevOps”. The built-in authenticate tool opens a visible browser window; you log in (MFA), and the session is persisted. (No separate terminal command needed.) From then on everything runs headless until the session expires — then just ask it to authenticate again.

The command every client runs is the same:

npx -y mcp-ado-browser --org <your-org>

Config is passed as CLI flags (--org, --project, …) or env vars (ADO_ORG, …); flags win. Then ask things like “list my active pull requests”, “show work item 1234 and its linked PR”, or “what feeds and packages are in this org?”.

Tip: prefer per-user/local config (not committed) so your org name doesn't land in a shared repo. Or omit --org from a committed config and set ADO_ORG in your env.

Use it from your MCP client

claude mcp add ado --scope local -- npx -y mcp-ado-browser --org <your-org>

Then ask Claude to “authenticate to Azure DevOps”.

{
  "mcpServers": {
    "ado": {
      "command": "npx",
      "args": ["-y", "mcp-ado-browser", "--org", "<your-org>"]
    }
  }
}
{
  "servers": {
    "ado": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "mcp-ado-browser", "--org", "<your-org>"]
    }
  }
}

Open Copilot Chat in Agent mode and pick the ado tools. (Avoid committing your org — use ${env:ADO_ORG} or a personal config.)

{
  "mcpServers": {
    "ado": {
      "command": "npx",
      "args": ["-y", "mcp-ado-browser", "--org", "<your-org>"]
    }
  }
}
[mcp_servers.ado]
command = "npx"
args = ["-y", "mcp-ado-browser", "--org", "<your-org>"]

After registering, trigger sign-in from the chat (“authenticate to Azure DevOps”), which runs the authenticate tool. Prefer a terminal instead? npx -y mcp-ado-browser authenticate --org <your-org> does the same thing. Tools return a structured AUTH_REQUIRED error when the session expires — re-authenticate and continue.

Tools (tools/list)

Tool

What it does

list_projects

All projects you can access (org-wide).

list_repositories

All Git repos across the org (or one project).

search_work_items

WIQL (org-wide by default) or full-text (almsearch); project to scope.

get_work_item

Work item with $expand=all + relations (hierarchy, Related, PR ArtifactLink resolved).

get_work_item_comments

The separate comments endpoint (project derived automatically).

get_comment_details

A comment plus its downloaded attachments (size, sha256).

search_pull_requests

PRs org-wide, by repo, or by project; filter by status/author/target.

get_pull_request

Metadata, branches, reviewers, linked work items (repo by id or name).

get_pull_request_comments

Threads, distinguishing system vs human.

search_feeds

Artifacts feeds → packages → versions.

download_artifact

.nupkg/.tgz from a feed (cross-host pkgs.dev.azure.com), with archive-integrity validation.

authenticate

Opens a visible browser for interactive sign-in (MFA); persists the session. Run it once, or whenever a tool returns AUTH_REQUIRED.

Commands

The single npx mcp-ado-browser binary has a few subcommands:

Command

What it does

npx mcp-ado-browser --org <org>

Start the MCP stdio server (default).

… authenticate --org <org>

Interactive sign-in (visible browser). Same as the authenticate tool.

… status --org <org>

Show the profile/cache paths, the org, and whether the session is signed in (and as who).

… logout

Clear the persisted session and the cache (a local sign-out). No org needed.

Switching org with the same account needs nothing special — just change --org; one sign-in covers every org that account can access. A different account → logout first, then authenticate against the other org.

Where it stores things

Everything is local to your machine, under a single dedicated folder (mode 700, never committed). Nothing is hosted remotely — the server is a local process spawned by your MCP client over stdio.

What

Path (default)

Browser session (profile)

macOS/Linux: ~/.mcp-ado-browser/profile/ · Windows: C:\Users\<you>\.mcp-ado-browser\profile\

Session snapshot (mode 600)

…/.mcp-ado-browser/session-state.json — the session cookies Chrome will not keep on its own. Treat it like a credential; logout deletes it.

SQLite cache

…/.mcp-ado-browser/cache.sqlite

Package code (npx cache)

macOS/Linux: ~/.npm/_npx/<hash>/…/mcp-ado-browser · Windows: …\AppData\Local\npm-cache\_npx\<hash>\… (see npm config get cache)

Reset everything (forces re-login): logout, or rm -rf ~/.mcp-ado-browser.

Configuration

Flag

Env

Default

Meaning

--org

ADO_ORG

—

Organization (required).

--project

ADO_PROJECT

—

Default project scope (optional; org-wide otherwise).

--user-data-dir

ADO_USER_DATA_DIR

~/.mcp-ado-browser/profile

Isolated persistent browser profile.

—

ADO_SESSION_STATE

~/.mcp-ado-browser/session-state.json

Session snapshot (mode 600) that keeps you signed in across browser restarts.

--channel

ADO_BROWSER_CHANNEL

chrome

chrome or msedge.

--cache-ttl

ADO_CACHE_TTL_SECONDS

900

Global cache TTL. Per-resource: ADO_CACHE_TTL_WORKITEM=60.

