Join the Waitlist
Get early-adopter pricing when this tier launches. We'll email you first — no spam, no obligation.
✓ You're on the list. We'll email you first.
PROMPT GOVERNANCE INFRASTRUCTURE · BETA

THE CI/CD PIPELINE
PROMPT MANAGEMENT
TOOL.

// propose → version → approve → serve · ci/cd native

Hardcoded prompts scattered across agent configs, codebases, and .env files are a liability. PromptMatrix is the ultimate CI/CD pipeline prompt management tool — pulling every prompt into a single governed layer. Review, version, and instantly update production prompts via API without triggering a full deployment.

✓ ANY STORAGE — browser · sqlite · self-hosted
◎ No server needed to start — runs entirely in your browser. Upgrade storage when you're ready.
BECOME A FOUNDER → SELF-HOST FREE
⬡ Launch local dashboard · ⚠️ ONLY -- OF 100 SPOTS LEFT · $499 one-time · no subscription ever
VIEW SOURCE ON GITHUB — MIT LICENSED · Open Source Core
Professional · open source · MIT
OSS · no signup needed
Zero setup to start
Every change gated
0DEPLOYS PER CHANGE
3STORAGE TIERS
MITOPEN SOURCE
∞PROMPTS · AGENTS
// LIVE APPROVAL SEQUENCE
◈
ENGINEER · INITIAL SETUP
Registered assistant.persona in production
LIVE One-time setup · never touches again
↓
◉
NON-ENGINEER · PROPOSING CHANGE
"The tone reads too corporate. Users are bouncing at the first response. Needs to be direct, human."
PENDING REVIEW 3 min ago · no code touched
↓
✓
ENGINEER · REVIEWING DIFF
Diff reviewed · semantics checked · approved · deployed in 9s
APPROVED Logged to immutable audit trail
◎ /pm/serve/:key — live prompt endpoint
LIVE
$GET/pm/serve/assistant.persona
→200 OK11ms
CURRENT APPROVED CONTENT — 9 SECONDS AGO
Your application calls this URL at runtime. When an engineer approves a change, the very next call returns the new content. Zero redeploy. Ever.
// OPERATOR_PROFILES

Three roles. One protocol.

// THE ENGINEER
◈
You wire it in. Once.

Swap the hardcoded string with one API call. That's the full integration cost. After that, your job is reviewing diffs and clicking approve — not fielding Slack messages about wording.

→ 30-minute one-time setup
→ Approve or reject any change in 10s
→ Zero deploys for prompt changes — ever
// PM · MARKETER · OPERATIONS
◉
You fix it yourself.

You see the live prompt in plain text. You edit it. You submit it. The engineer reviews the diff and approves. You never file a ticket for a word change again — and you never touch a codebase.

→ No Git. No terminals. No tickets.
→ Plain text edit in any browser
→ Approved changes live in 10 seconds
// SOLO BUILDER · INDIE ENGINEER
⑂
Prompts leave the codebase.

Even building alone, every prompt gets versioned, audited, and editable from a dashboard — not scattered across env files, Notion docs, and forgotten commits. Your system has memory.

→ All prompts in one governed registry
→ Complete history · 1-click rollback
→ Any LLM API · any framework
// STORAGE_ARCHITECTURE

Run it your way.
Your data. Your rules.

No forced cloud. No vendor lock-in. Start in your browser today. Move to SQLite when you're ready. Self-host when it's serious. Cloud when your team needs it.

🌐
// TIER 1 · INSTANT
Browser Storage

Zero setup. Open the HTML file, start adding prompts. Everything lives in localStorage — no backend, no install, no account. Your agents can read from it instantly.

✓ How the builder uses it today
✓ No install required
✓ Instant. Today.
🗃️
// TIER 2 · LOCAL FILE
SQLite

Download the open-source repo, point it at a local SQLite file. Persistent across browser sessions, queryable, portable. Ideal for solo developers and indie builders who want durability without a server.

✓ Data survives browser restarts
✓ Git-friendly · single file
✓ No network dependency
🖥️
// TIER 3 · SELF-HOSTED
Your Own Server

Point PromptMatrix at your own Postgres, MySQL, or any database backend. Full control, zero vendor relationship. Teams and orgs who need data residency, custom auth, or integration into an existing stack.

