From zero to your first running agent in less than 10 minutes.
Please know that the early-access program is for people who want to help make MCProspero great, and are willing to endure a few hiccups along the way. We will want to get your feedback and ideas and learn what works and what needs fixin’. But PLEASE do not use this early access platform for business-critical agents, and definitely not for life-critical agents.
We’re building MCProspero in the open and we genuinely want your input. Found a bug? Have an idea for a feature? Built a cool agent you want to show off? Head over to our feedback repo:
Or email us at hello@mcprospero.ai if that’s more your style.
MCProspero works with any MCP-compatible client. We have done most of our testing with Claude (desktop, web and mobile), but have tried it in ChatGPT as well. Here is the wrinkle, at least for the Anthropic suite, you need to be on one of the paid plans. Their free tier does not allow “custom” HTTP MCP Servers. We will eventually be an approved MCP Server and then should be usable there too. But not now. Sorry. But pay for Claude, it’s worth it.
Next, if you want agents that act on your email and calendar, it needs to be Gmail. We can send to any email, but can only connect to the Google suite right now. Adding Microsoft is in the works, but not deployed yet. If you need some other provider let us know and we can put it on the roadmap.
If you want to get notified in Slack, you will need to be able to add us to your Slack instance. It’s easy, but some companies don’t let you add custom apps without approval.
During the setup phase you will need to make an account – we only support Social Login right now for these accounts. You can sign in with a Google or GitHub account. We hope to add email+magic link in the future. But not there yet. So if you don’t have a Google/Gmail account, or a GitHub account, you won’t be able to join this early access program.
You will also need to get approved for the early access program. This is to prevent too many people from signing up before we have worked out all the teething pains. If you have not done so, head over to request early access and we’ll add you to the allowlist as quickly as we can.
Once you’re in, MCProspero provides everything else — infrastructure, LLM tokens, scheduling, tool integrations. No API keys required to get started.
https://api.mcprospero.ai/mcp
Once connected, you’ll see MCProspero’s tools available in your Claude conversations. Start a new conversation and Claude will know about your agents.
https://api.mcprospero.ai/mcp
https://api.mcprospero.ai/mcpYour agents work from any device. Create on desktop, check status from your phone.
https://api.mcprospero.ai/mcp
Cursor supports MCP servers. Add MCProspero to your .cursor/mcp.json:
{
"mcpServers": {
"mcprospero": {
"url": "https://api.mcprospero.ai/mcp"
}
}
}
Then authenticate when prompted. This lets you manage agents from within your coding environment.
MCProspero works with any MCP-compatible client that supports remote servers. The server URL is always:
https://api.mcprospero.ai/mcp
Authentication uses standard OAuth 2.0. Your client will handle the redirect automatically.
Before your agents can read email, send Slack messages, or check your calendar, you need to connect those services. This is a one-time setup. From then on, all your agents can use these connections.
In any conversation with Claude (with MCProspero connected), just say:
Ask Claude: “Connect my Gmail” or “Connect my Google account”
Claude will give you an OAuth link. Click it, sign in with your Google account, and approve the permissions. This connects both Gmail and Google Calendar in one step.
Important – we have not been officially approved by Google (in the works) – so you may get a page warning you and making sure you trust us. That’s ok. This is just temporary.
You can disconnect anytime by saying “Disconnect my Google account.”
If you have more than one Gmail account, you can also say things “connect my work email” and then later “connect my personal email.” Those labels are just examples. You can use anything you want.
Once connected you can do all sorts of things. For example, I have an agent that “every hour look for unread mail from my wife that seems REALLY important and then slack me a summary.”
What will happen behind the scenes in this example: MCProspero will wake your agent up every hour, fetch all the unread from your wife, and then pass those to the AI model that is running your agent (more on that later) that will do the “is the REALLY important” and “summarise” steps.
Ask Claude: “Connect my Slack”
You’ll get an OAuth link to authorize MCProspero in your Slack workspace. Once connected, your agents can:
I have only really tested sending to channels and DMs.
We just added generic Webhook support to the platform. So you should be able to configure Slack to call your agent to wake it up. This has not been tested.
Ask Claude: “Connect my GitHub”
Authorize via OAuth, and your agents can:
You can also configure an agent to respond to a Webhook from GitHub. For example, “Send me a slack every time we finish a deploy to production, and include a summary of all the pull-requests in the deployment”
Assuming Claude gets what you are asking, you will be given a URL and Signing Key, and instructions where in GitHub to put them.
IMPORTANT You will be asked to approve your webhook. If you configure GitHub before you approve the webhook agent, GitHub will report errors.
Ask Claude: “What services do I have connected?” or “List my connections.”
Claude will show you everything that’s linked and ready to use.
Just describe what you want. Claude is your design partner — it’ll ask clarifying questions, suggest the right tools, and help you get the details right.
Try something like:
“Can you help me make an agent to Check Hacker News every morning at 8 AM and email me a summary of the top AI stories.”
Claude will:
The manifest is how MCProspero keeps you in control. You see exactly what the agent will do before it runs. Anything you did not approve gets blocked at runtime. (and will put your agent into an error state - more on that later)
Agents support both simple intervals and cron expressions:
Agents can also “wake up” on a webhook firing. If you find your self wanting to run every minute polling something, you probably want a webhook instead.
