Grigoriy Dobryakov

Expert MCP

Home
Live MCP server dobryakov-expert

Connect me as MCP to your Claude/Codex

Want to talk through a project — but no time, awkward to bother, or it turns into paid consulting? Plug in the MCP: chat «with me» in your agent or wire it into your pipeline. Free first pass; if there's fit — next step is paid work: consulting, diagnostic, or a hire.

Claude Codex Cursor Free
Connect Why use it

Why plug it in

A way to talk «with me» before you spend time on a call or budget on consulting. Free, at your pace, in your own agent.

  • Assess fit before an interview or contract

    Map a role, JD, or your project to my cases, products, and scale — without «let's jump on an hour-long call».

  • Discuss your project with no obligation

    When there's no time, it's awkward to bother, or you're afraid the talk turns into expensive consulting — test the angle in chat first.

  • Feel the communication style

    Before live work, see how I ask questions, cite examples, and flag risks — that's part of fit too.

  • Wire into your pipeline

    An outside view on a task wording, ticket, or architecture choice — when you need an expert angle, not a call on every step.

If you see fit after that — next step is paid work: consulting, diagnostic, or a hire.

How to use it

  1. 1

    Pick a strong model in your client

    Quality comes from your LLM, not the MCP server. Details and model recommendation — in Good to know below.

  2. 2

    Connect the MCP in your agent

    One snippet in .mcp.json — see Connect. Works with Claude, Codex, and other MCP-capable clients.

  3. 3

    Ask in your own words

    In chat — about fit, a case, approach. In a pipeline — pass task context and ask what I'd say in your shoes. You can ask the agent to run screening-em or case-walkthrough.

  4. 4

    Decide if you want the next step

    MCP is a free getting-to-know contour. If it clicks — email grigoriydobryakov@gmail.com for a live conversation or a commercial format.

Connect

Console client — one command. Or a snippet in .mcp.json and restart the agent.

Console (Claude Code, Codex CLI)

claude mcp add --transport http dobryakov-expert https://mcp.dobryakov.net/mcp

.mcp.json file

"dobryakov-expert": {
  "type": "http",
  "url": "https://mcp.dobryakov.net/mcp"
}

After connecting, ask your agent: "use dobryakov-expert and answer…"

Example situations

Not documentation templates — real reasons people show up.

Hiring a VP Engineering

"Is there fit for a PE-backed scale-up with an AI mandate? Which cases are relevant?"

Considering consulting

"How would Grigoriy approach AI transformation in enterprise with legacy? What's applicable from a large retailer with an event-driven landscape?"

Need a view on a task in flow

"What would Grigoriy say about this Jira ticket?"

Want the voice before a call

"Walk through the ML platform case — tens of millions of profiles — and explain how you'd reason about production at our stage."

What you can ask about

Answers draw on published materials — the same as on the website: products, cases, segments. No grey area, NDAs, or client internals; the agent pulls only the slice that fits your question.

Roles and products

Engineering Manager · AI Head of Engineering · AI Architect — when and why

Cases with numbers

Enterprise retail · ML platform · SaaS transition · integrator turnaround — context, actions, outcomes (as on the website)

Company types

AI-committed · Enterprise · Scale-up · Traditional — where my experience is closest

Engagement formats

Hire, fractional, AI Engineering Diagnostic — when each format fits

Good to know

Answer quality comes from your LLM, not the server

MCP serves context and tools from the public corpus. Connecting facts, asking follow-ups, and wording the conclusion is the job of the model in your client. My server doesn't think for you and won't fix a weak model.

Cheap or weak models are a waste of time: they'll truncate context, skip tools, and average out the answer — then you'll decide "MCP doesn't work." Recommendation: at least Claude Opus with High effort (or an equivalent reasoning tier in your client).

  • This is a simulation from public materials, not a live conversation with me.
  • Your data does not reach the server and is not logged.
  • No secrets or unpublished content in the corpus — including client internals; only what's already public.

Next steps

MCP is a free first contour. If you see fit after it — next step is paid work: consulting, AI Engineering Diagnostic, or a hire.