✓ Any database backend
✓ Full data sovereignty
✓ MIT licensed · self-managed
📡
// ENTERPRISE READY
Managed Scaling

PromptMatrix is built to scale. Start locally, then move to your own production cluster. The protocol remains the same. Whether you have 5 prompts or 5,000, the governance engine handles the load with zero friction.

⚡ Horizontal scaling ready
⚡ Production-hardened core
⚡ 100% Open Governance
⟶
// MIGRATION PATH
Start in your browser this afternoon. When you outgrow it, export your prompt registry and move to SQLite or your own server — your data, your schema, your keys. No migration tax. No lock-in.
// SYSTEM_FAILURE_MODES

Your AI behavior is
running ungoverned.

The people closest to your users — product, marketing, ops — can't touch what the AI says. They file tickets. Engineers handle it between real work. Prompts drift from what's in Notion. Nobody knows what's actually live in production right now. You shipped an AI product without a control plane for it.

// 01
🔒
Only engineers can change what your AI says.
Tone is wrong. Response needs tuning. Brand voice drifted. Every fix demands a PR, a review cycle, and a deploy. The people who actually understand your users have no path in.
ITERATION SPEED: 2–6 DAYS PER CHANGE
// 02
🎫
Non-technical people file tickets for word changes.
Someone notices a problem. They can't fix it. They file a ticket. It enters the backlog. By the time it ships, the context is stale and the person who raised it has moved on to the next fire.
BACKLOG LATENCY: 5–14 DAYS
// 03
👻
Nobody knows which prompt is actually running in production.
The "approved" version lives in a doc. The live version is different. There's a draft in Slack. An engineer made a hotfix. The doc, the code, and production are three different realities.
SINGLE SOURCE OF TRUTH: ✗
// 04
🚨
Prompt changes ship with zero visibility or gate.
A string in a large PR. It merges. It hits production. AI behavior shifts. No one caught it in review. There's no rollback plan because there was no change log. You find out from users.
MEAN TIME TO DETECT: HOURS OR DAYS
// 05
📋
Compliance can't audit what your AI said or when it changed.
Legal and enterprise customers need a change log: who edited this, when, what did it say before. Git blame on a hardcoded string is not an audit trail that passes review. It's a liability that surfaces in deals.
AUDIT TRAIL: NONEXISTENT
// 06
⏳
Engineering hours spent on non-engineering work.
Three to six hours per week. Senior engineers updating tone, adjusting copy, shipping work that anyone with product context could handle — if there were a safe, governed channel to do it through.
ENGINEERING WASTE: 150–300 HRS/YEAR
// STATE_COMPARISON

Uncontrolled changes.
Or a control plane.

✗ WITHOUT GOVERNANCE
✗
Product spots a tone failure. Files a ticket. Waits 5 days. Engineer deploys a two-word change.
✗
Leadership asks "what is our AI actually saying right now?" — no one can answer with certainty.
✗
Marketing wants to test two onboarding variants. Needs an engineer. Gets backlogged.
✗
A prompt change slips through a large PR unreviewed. Quality drops. Support tickets surface it three days later.
✗
Compliance requests the AI's change history. You produce a Git log. The deal stalls.
✗
The people who understand users best have zero influence over how the AI behaves.
✓ WITH PROMPTMATRIX
✓
Product edits the prompt in plain text. Engineer reviews the diff and approves. Live in 10 seconds.
✓
The registry shows every prompt's live content, version history, and who approved the last change.
✓
Marketing proposes both variants. Engineer approves both. They're tracked with quality scores and full history.
✓
Every change is reviewed before it reaches production. Bad version? 1-click rollback in under 10 seconds.
✓
Full immutable log: who changed what, when, previous content, who approved. Export-ready. Enterprise-grade.
✓
Non-engineers contribute directly to AI behavior. Engineers maintain the gate. Everyone moves faster.
// THE_PROTOCOL

Four steps. Engineer runs two of them.
Everything else is governed automatically.

Engineers own steps 1 and 4. Everyone else uses steps 2 and 3. The gate never opens without engineer review.

