Przemysław Zalewski

For companies rolling out Claude or ChatGPT

Your company systems inside Claude and ChatGPT

You bought Claude or ChatGPT for the team. Now people paste data from the ERP into the chat by hand. A connector lets the assistant look things up and prepare changes in your own systems, on the logins you already have, asking before anything it cannot undo.

See price and start date →

Hourly at $36⁠–⁠54/h excl. VAT. The first connector takes one to four weeks. Details in the configurator →

Przemysław Zalewski

Sound familiar?

MCP connector problems I solve in 2026

Can an MCP connector show one customer’s data to another?

On 14 July 2026 an advisory for n8n‑MCP (CVE‑2026‑54052) disclosed that in multi-tenant HTTP deployments workflow version history was not isolated per tenant, so an authenticated tenant could read or delete other tenants’ backups. It was fixed in 2.56.1.

Every call goes through your existing API and each person’s own login, so the assistant sees exactly what that person may see. My last MCP server routed every operation through the product’s existing REST controllers for the same reason.

Is it safe to install a community MCP server?

On 20 August 2026 GitHub published a malware advisory for the npm package mcp‑server‑github: any computer that has it installed or running should be considered fully compromised.

For your own systems I write the connector: a small TypeScript server you own, with a handful of tools, no third-party bridge in between, and dependencies updated in waves you can revert. Sandtime’s users connect to its own MCP server, whose tools I shaped and stripped of destructive actions, so nobody needs a community bridge to reach their data.

Sound like you? Book a free 30‑min call →

Our MCP server eats the model’s context window

The MCP client guide for the 28 July 2026 spec revision says tool definitions from many servers can fill most of the context window before the model reads the message: about 150,000 tokens loaded up front against about 2,000 when tools are discovered on demand.

I start from the three questions people ask most and give each one a tool, instead of one tool per API endpoint. Usage per tool is measured from day one, so unused tools get removed. In 2026 I trimmed Sandtime’s MCP responses so listing projects no longer drags every member into the model’s context, and activities come back with times and durations already computed, so the assistant has nothing to recalculate.

How do we get our product into ChatGPT and Claude?

OpenAI’s help article on plugins, updated 26 October 2026, says a ChatGPT plugin bundles connected apps and skills, and that new custom GPT creation ended on 26 October 2026 for affected Enterprise workspaces.

I build one MCP server on your existing login that works in both ChatGPT and Claude, with adoption measured per client. My previous one had its first external users within a few weeks. In August 2026 I finished and hardened OAuth 2.1 for Sandtime’s MCP server, and since October 2026 connecting Claude, ChatGPT, Claude Code or Cursor takes one URL and a sign‑in instead of an API key in a config file.

Facing one of these? Book a free 30‑min call →

Scope

What the connector does

Narrow tools

A handful of tools for the questions your team asks every day: find an order, check stock, draft an invoice. Not every endpoint of the API.

Your existing login

People use their own accounts, so the assistant sees only what each person may see in the system.

Asks before it acts

Anything that cannot be undone, like deleting, refunding or sending, is marked so the assistant asks the person first.

Measured

Usage per person and per tool from the first day, so you see whether it is worth extending.

Process

How it works

  1. Pick three questions

    We start from what people ask most often and which system holds the answer.

  2. One to four weeks

    The connector on your API or database, deployed where your team works.

  3. Rollout

    A short guide for the team and a usage dashboard.

  4. Then more tools

    New processes become new tools, on a monthly care plan if you want.

Systems I connect most often

  • Your own API or database, including apps built with AI.
  • ERPs, CRMs and ticket systems with an API.
  • Mailboxes, documents and spreadsheets the team already uses.

Track record

Proof

The first case is the kind of connector you would get: built on an existing product’s login and API, with adoption measured per client from day one.

An MCP server from first merge to first external users

The problem
Users of Sandtime.io, the time tracker I built at Sanddev, used by 4,000+ freelancers and teams, wanted to log time from their AI assistants. The API existed, the agent surface did not.
What I did
I set the tool surface, removed destructive tools, routed every operation through the existing REST controllers so an agent gets no wider access than a person, and instrumented adoption per client in PostHog from day one.
Result
First external users. A bot-protection rule that rejected MCP clients by user agent was found and removed the same day because the dashboard showed it.

MCP OAuth adoption dashboard

See also: AI & agents · PostHog & product analytics

An AI coding agent as a team member, under enforced rules

The problem
Coding agents write code faster than a team can review it. A team of four wanted the speed without a slop swamp on main.
What I did
I authored the rule set shared by Claude Code and Codex: mechanical checks over written rules, lint and test gates for security, API surface and CSP, agents working on main with hooks, every PR reviewed by a human, AI‑assisted content disclosed as such.
Result
In 2026 about two thirds of merged pull requests across two products were built with coding agents, and nearly all went through my review or merge. I merged 260+ PRs of my own and reviewed 300+ of other engineers’. My commits across those two products went from under 1,000 in 2025 to 7,500+ by October 2026.

