You will get your MCP server fixed so it installs and runs on a clean machine
Project details
It works on your machine. It fails on everyone else's.
This is far more common than people think. While auditing public MCP servers I found three that
would not start at all on a clean install - and two of them were official reference servers from
major companies. The cause was the same every time: a dependency declared with a lower bound and
no upper bound, a major version shipped, and the server died on startup with an import error
that had nothing obviously to do with dependencies.
WHAT I DO
• Reproduce the failure in a genuinely clean environment, not a cached dev machine
• Find the actual cause - usually dependency pinning, but not always
• Fix it, pin it properly, and verify a from-scratch install works
• Send you a short note on what broke and why, so it does not happen again
Cheap and fast, and it removes the most embarrassing failure a published server can have:
somebody tries it for the first time and it does not even start.
This is far more common than people think. While auditing public MCP servers I found three that
would not start at all on a clean install - and two of them were official reference servers from
major companies. The cause was the same every time: a dependency declared with a lower bound and
no upper bound, a major version shipped, and the server died on startup with an import error
that had nothing obviously to do with dependencies.
WHAT I DO
• Reproduce the failure in a genuinely clean environment, not a cached dev machine
• Find the actual cause - usually dependency pinning, but not always
• Fix it, pin it properly, and verify a from-scratch install works
• Send you a short note on what broke and why, so it does not happen again
Cheap and fast, and it removes the most embarrassing failure a published server can have:
somebody tries it for the first time and it does not even start.
AI Algorithms
Regression AnalysisAI Applications
Sentiment AnalysisAI Development Language
PythonAI Models
OpenAI CodexWhat's included $100
These options are included with the project scope.
$100
- Delivery Time 2 days
- Number of Revisions 1
- Source Code
Optional add-ons
You can add these on the next page.
Fast 1 Day Delivery
+$40About Junaid
MCP Servers & AI Agent Tools | Python, FastAPI | I also audit them
Lahore, Pakistan - 12:53 am local time
Lately, most of my work has been with LLMs and AI agents:
- MCP servers: the piece that lets AI agents like Claude actually use your tools and data
- Auditing MCP servers: finding out what breaks before your users do
- RAG systems: "chat with your documents" assistants that answer from your files and cite the
exact source, so nothing gets made up
- Python / FastAPI / Flask backends and REST APIs
Here is the part that makes me different from everyone else on this page.
I wrote mcp-probe, an open-source tool that audits MCP servers. It connects to a server,
discovers every tool, then generates attacks from each tool's own schema: missing required
fields, wrong types, empty strings, 100,000-character strings, injection-shaped payloads. Then
it scores what survived.
I have run it against 13 publicly installable MCP servers - including ones maintained by
Anthropic, Google, Microsoft and MongoDB.
What I found:
- Two servers had real input-validation gaps. Both were reported privately to their maintainers,
through the channel each of them asked for. I am not naming them until they have had time to
respond.
- Three more would not start at all on a clean install, including two official reference
servers. Their code was fine. Their dependency pinning was not.
- Five false positives in my own tool. I fixed them and wrote all of it up publicly, because a
scanner that fails well-built servers is worse than no scanner at all. Existing tools in this
space run roughly a 78% false-positive rate.
That last one is the point. Anyone can run a scanner and send you a list of scary-looking
findings. Knowing which of them are real is the actual skill, and I have shown my working in
public: github/junaidshahid-dev/mcp-probe
Why this matters for your project: MCP tools are called by AI agents, and agents produce
malformed arguments constantly. That is the normal case, not the edge case. A tool that trusts
its inputs will break in production, and it will break in a way that is very hard to trace back.
I build servers that validate properly, and I hand you the audit report that proves it.
How I work:
- I ask specific questions before writing code, so you don't pay for a wrong guess
- Tests and documentation come with the job, not as an upsell
- You get the full source, and it runs on your infrastructure, not mine
- If I think your idea has a problem, I tell you before you pay me, not after
Computer Science graduate, FAST-NUCES. Based in Lahore, and I work across US and EU hours.
Tell me what you're building or what's breaking, and I'll tell you honestly whether I'm the
right person for it.
Steps for completing your project
After purchasing the project, send requirements so Junaid can start the project.
Delivery time starts when Junaid receives requirements from you.
Junaid works on your project following the steps below.
Revisions may occur after the delivery date.
Reproduce it on a genuinely clean machine
Fresh virtualenv or container, no cache, nothing already installed. Most of these bugs are invisible on a dev machine because the working version is already sitting there.
Find the actual cause
Usually an unpinned dependency that shipped a breaking major version, but not always. I read the traceback rather than guessing.