01
Engineer registers the prompt key

One API call replaces the hardcoded string. Done once per prompt. The key appears in the dashboard immediately. Engineers never touch it again.

◈ ENGINEER ONLY
02
Anyone on the team proposes a change

Product, marketing, operations — they see the live prompt in plain text in the dashboard. They edit it. They submit. No code. No Git. No ticket.

◉ ANY TEAM MEMBER
03
Engineer reviews the diff

A clean line-by-line diff. Old version in red. New version in green. The engineer sees exactly what changed, approves if it's correct, rejects with a note if it's not.

◈ ENGINEER APPROVES
04
Approved content goes live in 10 seconds

The serve endpoint returns the approved content on the next call. No application redeploy. The audit trail is written. The version is snapshotted for rollback.

✓ LIVE · AUDITED · ROLLBACKABLE
// THE_REVIEW_INTERFACE

The engineer sees
exactly what changed.

Not a wall of text. A precise, line-by-line diff. Old content in red. New content in green. The engineer knows the full scope of what they're approving — and can reject with a written note if anything is off-brand, incorrect, or unsafe.

assistant.persona — proposed change · product team · 6 min ago
+1 line -1 line
1 You are an AI assistant for {{product_name}}.
2−Respond in a formal, structured tone. Use complete sentences.
2+Be direct and human. Skip the formality. Get to the point.
3 Always cite sources when making factual claims.
// CHANGE REQUEST proposed 6 min ago · no code touched

"Data shows users are dropping off after the first response. The formal tone is creating friction. This makes it feel like a tool, not a wall. Easy to roll back if it moves the wrong metric."

// PLATFORM_ARCHITECTURE

The full governance
stack for AI behavior.

Not just an editor. A controlled change pipeline — with history, diff review, quality scoring, and an audit trail that compliance teams actually accept. Every layer of the governance stack, shipped.

// 01
⑂
The Approval Workflow

The core of the system. Non-engineers propose changes in plain text. Engineers see the full diff before approving — old content vs new, line by line. Reject with a note, or approve and it's live in 10 seconds. Nothing reaches production without an explicit human decision.

proposed_by: hello@promptmatrix.io
status: PENDING_REVIEW
diff: -1 line / +1 line
→ engineer approves
status: LIVE · deployed in 9s · trail written
◈ ENG — holds the gate ◉ PM — proposes freely $ EXEC — full visibility
GOVERNANCE · DIFF VIEW · 10s APPROVAL
// 02
◎
The Serve Endpoint

Your application calls /pm/serve/key at runtime. It always returns the current approved version. When a change is approved, the next call returns the updated content. No application redeploy, ever.

LIVE API · RUNTIME FETCH · SDK SUPPORT
// 03
◧
Version History + Rollback

Every approved change creates an immutable snapshot. Full history of what was live and when. 1-click rollback to any previous version in under 10 seconds. No incident call. No engineer on-call required.

SNAPSHOTS · HISTORY · 10s ROLLBACK
// 04
◈
The Prompt Registry

Every prompt your system uses, in one place. Live content, current status, last change, who approved it, what version it's on. The single source of truth — visible to engineers, product, and leadership alike.

REGISTRY · STATUS · SINGLE SOURCE
// 05
📋
Compliance-Grade Audit Trail

Immutable log of every event: who proposed, who approved, previous content, new content, exact timestamps. Export-ready for compliance review, enterprise contracts, and legal discovery. Not a Git log — a real audit trail.

AUDIT LOG · COMPLIANCE · CSV EXPORT
// ALTERNATIVES_ANALYSIS

You've tried the
workarounds. Here's why they break.

Every team reaches for the obvious shortcuts first. Each one has a structural failure. Here's what each actually costs — and what PromptMatrix does instead.