Once an agent is running, you can manage it through conversation:
If your agent stops unexpectedly, or is in an ERROR state, or did not work the way you expected, get Claude to help. “Why didn’t I get the email from my weather agent?” Claude will help you debug.
You can ask your agent to run immediately, to test it, outside of the schedule. “Let’s test the weather agent now…”
AGENT LIFECYCLE: When you first ask Claude to create an agent, you likely will be asked some clarifying questions, and then Claude will submit your agent for a “dry run.” This dry run does not actually make any external calls, but most importantly, it determines what the agent will do when you finally run it. MCProspero returns back to Claude a “manifest” that lists all the tools your agent will use and all the endpoints (web sites, …) it might go to. This puts the agent in PENDING state. Claude will then show you the manifest and you need to approve it (or not). Once you approve it, you go to a RUNNING state, and then only those tools and endpoints you approved are allowed for use. If your agent tries to use something else, your agent will stop and be put in an ERROR state. You can ask Claude what happened and hopefully get a good answer. One way this can happen, is if there is conditional logic in the agent that Claude did not tell MCProspero about during that initial dry run.
The other way an agent can go to an ERROR state is if MCProspero detects passwords, secrets, signing keys, tokens, etc. in any of the information you send or receive from any of the tools or endpoints.
Here are some agents running on MCProspero today. Use these as inspiration — or just paste them into your conversation.
“Every morning, send me today’s weather for Ridgewood, NJ and tell me which Boston sports teams are playing.”
“Check Hacker News every morning at 8 AM and email me a summary of the top AI and startup stories. Keep it to 5-7 items with a few sentence summary.”
“Every morning at 8 AM Pacific, pull from Hollywood Reporter, Variety, Deadline, and IndieWire and send me a digest of new films going into production, getting greenlit, or getting major casting attached.”
“Every morning, look at my calendar and see if there’s any recent email from anyone I’m meeting with that I should read first.”
“When a GitHub deploy finishes, summarize the commits for the #dev channel in Slack, then write a customer-friendly version and email it to marketing.”
“Every hour, check my Gmail for unread messages from my wife. If anything looks urgent, Slack me immediately.”
Your agents will work out of the box, but Claude can help you make them better. Here are some things to try:
Say: “Look at my HN agent’s recent runs. Is it working well? Any suggestions?”
Claude can read the agent’s logs and transcripts, spot issues, and suggest improvements — like better search queries, smarter filtering, or more efficient tool usage.
Agents can save state between runs using checkpoints. This is powerful for things like:
Just describe the behavior you want and Claude will set up the checkpoints automatically.
You really should tune your agents and checkpoints after your agent has run a few times. It is amazing how sometimes a small change can dramatically reduce the number of AI tokens your agent needs (which is important as the platform only gives you a fixed amount of tokens per day, depending on your plan – for early access it is 100K/day).
One of MCProspero’s unique features: you can build a watchdog agent that monitors your other agents. “Create an agent that checks once a day if all my other agents are healthy. Slack me if any are in error state.”
MCProspero uses Claude Haiku 4.5 by default — it’s fast, affordable, and great for most agent tasks. But you can upgrade to more powerful models when you need deeper reasoning.
| Model | Tier | Best for | Cost |
|---|---|---|---|
| Haiku 4.5 | Standard | Monitoring, alerts, email triage, data pulls | 1x |
| Sonnet 4.6 | Advanced | Classification, summarization, analysis | 3.75x |
| Opus 4.6 | Deep | Research, complex reasoning, nuanced writing | 18.75x |
To use a different model, tell Claude when creating or updating an agent:
You can reference models by name (Haiku, Sonnet, Opus) or tier (standard, advanced, deep). More powerful models use more of your daily AI token budget, so use them where the quality difference matters.
In addition to schedules, agents can be triggered by webhooks — external events that fire your agent on demand.
Tell Claude: “Set up a webhook trigger for this agent”
You’ll get a unique webhook URL and signing secret. Point your external service (GitHub, Stripe, your own app) at that URL, and your agent runs whenever the webhook fires.
Agents can use both schedule and webhook triggers simultaneously — for example, a daily summary on schedule plus immediate runs when a GitHub issue is created.
Webhook payloads are validated (HMAC-SHA256 signature), size-capped (4KB), and scanned for sensitive data before being passed to your agent.
During the preview, MCProspero provides platform-hosted LLM tokens at no cost. To keep things fair, there are daily limits:
| Limit | Preview |
|---|---|
| Agents | Up to 10 |
| Daily budget | 100,000 cost units (resets at midnight UTC) |
| Tool calls per run | 20 |
| Platform email sends | 10 per agent per hour |
What are cost units? One cost unit equals one Haiku token. If you use Sonnet (3.75x) or Opus (18.75x), each token costs proportionally more units. External tool calls (HTTP requests, email sends) also consume a small number of units.
Check your usage anytime: “What’s my budget status?” or “How much have I used today?”
If you hit your daily limit, agents will pause and resume the next day. No data is lost.
We do support bring-your-own API Key, to let you use AI models on the provider platforms. Contact support for instructions.
Your agents run in a managed environment with a few guardrails worth knowing about:
Manifest enforcement: Your agent can only use the tools and endpoints you approved in the manifest. If it tries to reach something new, it gets stopped. Update the agent and re-approve if you need to expand its access.
If your agent hits any of these limits, ask Claude what happened — it can read the logs and help you adjust.
Ready to build something? Open Claude, connect MCProspero, and describe your first agent. Claude will take it from there.