production-master
OfficialThis server does not currently have a deployable release on Glama. Follow the repository's setup instructions to install and configure it for your MCP client, including any required credentials and local dependencies.
Trigger, stream, and review autonomous production-incident investigations without leaving your editor.
Production Master is a thin client for the Production Master hosted service. You point it at an incident, it starts an investigation on the service, and the results stream back into your IDE in real time. When the investigation proposes an action that changes something, you approve or reject it — nothing runs without your sign-off.
The investigation itself runs entirely on the hosted service. This repository is the thin client: it handles sign-in (a device-code login in Cursor, Codex, OpenCode and pmctl; an access token you configure in Claude Code), starts a run, streams live progress, renders the report, and relays your approve/reject decisions. No investigation logic, model provider SDKs, or credentials for the analysis live here.
Features
IDE-native investigations — start and follow a run from Claude Code, Cursor, Codex, or OpenCode. No context switch to a separate dashboard.
Live streaming — progress, findings, and the final report stream over Server-Sent Events (SSE) as the hosted service works.
Human-gated actions — every proposed action that mutates a system is surfaced for explicit approval; you approve or reject before anything happens.
Multi-IDE support — one thin client, registered through each editor's native extension mechanism (plugin, MCP config, or config file).
Related MCP server: agent-debugger
Data flow
From the plugin: the Claude Code plugin sends data to exactly one place, the Production Master hosted service (
https://api.productionmaster.dev), over HTTPS and authenticated by the access token you configure when you install it. Everything you enter in/production-master:investigate,/production-master:connectand/production-master:update, and the run's progress and report, goes through that service. Claude Code keeps the token in your OS credential store. The Cursor, Codex and OpenCode adapters send to the same service (PM_SERVICE_URL, defaulthttps://api.productionmaster.dev), authenticated by the device-code login, and store only your login token locally.On the hosted service: an investigation (a) reads evidence from the connectors your workspace admin configured on the service side, for example Jira, GitHub, Slack, Grafana, PagerDuty and Prometheus; (b) sends that evidence and prompts to Google Vertex AI (Gemini), in Production Master's own Google Cloud project, to reason over them; and (c) stores the run, its evidence and report, and an audit trail in Production Master's database (Supabase Postgres) and evidence store (Google Cloud Storage). Proposed actions that would change a system run only after a person approves them in the plugin.
Retention: evidence is kept for 365 days by default, configurable per workspace; evidence classified as PII is capped at 30 days. Run reports and audit records are kept for the life of the account.
Self-hosted (Enterprise edition): deployments can point the service at other model providers. The hosted service uses Google Vertex AI only.
See the privacy policy for details.
Prerequisites
Node.js 22 (pinned in
.nvmrc) — for Cursor, Codex, OpenCode andpmctl; the Claude Code plugin needs no Node.jsAn account on the Production Master hosted service — Claude Code authenticates with an access token created in the console (Settings, then Access tokens); the other editors use device-code login.
One of the supported editors: Claude Code, Cursor, Codex, or OpenCode.
Quick Start
Claude Code uses an access token (see below). For Cursor, Codex and OpenCode the flow is: register the client → log in with a device code → start an investigation. Point the client at your service with PM_SERVICE_URL (default https://api.productionmaster.dev).
For Cursor, Codex, OpenCode and pmctl, build the client first (workspaces compile the host-neutral core and each adapter); Claude Code needs no build:
nvm use && npm ci && npm run buildClaude Code
Claude Code is wired end-to-end. This repo is its own plugin marketplace (.claude-plugin/marketplace.json), so a fresh clone or a bare owner/repo reference is all /plugin marketplace add needs. The plugin (the self-contained plugin/ folder — a manifest and prompt-only commands/) runs nothing on your machine: it connects Claude Code to the hosted service as a remote MCP server, so it needs no Node.js and no build step.
/plugin marketplace add ProductionMasterAI/production-master
/plugin install production-master@production-masterClaude Code asks for a Production Master access token when you install or enable the plugin. Create one in the console under Settings, then Access tokens; Claude Code keeps it in your OS credential store. Then use the slash commands:
/production-master:investigate PROJ-1234The full set is /production-master:setup (set up, change or remove your access token), /production-master:investigate <ticket>, /production-master:connect <id> and /production-master:update <id> <what to do>. The commands pre-approve no tools, so Claude asks before each call.
On Claude Code 2.1.224+, the plugin can also be installed from an archive source — a zip fetched over HTTPS, with optional SHA-256 pinning — for machines without git or npm; see the note in the quick start.