APPROACH STRUCTURAL FAILURE NON-ENG EDIT APPROVAL GATE AUDIT LOG
Hardcoded in source Every change is a PR and a deploy. Non-engineers are permanently locked out. No change history outside of git blame. ✗ ✗ ✗
Env variables / config files Still requires a deploy to change. No diff view. No approval gate. Who actually edits env vars on your team? Not the product manager. ✗ ✗ ✗
Notion / Google Doc Disconnected from the application. The doc drifts from production immediately. No enforcement, no history that maps to what actually ran. ~ ✗ ✗
LaunchDarkly / feature flags Designed for boolean feature flags, not prose text governance. No diff view for language. No non-engineer editing workflow. Expensive for what you actually get. ~ ✗ ~
LangSmith / LangFuse Observability and tracing tools — they watch what's happening, not govern what happens next. Still no approval gate. Still no safe editing path for non-engineers. ~ ✗ ~
PromptMatrix Built exactly for this problem: non-engineers edit safely, engineers review and approve, everything is logged immutably, live in 10 seconds. ✓ ✓ ✓

~ = partial or limited · ✗ = not supported · ✓ = built for this

// INTEGRATION_PROTOCOL

Register once.
Serve forever.

The entire engineer integration is two code changes. Register the key once. Replace the hardcoded string with the serve endpoint. That is the complete engineering cost. Everything after that happens in the dashboard — no further deploys, ever.

Python · Before ✗ HARDCODED · NO GOVERNANCE
# prompt buried in application code # no version history · no approval gate · no rollback # works until someone asks "what is our AI actually saying?" def get_prompt(): return "You are a helpful AI assistant. Be concise. Stay on topic. Never fabricate information." # to change this: PR → review → deploy # average turnaround: 2–6 days # non-engineers: locked out permanently
Python · After ✓ GOVERNED · LIVE IN 10s
# one-time setup — register key in codebase once import requests # same pattern for any prompt — chatbot, agent, pipeline def get_prompt(): res = requests.get( "http://localhost:8000/pm/serve/assistant.persona", headers={"x-pm-key": PM_API_KEY} ) return res.json()["content"] # now any team member can propose a change # engineer reviews the diff → approves → live in 10s # zero redeploy · full audit trail written automatically
// SDK · PYTHON + NODE.JS — AVAILABLE NOW
Python SDK pip install promptmatrix
from promptmatrix import PromptMatrix pm = PromptMatrix(api_key="pm_your_key") # register once — replaces the hardcoded string pm.register("assistant.persona") # fetch current approved version at runtime prompt = pm.serve("assistant.persona") # drop directly into your LLM call response = openai.chat.completions.create( model="gpt-4o", messages=[{"role": "system", "content": prompt}] )
Node.js SDK npm install promptmatrix
import { PromptMatrix } from 'promptmatrix' const pm = new PromptMatrix({ apiKey: 'pm_your_key' }) // register once — replaces the hardcoded string await pm.register('assistant.persona') // fetch current approved version at runtime const prompt = await pm.serve('assistant.persona') // drop directly into your LLM call const res = await openai.chat.completions.create({ model: 'gpt-4o', messages: [{ role: 'system', content: prompt }] })
// RAW HTTP · ANY LANGUAGE · ANY FRAMEWORK
curl / fetch / requests — no SDK required, works everywhere ✓ FRAMEWORK AGNOSTIC
# one GET to fetch the current approved prompt curl -H "x-pm-key: pm_your_key" http://localhost:8000/pm/serve/assistant.persona { "key": "assistant.persona", "content": "I'm your AI assistant. Ask me anything — I'll give you a direct answer.", "version": 14, "approved_by": "team@acme.internal", "approved_at": "2026-03-12T09:41:02Z", "latency_ms": 11 }
// ENGINEER SETUP · TOTAL TIME: ~30 MINUTES
01
Install and initialise
pip or npm install. Run pm init with your API key. Two minutes total.
$ pip install promptmatrix $ pm init --key pm_your_key ✓ PromptMatrix initialised
ENGINEER ONLY · ONCE
02
Register your prompt keys
Find every hardcoded prompt string. Call pm.register("key"). Each key appears in the dashboard immediately and is ready for governed editing.
$ pm register assistant.persona $ pm register agent.briefing $ pm register email.composer ✓ 3 prompts registered · live in dashboard
ENGINEER ONLY · ONCE PER PROMPT
03
Replace and deploy once
Swap each hardcoded string with pm.serve("key"). Deploy this change once. You never deploy a prompt change again. The governance layer handles everything from here.
# before prompt = "You are a helpful AI assistant..." # after — your final deploy for this prompt prompt = pm.serve("assistant.persona")
THE LAST DEPLOY FOR THIS PROMPT

After step 3 — any team member can propose a change in the dashboard. The engineer reviews the diff and approves in 10 seconds. The new content is live. Zero further deploys. Ever.