~2/3 PRs with agents 300+ reviews 7,500+ commits in 2026

See also: AI & agents · Full‑stack & custom software

Offer: AI engineer for your team (Node.js, LLM agents)

AI content platforms: R&D and an agentic content platform, quality measured not assumed

The problem
Thousands of articles were needed, with AI disclosure built in and without publishing slop.
What I did
Late 2024 I built the agent driver: enforced stop conditions, tool‑call limits, parallelization, delegation chains. Quality came from best‑of‑n, LLM‑as‑judge, scoring against fixtures and plagiarism checks, with many models evaluated across their releases.
Result
Thousands of articles shipped through the platform. Content deployed automatically across 300+ generated sites with git‑backed bulk operations and reputation monitoring. From November 2023 to May 2025 it was a two‑person team. Invented author bylines came off all 300+ sites within a week, and one measured weekend in production saw about 94% of runs succeed at roughly a dollar per article.

thousands of articles 300+ sites LLM‑as‑judge

See also: AI & agents · AI in your company

Want the number before the call?

Hourly at $36⁠–⁠54/h excl. VAT. The first connector takes one to four weeks.

See price and start date → Book a free 30‑min call

Good to know

MCP details that decide whether a connector ships

The 2026-07-28 spec is stateless

The current MCP revision removed the initialize handshake and protocol sessions; each request carries its version (MCP changelog).

Dynamic Client Registration is deprecated

Clients move to Client ID Metadata Documents, and Dynamic Client Registration can be removed from revisions released on or after 28 July 2027 (MCP deprecations).

Never pass the user’s token through

The 2026‑07‑28 spec still says an MCP server must not pass through the token it received to your ERP or API; upstream calls use a separate token issued by that system (MCP security considerations).

Annotations decide confirmations in Claude

Read‑only tools run without per‑call confirmation and destructive tools always prompt. Directory tools need a title and a name of 64 characters or fewer (Claude review criteria).

ChatGPT wants all three hints set

OpenAI’s app review requires readOnlyHint, destructiveHint and openWorldHint as explicit booleans on every tool, plus a demo account (OpenAI guidelines).

Errors belong in the tool result

Errors from your system should return as a tool result with isError true, so the model can fix the input and retry (MCP tools).

FAQ

Questions

Does it work with Claude Team or Enterprise and ChatGPT?
Yes. MCP connectors are supported by Claude and by ChatGPT; the same server serves both.
Will the assistant see data a person should not see?
It uses each person’s own login, so it sees what they can see in the system and nothing more.
Is there a live MCP server of yours I can connect to?
Yes. ciunkos.com/mcp is a read‑only MCP server over my data set of Polish software houses and IT jobs: four tools, no key, input and result size capped, no AI and no paid services behind it. Add it as a remote HTTP server in Claude or Cursor.
Can we try it on our API first?
Yes, paste your OpenAPI spec into the free OpenAPI to MCP tools generator to see the proposed tools.
How long does it take?
One to four weeks for the first connector, depending on the number of tools and the login.
What does it cost?
Hourly, 36 to 54 USD, depending on how complex the system is. The configurator shows the price for your scope before we talk.
Who keeps it working when MCP changes?
You own the code: a small TypeScript server in your repository, with a runbook, and the copyright transferred in writing. The spec keeps moving (the 2026‑07‑28 revision dropped sessions, Dynamic Client Registration is deprecated), so you can keep me on a monthly care plan for updates and new tools, or hand it to your team with the runbook.
Why does Claude ask for confirmation on every call to my connector?
Claude uses tool annotations: readOnlyHint tools run without a prompt and destructiveHint tools always ask (Claude review criteria). Annotations are hints, not security, so authorization stays on the server (MCP tools).
Can an MCP connector move money?
Not in Anthropic’s directory: connectors that transfer money or other financial assets are not accepted (Claude review criteria). Local stdio servers read credentials from the environment instead of OAuth (MCP authorization).
Can an MCP connector show one customer’s data to another?
On 14 July 2026 an advisory for n8n‑MCP (CVE‑2026‑54052) disclosed that in multi-tenant HTTP deployments workflow version history was not isolated per tenant, so an authenticated tenant could read or delete other tenants’ backups. It was fixed in 2.56.1. Every call goes through your existing API and each person’s own login, so the assistant sees exactly what that person may see. My last MCP server routed every operation through the product’s existing REST controllers for the same reason.

In 30 minutes you will know if we are a fit

You bring the problem, I bring the questions. Want a number first? The configurator shows hours, a price range and the start date.

Book a free 30‑min call Email me instead

or write to ciunkos@gmail.com

After you book you get a calendar invite. Nothing else to fill in.

My pledge

  • Reply within 24 hours. I reply within 24 hours on working days, personally, not a bot. Weekend messages: by Monday noon.
  • Late means free. If I am late, the first hour of work is free. You do not have to ask.
  • Scope in writing within 24 hours. A free 30‑minute call with no company deck, then the scope and the price in writing within 24 hours. You decide.
  • The person who does the work. You talk to the person who does the work. No sales layer, no handoff.
  • Follow‑up only with news. I follow up only when I have something new for you. Never a “did you see my offer”.
  • Honest “not a fit”. If I am not the right person, I say so on the call and point you to someone who is.

Calls this week · work can start now · B2B contract · Wrocław, Poland (CET), remote

You own the code, domain and accounts from day one. Nothing is held hostage.