Cursor · Codex · OpenCode
Each of these editors registers the client as an MCP server that it spawns from the built binary. After npm run build, the config file in this repo points the editor at the client's mcp entry point:
Editor | Registration file | Backed by |
Cursor | ||
Codex | ||
OpenCode |
Log in once with node packages/adapter-<editor>/dist/cli.js login, then start investigations from the editor's own agent — it calls the client's investigation tools over MCP. Point the client at your service with PM_SERVICE_URL (default https://api.productionmaster.dev); secrets are ${ENV} references only, never literals.
pmctl (operator CLI)
For scripting and CI, packages/pmctl is a standalone pmctl binary — the
same thin client over the same BFF, without an editor. After npm run build:
node packages/pmctl/dist/cli.js login
node packages/pmctl/dist/cli.js start PROJ-1234 --title "..." --mode standard
node packages/pmctl/dist/cli.js status <run-id>
node packages/pmctl/dist/cli.js events <run-id> --follow
node packages/pmctl/dist/cli.js report <run-id> --format mdRun node packages/pmctl/dist/cli.js --help for the full command and flag reference. --output json
emits a versioned envelope ({schema:"pmctl/v1", ok, data?, error?}) for scripting; auth, transport,
and streaming are all reused from @production-master/plugin-core — pmctl implements no HTTP, auth,
or SSE of its own, and imports no LLM/provider SDK.
Full walkthrough: docs/user/quick-start.md.
Python
For scripting and CI — a pure-stdlib SDK that speaks the same BFF as the editors, with no third-party runtime dependencies:
pip install "git+https://github.com/ProductionMasterAI/production-master.git@main#subdirectory=sdk/python"from production_master import Client
client = Client() # honours PM_SERVICE_URL
inv = client.start_investigation({"ticket": "ACME-123"})
for event in inv.stream_events(): # SSE with Last-Event-ID resume
print(event.sequence, event.type)
print(inv.get_report(format="json"))Full reference: sdk/python/README.md.
Architecture
The client is a thin transport-and-render layer. All investigation logic lives on the hosted service; the client talks to it over HTTPS (control) and SSE (streaming).
flowchart LR
subgraph IDE["Your IDE"]
C["production-master<br/>thin client"]
end
S["Production Master<br/>hosted service"]
C -- "sign-in (HTTPS)" --> S
C -- "start / approve / reject (HTTPS)" --> S
S -- "live progress + report (SSE)" --> CThe client owns four concerns: auth (device-code login + token storage in the Cursor, Codex and OpenCode adapters and pmctl; in Claude Code, an access token you configure), MCP transport (exposing thin-client commands to the editor), streaming (consuming SSE and rendering progress), and render adapters (per-IDE presentation). It owns none of the analysis.
Documentation
Doc | Purpose |
Install, log in, run your first investigation | |
Common workflows — start, connect, approve/reject | |
Thin-client command reference | |
Auth, service URL, and MCP registration issues | |
Editors and versions the client is validated against | |
Scripting/CI client — install, auth, streaming, publishing | |
How to contribute | |
Release history |
License
MIT — see LICENSE.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
This server cannot be deployed
Maintenance
Related MCP Connectors
Approved test intent, reviewed Playwright automation and run evidence, inside your editor.
Security reviews, threat models over a repo or website, and remediation tracking, in your editor.
AI SRE: investigate production incidents across your observability stack.
Deep security scans of repos you own from your editor: dependency CVEs, SAST, git-history secrets.
Related MCP Servers
- AlicenseAqualityDmaintenanceEnables LLMs to automatically diagnose coding errors through codebase search, test execution, and live debugger integration (DAP/V8 CDP). Provides a secure, policy-gated environment for investigating failures while preventing destructive operations.9ISC
- AlicenseNot gradedqualityDmaintenanceEnables AI agents to investigate backend incidents by executing runbooks that gather evidence from observability and storage systems.9MIT
- FlicenseNot gradedqualityCmaintenanceEnables users to investigate infrastructure incidents in plain English, correlate observability and deploy data with runbooks, and get evidence-backed root-cause proposals with approval-gated remediation.5-
- AlicenseAqualityBmaintenanceConnects AI agents to live observability stacks including Sentry, GitHub, Vercel, Better Stack, and Cloudflare, enabling end-to-end incident investigation, deployment correlation, root cause analysis, and regression triage.14MIT