// LATENCY — SOLVED WITH CACHE
A 30-second cache. Overhead rounds to zero.

Without caching, a serve call adds ~10ms per LLM request. With a local 30-second TTL, the overhead is negligible — and every approved change reaches your app within 30 seconds of going live. Cache once, serve instantly.

# Python — 30s TTL cache
import time
 
_cache = {}
def get_prompt(key):
  now = time.time()
  if key in _cache and now - _cache[key][1] < 30:
    return _cache[key][0]
  val = pm.serve(key)
  _cache[key] = (val, now)
  return val
// PRODUCTION SAFETY
Always keep a hardcoded fallback.

If the serve endpoint is unreachable — network blip, maintenance window — your application should not fail. Wrap the call in a try/except with a sensible default. The fallback fires only if PromptMatrix is unreachable. Your app stays live regardless.

# The correct production pattern
FALLBACK = "You are a helpful AI assistant."
 
def get_prompt(key):
  try:
    return pm.serve(key)
  except Exception:
    return FALLBACK
// GOVERNANCE_SCOPE

Any AI product.
Every prompt. Now.

If your product sends a string to a language model, it belongs in PromptMatrix. Register any key in 30 seconds. These are the prompt categories teams are putting under governance today.

◉
Chatbots & Assistants
System prompt, persona, fallback responses, guardrails. Tune how your chatbot behaves in production — approved changes live without touching code.
assistant.persona assistant.guardrails assistant.fallback
◈
AI Agents & Agent Networks
Control the briefing, operating rules, and persona of every agent. Edit any agent's instructions from the dashboard — no code change, no restart, no hunting through config files. Works with OpenClaw, LangChain, AutoGen, or any orchestrator.
agent.briefing agent.rules agent.persona
✉
Email & Outreach Writers
Version-control the prompts behind your email sequences, cold outreach, and follow-ups. Propose, review, and approve before any change reaches a live send.
email.composer email.subject outreach.cold
◧
Classifiers & Routing Logic
Update classification rules and intent routing without a code change. Full diff review before any rule change touches live traffic.
classifier.rules router.intent triage.priority
✍
Content & Writing Pipelines
Manage tone, format, and style rules for every content workflow. Marketing proposes, engineer approves — no more "can you just fix the prompt" Slack messages.
content.tone content.format newsletter.voice
⑂
Anything with a Hardcoded Prompt
If a string in your code gets sent to a language model, it belongs under governance. Works with OpenAI, Anthropic, Gemini, Mistral — any API, any model, any team.
any.key any.model any.team
// ORIGIN · WHY THIS EXISTS

Built because every team
shipping AI hits this wall.

// PATTERN · CONTENT TEAMS

A marketer sees a tone problem in the AI's response. It's a 3-word fix. She files a ticket. It sits for 9 days. By the time the engineer ships it, the campaign is over and the moment has passed. The problem isn't engineers — it's the absence of a governance layer between them.

✍
Content & Marketing Teams
Propose in plain text. No Git. No tickets. No waiting.
// PATTERN · ENGINEERING TEAMS

An engineer gets a Slack message: "can you change the wording on the support bot?" Twenty minutes — PR, review, deploy. Multiply by three times a week, every week. That's 50+ engineering hours per year spent on word changes.

◈
Engineering Teams
Review a diff. Click approve. Back to real work.
// PATTERN · AI AGENT SYSTEMS

You have 26 agents — each with a different briefing, persona, and operating rules. They're running on OpenClaw, LangChain, or raw API calls. Changing one agent's instructions means finding the hardcoded string, editing it, restarting the process. PromptMatrix puts every agent's instructions in one governed registry. Edit any of them from a dashboard. No code. No restart. No hunting through files.

⚡
Agent System Builders
OpenClaw · LangChain · AutoGen · raw API · any orchestrator.
// LIVE DEMO

The dashboard.
Running. Right here.

