MCP workflow planner for AI agents
Plan AI agents you can inspect — before you build.
Describe your agent goal in plain English and get an implementation-ready workflow draft from registry-backed building blocks — with unsupported steps, safety gaps, and approval needs called out before you build.
Works with the agent clients you already use
What you can build
Real agents, planned in seconds
Describe the goal — OrchestrateMCP designs the workflow. Here are a few it plans today.
Invoice processing agent
“Process 100 invoices in parallel, roll back the batch if any fails.”
Payment → Slack alerts
“When a Stripe payment lands, post it to the team's Slack.”
PR review bot
“Review pull requests, flag risks, never edit the code.”
Lead → CRM agent
“New lead in → enriched note written straight to the CRM.”
Content repurposing
“One blog post → three social variants, you approve before posting.”
Site / price monitor
“Watch a page on a schedule, alert the moment it changes.”
Why OrchestrateMCP
Faster to build. Harder to break.
Start from documented building blocks, keep the evidence level visible, and know which parts still need proof.
Registry-grounded starting point
Plans draw from the documented component registry, then show unsupported pieces instead of silently filling the gaps.
Designs the hard parts
Loops, parallel fan-out, retries, rollbacks, and approval gates are included in the design when the goal needs them, with evidence state and required controls kept visible.
Know what can break
It flags unproven steps, missing safety, and approval needs before implementation. The finished workflow still needs integration testing and review.
Honest positioning
What it is — and what it isn't
OrchestrateMCP does one thing well. Knowing what it doesn't do is just as important.
It IS
- A build-time advisor — plans the workflow before you write code
- Grounded in 65 documented components and 156 registry connections
- Honest: flags the unproven steps and safety gaps explicitly
- Stateless, read-only, open-source — you can audit every recommendation
- Deterministic tool logic — client wording and rendering may differ
It is NOT
- ✕An agent runtime — it never executes your workflow
- ✕A credentials vault — it never stores or touches your secrets
- ✕A coding assistant — it plans architecture, not implementation
- ✕An LLM judge — no model calls, no inference, fully deterministic
- ✕Exhaustive — 65 components cover the most common agentic patterns; novel integrations need to be added to the registry first
How it works
From idea to grounded plan in three steps
Connect the MCP
Add one hosted URL to ChatGPT, Claude, or Cursor. Thirty seconds, no install.
Describe your agent
Say what you want it to do, in plain English. No diagrams, no jargon.
Get a plan, risks flagged
A grounded, ready-to-build workflow — with the unproven and unsafe parts called out.
Connect in 30 seconds
Pick your client. One URL, no install, no keys.
- 1Open Settings → Connectors → Advanced and turn on Developer Mode.
- 2Back in Connectors, click Create and choose a remote MCP server.
- 3Paste the server URL below and save.
- 4Start a new chat and ask it to plan a workflow — it'll call OrchestrateMCP.
https://mcp.orchestratemcp.dev/mcp
One hosted URL — no install, no keys, nothing to deploy. Works the same in every client.
Try this first
Paste one goal and get a product card
These prompts hit current validated paths: brief steps, connections, host/monitor choice, safety note, and one recommended next action.
Competitor price monitor
“Build an agent that checks 5 competitor pages every morning, detects price changes, and sends me a Slack summary. I want to approve before anything external is changed.”
Gmail lead to CRM
“On each new Gmail message, detect sales leads, draft a reply, write a CRM note, and notify the sales Slack channel after a human approves.”
Read-only PR reviewer
“When a pull request opens on GitHub, review the diff for bugs and risky changes, notify reviewers with a summary, and never edit or commit code.”
Why you can trust the plan
Grounded, deterministic, open
65
documented components
156
connections with test references
12
published playbooks
7 / 7
public benchmark prompts passed
Deterministic
The same versioned inputs produce the same structured tool result. ChatGPT, Claude, and Cursor may phrase or render it differently.
Grounded
Built from 65 documented components and 156 registry connections. Every published connection carries a test reference; that is not end-to-end production certification.
Stateless & open
Read-only, stores nothing, holds no credentials — and fully open-source, so you can audit every recommendation.
Built with OrchestrateMCP
Shipping something with it? Show it off.
Be one of the first to share what you built. Submissions are reviewed before they go live.
Keep it free
This stays free. Here's how to keep it alive.
No paywall, no pro tier. If OrchestrateMCP saved you time, any of these helps — none is expected.
Star on GitHub
The cheapest way to help — it makes the project easier for the next builder to find.
Star on GitHub →Sponsor
Recurring support keeps the hosted endpoint always-on and free for everyone.
GitHub Sponsors →Buy a coffee
A one-off thank-you. No tiers, no perks gated behind it — just appreciated.
Buy Me a Coffee →Common questions
Frequently asked questions
Still curious? The source is open — read it.
Is it really free?
Yes — free and open-source, forever. No paid tier, no paywall. If it helps, a GitHub star or a small sponsorship keeps it alive, but nothing is gated.
Do you store my data or workflows?
No. OrchestrateMCP is stateless and read-only. It never stores your prompts, never holds credentials, and never calls your tools — it only recommends a design.
Which clients work?
Any MCP-capable client. Claude and Cursor connect directly; ChatGPT works via Developer Mode. Same hosted URL in each — see the connect guide above.
How is this different from just asking the model?
OrchestrateMCP grounds the plan in a versioned registry, keeps unsupported demand visible, and returns deterministic structured tool results. Your client may still phrase or render those results differently.
Can I self-host?
Yes — clone the open-source repo and run it locally over stdio, or deploy your own Cloudflare Worker. The hosted URL is just the convenient default.
Plan your next agent with your eyes open.
Paste one URL into your AI client and describe what you want. Free, open-source, nothing stored.