@assertthat/at-tm-mcp-server
v0.1.0
Published
MCP server for the AssertThat AI Test Manager (AT AI TM) Forge app, authenticated with the caller's own Atlassian account.
Readme
at-tm-mcp-server
MCP server for the AssertThat AI Test Manager (AT AI TM) Forge app. Instead of pasting an app-generated token, each user signs in with their own Atlassian account (OAuth 2.0 / 3LO). From then on:
- Every tool call carries that user's own Atlassian access token, so Jira's own permission model decides what they can touch — a user with access to projects A/B/C who tries to act on project E is rejected by Jira itself, not by anything this server invents.
- Access + refresh tokens are cached locally, so the user doesn't have to log in again next session (refresh tokens are used to silently renew the access token).
- This is a global CLI + a stdio MCP server: install once, run
at-tm-mcp login, wire it into your AI agent.
Cloud-only. Jira Data Center/Server has no equivalent OAuth 3LO / accessible-resources /
mypermissions APIs, so this design doesn't extend to it.
How it fits together
Claude / MCP host --stdio--> at-tm-mcp-server --HTTPS--> AT AI TM Forge app webtrigger
| |
| X-User-Access-Token | assertProjectAccessWithToken()
| X-Cloud-Id | -> /rest/api/3/mypermissions
| Authorization: Bearer <shared secret>The webtrigger side of this already exists in the at-ai-tm repo
(src/resolvers/api-tools.js → webTriggerToolsHandler, src/resolvers/auth-remote.js) — it
was already built to accept a caller-supplied Atlassian bearer token + cloud ID and check Jira
permissions live. This project is the missing piece: it gets that bearer token by having the
user log in with Atlassian directly, and keeps it fresh.
Quick start
1. Install
npm install -g @assertthat/at-tm-mcp-serverNot published to npm yet? From a clone of this repo:
npm install && npm run build && npm link— same result,at-tm-mcpandat-tm-mcp-serverland on your PATH either way.
2. Log in
at-tm-mcp loginFirst run walks you through everything interactively, then opens your browser to sign in with Atlassian:
- Atlassian OAuth 2.0 (3LO) app client ID + secret — one-time per organization. If a
teammate already registered one, ask them for these two values. Otherwise, register one
yourself at https://developer.atlassian.com/console/myapps/:
- Authorization type: OAuth 2.0 (3LO).
- Permissions/scopes:
offline_access,read:me,read:jira-user,read:jira-work. - Callback URL:
http://localhost:8765/callback(only needs to change if you customizeredirectPortviaat-tm-mcp config set redirectPort <port>— keep both in sync).
- AT AI TM webtrigger URL + shared secret — self-service, no Forge CLI needed. As a Jira
admin: Jira Settings → Apps → AT AI TM → AI TM Global Config → MCP connection
details → copy the webtrigger URL, then Reveal/Copy the shared secret. This pair is
specific to your Jira site; every teammate connecting to it uses the same values (not tied to
an individual user — that's what the Atlassian sign-in step handles). If it ever leaks, click
Regenerate secret on that page and have everyone re-run
at-tm-mcp login --reconfigure.
Answers are saved to ~/.at-tm-mcp/config.json, then your browser opens for the actual
Atlassian sign-in, and refresh/access tokens land in ~/.at-tm-mcp/credentials.json. Re-running
at-tm-mcp login later skips straight to the Atlassian sign-in step (reusing saved config) —
use at-tm-mcp login --reconfigure to re-enter the four values instead (e.g. after rotating the
shared secret, or to switch organizations).
Prefer environment variables instead (e.g. for CI, or to avoid writing secrets to disk)? Set
AT_TM_MCP_CLIENT_ID,AT_TM_MCP_CLIENT_SECRET,AT_TM_MCP_WEBTRIGGER_URL,AT_TM_MCP_SHARED_SECRET— env vars always take priority over the config file, and skip the setup prompt entirely once all four are set.
Access tokens refresh automatically and transparently on every tool call once they're close to
expiry — you won't need to log in again unless Atlassian revokes the refresh token (long
inactivity, or you revoking this app's access from your Atlassian account settings). Run
at-tm-mcp login again any time to switch accounts or sites.
3. Add it to your AI agent
login prints these once it finishes, but for reference:
Claude Code:
claude mcp add at-tm -- at-tm-mcp-serverClaude Desktop — add to claude_desktop_config.json
(%APPDATA%\Claude\claude_desktop_config.json on Windows):
{
"mcpServers": {
"at-tm": {
"command": "at-tm-mcp-server"
}
}
}Other MCP hosts: point them at the command at-tm-mcp-server with no arguments — no path or
env block needed, since config/credentials already live on disk from step 2.
Restart Claude Desktop (or reload the MCP connection in Claude Code), then confirm the at-tm
server shows as connected and its tools (jira_issue, test_cases, etc.) are available.
The server process itself only serves tool calls — it doesn't run the interactive login flow (a stdio MCP process can't prompt for browser interaction the way a CLI invocation can). Always run
at-tm-mcp loginfrom a terminal first.
Everyday commands
at-tm-mcp whoami who's currently signed in, and to which site
at-tm-mcp status config + credential file locations, whether everything's configured
at-tm-mcp login --reconfigure re-enter client ID/secret or webtrigger URL/secret
at-tm-mcp logout sign out AND wipe all local config (asks to confirm; -y to skip)
at-tm-mcp config get show current config (secrets masked)
at-tm-mcp config set <key> <value> set one config value manuallyTools exposed
jira_issue, jira_subtasks, folders, test_cases, test_cases_bulk, test_plans,
executions, analytics, tags, issue_links — same surface as the in-Jira Rovo agent (see
manifest.yml's rovo:agent block in at-ai-tm), reachable here over the webtrigger instead.
Development
Running from a local clone instead of the global install:
npm install
npm run build # compiles src/ -> dist/
npm run dev # tsc --watch, recompiles on save
node dist/cli.js login # or: npm run loginnpm link (after building) puts at-tm-mcp/at-tm-mcp-server on your PATH pointing at this
clone's dist/, so the Quick start commands above work identically against your working copy.
After changing src/, rebuild (npm run build) and restart the server in your MCP host to pick
up the change — Claude doesn't hot-reload a running stdio server.