This is the real PromptMatrix interface — running in demo mode with simulated data. Click through every view. Edit prompts in the studio. Approve pending changes. No backend, no account required. The full product, live.

promptmatrix · dashboard · v3 ◎ DEMO MODE · ALL DATA SIMULATED
PMv3
// WORKSPACE
◈Dashboard
◧Prompt Studio
◫Registry 24
◉Evaluations
⑂Versions
// OBSERVE
◎Analytics
◑Trace Viewer 3
⚠Alerts 2
PM
demo_workspace
[DEMO]
SYS:DASHBOARD
DEMO
// SYSTEM_STATUS — ONLINE
prompt_intelligence · session_active
TOTAL_PROMPTS
247
↑ 12 this cycle
AVG_QUALITY
8.4
↑ 0.3 delta
API_CALLS
18.2k
↑ 4.1% T-1
COST_EST
$284
↑ $12 var
ACTIVE_PIPELINE — support-agent-v3 ● LIVE
✓
INPUT
──▶
✓
BUILD
──▶
◉
EXEC
──▶
○
EVAL
──▶
○
ROUTE
RECENT_PROMPTS
support-agent-v3
Sonnet 4.6 · 2h
9.1
content-summarizer
Sonnet 4.6 · 5h
7.8
lead-qualifier
Haiku 4.5 · 1d
8.6
CALL_VOLUME [7D]
↑ 14% avg · sonnet 52% · haiku 31%
DEMO MODE · All data is simulated · No backend required · Self-host on GitHub →
// BUILD_SCHEDULE

What ships
next.

Public roadmap. Backend first, then eval engine, then enterprise-grade controls. No vaporware — only what makes the governance pipeline more reliable and complete.

✓ COMPLETE · RELEASING NOW

Core Governance

  • ✓ Approval workflow (propose → review → approve)
  • ✓ Line-by-line diff view for engineers
  • ✓ Serve endpoint (/pm/serve/:key)
  • ✓ Version history + 1-click rollback
  • ✓ Role-based access control
  • ✓ Immutable audit trail
⚡ BUILDING NOW · Q2 2026

Backend + Eval Engine

  • → Persistent backend (real database)
  • → Python + Node.js SDK
  • → LLM-as-judge eval scoring on proposed changes
  • → Team auth + SSO foundations
  • → Email notifications for pending approvals
  • → Slack integration
◦ PLANNED · Q3–Q4 2026

Enterprise Controls

  • ◦ Enterprise SSO / SAML
  • ◦ Audit log export (CSV, API)
  • ◦ On-premise deployment option
  • ◦ SOC 2 Type II certification
  • ◦ Multi-region data storage
  • ◦ Custom approval workflow configuration
// COMMUNITY_GOVERNANCE

100% Open Source.
Strictly Local.

PromptMatrix is a community-first protocol. Download, self-host, and govern your AI systems with total autonomy.

// COMMUNITY / SELF-HOST
$0/life
MIT LICENSED · LOCAL-FIRST SQLite

The full governance protocol for local-first engineers. Fully open source, transparent, and community-driven. Run it on your iron, keep your data, own your logic.

  • ✔ Full repository & source access
  • ✔ Local-first SQLite registry
  • ✔ Unlimited seat count (Self-hosted)
  • ✔ Rule-based evaluation engine
  • ✔ AES-256-GCM Cryptographic security
  • ✔ Docker, Start-scripts, and CLI tools
  • ✔ 100% Privacy & Data Sovereignty
★ CLONE ON GITHUB →
// THE LOCAL-FIRST MANIFESTO
100%
Open Source. MIT Licensed. No proprietary hooks.
SEC
On-Prem Security. Your keys, your database, your iron.
∞
Free for teams, startups, and enterprises locally.
// FAQ

Questions
engineers ask.

Can't find your answer here? Reach out directly.

hello@promptmatrix.io →

A bare serve call adds roughly 10–15ms. In practice, you cache locally with a 30-second TTL — the overhead rounds to zero for the vast majority of requests. The SDK handles this automatically. You also keep a hardcoded fallback so your app stays live if the endpoint is unreachable. The Integration section shows the exact caching pattern.

Yes — that's the default mode. Open the HTML file in your browser, start adding prompts. Everything persists in localStorage. Your local agents and scripts can call the serve endpoint directly. No installation, no backend, no account required. The browser mode is the zero-infrastructure starting point. When you need durability across browser restarts, move to SQLite mode — one flag in the config.