--api-version

ADO_API_VERSION

discovery/defaults

Force an api-version for all areas.

--no-app-window

ADO_APP_WINDOW=0

app mode

Use a normal browser window for sign-in.

--headed

ADO_HEADLESS=0

headless

Run work with a visible window.

Development & verification

npm install
npm run build
npm run verify           # all offline gates (browser stack, MCP, tools, cache, artifacts, no-hardcoding)
npm run verify:live      # adds the live acceptance pass against real Azure DevOps
npm run scan:secrets     # pre-push secret / sensitive-data scan
npm run demo:live        # drive the real stdio server as an MCP client (env-driven)

npm run verify prints a detailed report, gate by gate, assertion by assertion. BLOCKED_ON_AUTH is transitory: the run is not done until the live pass is green; the only tolerated terminal exclusion is EMPIRICALLY_BLOCKED (with evidence), for the cross-host artifact download only.

Security & privacy

  • Authentication is only your real browser session on a dedicated, isolated profile — no PAT or token is ever created, stored, or transmitted by this tool.

  • The session lives in ~/.mcp-ado-browser/profile (machine-local, gitignored).

  • Fixtures and reports are anonymized; npm run scan:secrets blocks pushes that would leak personal/org data or secrets (also enforced in CI).

License

MIT © VMargan

Available Tools

12 tools
authenticateA

Sign in to Azure DevOps by opening a VISIBLE browser window for interactive login (MFA included). Run this once (or whenever a tool returns AUTH_REQUIRED). The session is persisted on an isolated profile and reused headless afterward — no PAT or token is ever stored.

Inputs: timeoutSeconds.

download_artifactA

Download a package artifact (.nupkg / .tgz) from a feed via the browser session (pkgs.dev.azure.com). Validates archive integrity (size, sha256, valid zip/tgz) for re-hosting.

Inputs: feedId (required), saveDir (required), version (required), protocol (required), +2 more.

get_comment_detailsA

Resolve a work item (and optionally a specific comment) AND download all related attachments (work-item AttachedFile relations + attachments referenced in the comment body). Returns metadata + downloaded content stats (size, sha256).

Inputs: workItemId (required), saveDir, commentId.

get_pull_requestA

Get a pull request with metadata, branches, reviewers and linked work items. Resolved org-wide — repoId may be a GUID or a repository name.

Inputs: prId (required), repoId (required).

get_pull_request_commentsA

Get PR threads (system vs human). Resolved org-wide — repoId may be a GUID or a repository name.

Inputs: prId (required), repoId (required).

get_work_itemA

Get a single work item with $expand=all, including relations (hierarchy, Related, and ArtifactLink PR references resolved).

Inputs: id (required), bypassCache.

get_work_item_commentsA

Get the full discussion for a work item (GET _apis/wit/workItems/{id}/comments — a SEPARATE endpoint, not part of $expand).

Inputs: id (required).

list_projectsA

List ALL Azure DevOps projects the user can access in the organization (GET _apis/projects). Use this to discover projects to browse.

Inputs: no named parameters.

list_repositoriesA

List ALL Git repositories the user can access across the organization (GET _apis/git/repositories), or within one project. Returns id, name and owning project for each repo.

Inputs: project.

search_feedsA

Browse Azure Artifacts feeds (GET feeds.dev.azure.com/_apis/packaging/feeds). Pass feedId to also list packages + versions. Each feed reports its project (null for org-scoped feeds); project-scoped feeds are addressed under their project automatically.

Inputs: feedId.

search_pull_requestsA

Search pull requests ORG-WIDE by default, or within a repo (repoId) or a project. Filters: status, creatorId, targetRef. (GET _apis/git/pullrequests)

Inputs: top, repoId, status, project, +2 more.

search_work_itemsA

Search/browse work items ORG-WIDE (cross-project) by default. Backend is WIQL (POST _apis/wit/wiql); pass text for full-text Search (almsearch), or project to scope to one project. Returns id + summary fields.

Inputs: top, text, wiql, project.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A4.1/5.0

    Scored across 12 tools

    Disambiguation5/5

    Each tool targets a distinct resource/action: projects, repositories, work items, PRs, feeds, artifacts, and authentication. The only possible overlap is between get_work_item_comments and get_comment_details, but their descriptions clearly separate discussion retrieval from comment resolution plus attachment download.

    Naming Consistency4/5

    Most tools follow a consistent verb_noun pattern (list_projects, search_work_items, get_pull_request). Minor deviations include get_comment_details (which omits 'work_item') and authenticate (verb-only), but these do not significantly hinder readability.

    Tool Count5/5

    With 12 tools, the set is well-scoped for a browser-oriented Azure DevOps server. Each tool covers a meaningful browsing operation, and there is no redundancy or bloat.

    Completeness4/5

    The toolset provides good read-only coverage: listing and searching projects, repos, work items, PRs, and feeds, plus getting details and comments. Minor gaps include no explicit get_project or get_repository (though list tools return sufficient info) and no code search, but core browsing needs are met.

    Maintenance

    ActivityInactive
    ResponsivenessNo recent community issues

    Related MCP Connectors

    Related MCP Servers