Ask a human expert (MCP server)
A remote MCP server that lets you, or your AI agent, bring in a human expert when stuck on a
problem that needs a creative solution. I'm Cristian Desivo: mathematics and computer science
background, 10 years taking projects from start to finish, specialized in mathematical
optimization and AI systems. I read every request and write every reply myself.
- Endpoint:
https://cristiandesivo.com/mcp(Streamable HTTP, no auth) - Registry:
com.cristiandesivo/ask-a-humanin the official MCP Registry - Terms as a page: https://cristiandesivo.com/ask/ (also as Markdown and JSON)
This repository only documents the server; the server itself runs on my website.
Connect
Claude Code```bash
claude mcp add --transport http ask-a-human https://cristiandesivo.com/mcp
**Cursor** (`~/.cursor/mcp.json`)```json
{
"mcpServers": {
"ask-a-human": { "url": "https://cristiandesivo.com/mcp" }
}
}
VS Codebash
code --add-mcp '{"name":"ask-a-human","type":"http","url":"https://cristiandesivo.com/mcp"}'
Claude Desktop or claude.ai: Settings → Connectors → Add custom connector, with the URL
https://cristiandesivo.com/mcp.
Tools
| Tool | What it does |
|---|---|
get_status |
Whether I'm online right now, or when I'm back, and whether requests are paused |
get_terms |
The full terms: price, hours, reply times, how to pay, how to submit |
submit_request |
Sends your problem (a summary of up to 300 words) with the payment's transaction ID (PayPal, or crypto_tx and network); returns a request_id |
get_reply |
My reply to a request, by request_id: waiting until I've replied, then the text |
send_followup |
Continues in the same thread (answer my question, clarify, ask the next one), at no extra cost |
find_request |
Recovers a lost request_id from the payment's transaction ID and the reply email you used (both must match) |
How a request works
- Check
get_statusandget_terms. - Pay, one of:
- 20 USD by PayPal at https://paypal.me/CristianDesivo/20USD, or sent by PayPal to
cdesivo92@gmail.com. An agent with access to a PayPal account (for example through
PayPal's Payouts API) can do this on its own, with no human.
- 25 USDT on TRON (TRC-20) from a crypto wallet, toTJCSWgHwkQ4gYTotSQxz4dVGcgnkecGtZ5.
Only USDT on the TRON network can be matched;get_termsalways lists the current address. - Submit with
submit_request: the payment, an email for my reply, and a summary of the
problem in 300 words or fewer (what it is, what has been tried, the key constraints, the
deadline).
- Paid by PayPal:transaction_id. If PayPal shows you a different transaction ID than it
shows me,payer_email(the PayPal account that paid) lets the payment still be matched.
- Paid in USDT:crypto_tx(the transaction hash) andnetwork: "tron". The transfer is
checked on the chain right away, so the request reaches me within seconds. - Read my reply with
get_replyand therequest_id. The reply is also emailed to the
address you gave, and a confirmation goes there right away. - Follow up with
send_followupif you need to; a new problem is a new request. - Lost the
request_id?find_requestgives it back from the payment's transaction ID and
the reply email. It's also in the confirmation email.
Payments are checked before a request reaches me (PayPal against PayPal's own records, USDT on
the TRON blockchain), and one payment pays for one request.
Terms in short
- What you get: my most creative solution idea for the problem, based on your summary.
- Price: 20 USD per request by PayPal, or 25 USDT on TRON. If the problem needs more (reviewing the full context, research,
a prototype or an implementation), my reply says so and includes a tentative budget; taking it
is optional. - Hours: every day, 09:00 to 24:00 UTC. While I'm online I reply within 15 minutes of
receiving your request, once your payment shows up; otherwise by 09:30 UTC. - Limits: 300 words per summary or follow-up; 5 requests, 10 follow-ups and 120 reply checks
per hour from one address. - Cancelling: requests can't be cancelled once submitted; if I can't help with it, I'll say so
in my reply. - Not a fit: physical or on-site tasks; anything illegal or meant to deceive people.
Without MCP
The same actions over plain HTTP:
| Action | Request |
|---|---|
| Status | GET https://cristiandesivo.com/status.json |
| Submit | POST https://cristiandesivo.com/ask/submit/ (JSON or form fields) |
| Read the reply | GET https://cristiandesivo.com/ask/reply/?request_id=... |
| Follow up | POST https://cristiandesivo.com/ask/followup/ with {"request_id", "message"} |
| Recover a request ID | POST https://cristiandesivo.com/ask/find/ with {"transaction_id", "reply_email"} |
Privacy
I read your summary and follow-ups myself; they travel by email. Don't include passwords, keys
or other secrets. Full privacy policy: https://cristiandesivo.com/privacy/
Example
Illustrative: a made-up request and the kind of reply I write. Not a real client.
Request: We schedule about 40 jobs a day on 6 CNC machines. Each job has a due date, can run
only on some machines, and has a setup time that depends on the material of the job before it.
Our time-indexed MILP solves 10 jobs fine, but at 40 jobs it hits the 10-minute limit with a 30%
gap, and its plans are worse than our planner's spreadsheet. We tried tighter big-M constants and
symmetry-breaking constraints. The schedule has to be computed in under 5 minutes, before the
6 am shift. What should we change?
Reply: Your model is spending its effort in the wrong place. With setup times that depend on
the previous job, a time-indexed MILP grows with the horizon and with every pair of jobs, and
most of that 30% gap comes from a weak relaxation, not from missing constraints. Two changes I'd
make:
- Model it as intervals in CP-SAT: one optional interval per job and eligible machine, and a
circuit constraint per machine that orders its jobs and adds the setup time between
consecutive ones. Minimize weighted lateness. CP-SAT is much stronger on this structure. - Give it a good start: pass the planner's spreadsheet schedule as a hint, and batch jobs by
material first, since changeovers between materials dominate. That hands the solver a
near-optimal skeleton to improve.
At 40 jobs and 6 machines this should beat the spreadsheet well within your 5 minutes. If you want
it built and tested against a few weeks of your past schedules, I'd send a fixed quote after
seeing a sample of the data.
Track record
- Reviews from past clients, on Freelancer.com: https://www.freelancer.com/u/cdesivo92
- Container loading optimizer: plans how to load mixed cargo into as few containers as
possible, under weight, upright-only and stacking rules. Demo: 204 items in 2 containers instead
of 3. I built a similar tool for a client. https://cristiandesivo.com/case/optimization/ - Factory scheduling optimizer: assigns orders to assembly bays and tasks to workers, hour by
hour, with dependencies, trades and parts deliveries. Demo: late orders down from 10 to 3.
I've built schedulers like this for clients. https://cristiandesivo.com/case/optimization/ - LLM features that can't make things up: a library where the model only picks one of the
owner's approved answers, word for word, or says nothing. https://cristiandesivo.com/case/llm/ - Live demo, answers only from approved content: https://cristiandesivo.com/ask_question/
- Case studies on GitHub: https://github.com/crisdesivo/case-studies
Contact
- Website: https://cristiandesivo.com/
- Longer projects: https://cristiandesivo.com/hire_me/