There are three tiers: Browser localStorage — zero setup, start immediately, data lives in your browser. SQLite — download the repo, point it at a local file, data persists across sessions and is git-friendly. Self-hosted database — Postgres, MySQL, or any backend you already run, full data sovereignty. You can export your prompt registry and migrate between tiers at any time. Your prompts, your schema, no lock-in.

Env vars still require a deploy to change — that's the exact problem this solves. A Postgres config table is structurally closer, but it has no approval gate, no diff view, no audit log, and no interface that a non-engineer can use safely. PromptMatrix is purpose-built for prompt text governance: it understands what a "before" and "after" look like, renders a human-readable diff, and routes every change through a review decision. The storage is the least interesting part. The governance layer is the product.

All of them. PromptMatrix stores and serves text — it has no opinion about which model receives it. Fetch the prompt string, pass it to any LLM API you use, done. Works with OpenAI, Anthropic Claude, Google Gemini, Mistral, Llama via any provider, and anything else with a chat completion interface.

GitHub requires every proposer to understand Git, open a PR, and wait for a code review cycle. PromptMatrix lets non-engineers edit prompts in plain text through a UI with zero code context required. Engineers still review and approve — but the proposer never needs to touch the codebase. The workflow is designed specifically for collaboration between technical and non-technical operators.

No. By design. The approval gate is not optional — it's the product. Non-engineers can propose freely, but nothing reaches the serve endpoint without explicit engineer sign-off. This is what makes it safe to give non-engineers editing access, and it's what enterprise compliance teams require: a documented human review before any AI behavior change goes live.

Replace the hardcoded prompt string with a single API call: GET /pm/serve/your.prompt.key. Your application always receives the current approved version. When someone proposes a change and an engineer approves it, the very next call to that endpoint returns the updated content. No application redeploy. No code change in the application layer — ever again, for that prompt.

Every event is logged and immutable: who proposed the change, the previous content, the new content, who approved or rejected, and the exact timestamp for each action. Exportable as CSV or queryable via API. For enterprise customers, this is often a hard requirement before AI-powered features can be deployed to end users — "git blame on a string" doesn't pass legal or compliance review.

Multiple engineers can be designated as approvers for any prompt. Any one of them can approve a pending change. The pending change queues until reviewed — it never auto-approves, which is the entire point of the system. Email and Slack notifications for pending approvals are on the cloud roadmap.

The self-hosted version gives you the full governance system with unlimited prompts — the complete approval workflow, serve endpoint, version history, and audit trail. It's not a limited demo. Run it on your own infrastructure and wire it into a real application. Full data sovereignty, zero telemetry, and 100% open source.

// OPERATOR_FEEDBACK

Engineers who stopped hardcoding.

"Took 30 minutes to wire up. Now our PM can tweak the onboarding prompt herself and I approve it from my phone. We shipped 3 prompt iterations last week without a single deploy. This should exist for every LLM project."

ALEX K.
SENIOR ML ENGINEER · AI STARTUP

"We moved from 47 hardcoded system prompts scattered across 6 repos to one governed registry. The diff view alone caught two regressions before they hit prod. Eval gating is the feature I didn't know I needed."

SARAH M.
LEAD ENGINEER · ENTERPRISE AI PLATFORM

"As a solo builder, I finally have a single place for every agent prompt in my system. The audit trail has saved me twice when I needed to understand why my agent changed behavior. Version control for prompts is not optional anymore."

MARCUS L.
INDIE AI DEVELOPER · OPEN SOURCE BUILDER
// OFFICIAL_SDK

One line.
Prompts leave the codebase.

Install the official Python SDK. Replace every hardcoded prompt string with a single call. The SDK handles caching, fallbacks, and async — you just write pm.serve("your.key").

agent.py PYTHON
# pip install promptmatrix-sdk from promptmatrix import PromptMatrix pm = PromptMatrix( api_key="pm_live_xxxxxxxxxxxxx", ) # Before: hardcoded string # SYSTEM = "You are a helpful assistant..." # After: always live, never deploy again SYSTEM = pm.serve("assistant.system") # With variables subject = pm.serve( "email.subject", variables={"product": "PromptMatrix"} ) # Async SYSTEM = await pm.aserve("assistant.system")
// PRICING_ARCHITECTURE

Five tiers. One protocol.
Built to scale with you.

All cloud tiers are currently on waitlist — join early to lock in launch pricing. The Founder tier is hardcapped at 100 orgs with lifetime access and no subscription, ever. Self-host forever free with the MIT-licensed OSS version.

// PLAN_01 · STARTER
Starter
WAITLIST
$29 /month

Solo developers and small teams who need managed cloud prompt governance without ops overhead.

  • ✓ 100 prompts on cloud
  • ✓ 3 seats
  • ✓ 60 RPM serve API
  • ✓ Approval workflow + audit log
  • – LLM-as-a-judge evals
  • – Email notifications
LOADING.
// PLAN_02 · FOUNDING_MEMBER
Founder
LIVE NOW
$499 one-time
Lifetime access — no subscription, ever
SPOTS CLAIMED loading...

Everything in Starter — plus fully managed cloud with LLM evaluation gating, priority support, and a permanent spot in our founding cohort. Pay once. Own it forever.

  • ✓ Everything in Starter
  • ✓ 500 prompts on managed cloud
  • ✓ 5 team seats with RBAC
  • ✓ 100 RPM serve throughput
  • ✓ LLM-as-a-judge evaluation engine
  • ✓ Email approval notifications
  • ✓ Dashboard + invite flows
  • ✓ Priority email support
  • ★ Name in FOUNDERS.md forever
  • ★ Direct input on roadmap

Step 1: Create workspace · Step 2: Payment activates cloud access

// PLAN_03 · PRO
Pro
WAITLIST
$79 /month

Growing AI teams shipping multiple products who need multi-seat governance and LLM-powered quality gates.

  • ✓ 400 prompts on cloud
  • ✓ 5 seats with RBAC
  • ✓ 200 RPM serve API
  • ✓ LLM-as-a-judge evals
  • ✓ Approval workflow + audit log
  • ✓ Email notifications
// PLAN_04 · SCALE
Scale
WAITLIST
$199 /month

AI-first companies with multiple model pipelines needing production-grade governance at real scale.

  • ✓ 1,000 prompts on cloud
  • ✓ 10 seats with RBAC
  • ✓ 500 RPM serve API
  • ✓ Full LLM eval gating
  • ✓ Priority support + SLA
// PLAN_05 · ENTERPRISE
Enterprise
CONTACT US
Custom pricing

For teams needing custom seat counts, SLAs, SSO/SAML, dedicated infrastructure, or compliance contracts.

  • ✓ Everything in Scale
  • ✓ 5,000+ prompts + unlimited seats
  • ✓ Dedicated infrastructure
  • ✓ SSO / SAML
  • ✓ SLA + compliance contracts
CONTACT US →
// PLAN_00 · OPEN_SOURCE · MIT LICENSE
Community $0 /forever
Self-hosted · Your server, your data · Unlimited local prompts · Clone and run in 60 seconds
✓ Unlimited prompts (local) ✓ Approval workflow ✓ AES-256-GCM encryption ✓ SQLite + Postgres + MySQL
★ CLONE ON GITHUB →
CAPABILITY OSS STARTER FOUNDER PRO SCALE ENTERPRISE
Cloud prompts ∞ local 100 500 400 1,000 5,000+
Team seats ∞ self-hosted 3 5 5 10 Custom
Serve RPM self-hosted 60 100 200 500 1,000
LLM-as-a-judge evals Rule-based – ✓ Full ✓ ✓ ✓
Approval workflow ✓ ✓ ✓ ✓ ✓ ✓
Email notifications – – ✓ ✓ ✓ ✓
SSO / SAML – – – – – ✓
Price Free forever $29/mo
waitlist		 $499 once
LIVE NOW		 $79/mo
waitlist		 $199/mo
waitlist		 Custom
∞
LIFETIME ACCESS
Founder pays once. Uses forever. No renewal, no price hike.
100
HARDCAPPED ORGS
When the last Founder org is claimed, that tier closes permanently.
MIT
OPEN SOURCE ALWAYS
The OSS core is free and public forever. No lock-in, ever.