Tendem MCP Connector - Call a human expert like any other tool — inside your AI

Hi everyone! Egor here, Head of Engineering at Tendem by Toloka 👋

You know that little “AI can make mistakes. Please double-check responses” line under every chat box? We built something for exactly that.
Your AI can now call a real human the same way it calls any other tool with the new Tendem MCP connector.
AI agents are great at planning and chaining tools together. But some tasks need real judgment and taste, and that’s where AI on its own falls short. So when your agent hits that wall, you just say “Use Tendem” and it hands the work off to a real human expert.
It’s a remote MCP server, not a local setup, so it runs right where you already work — Claude, Cursor, Codex, ChatGPT, OpenClaw, Hermes and more.

How it works: the AI scopes your task, routes it to a domain expert, and you get human-verified output back in the same chat. No switching tabs, no managing freelancers, no doing the last mile yourself.

The whole point is less slop: real human experts working inside your AI models and agents. As long as that “please double-check responses” disclaimer is still around, we think there’s a human who should be in the loop.

Happy to talk shop in the comments 🛠️

Sharing a few lessons we picked up while building this connector. A technical note for anyone building MCPs - especially the async kind.

1. Boring protocol features win.
We deliberately skipped some shiny stuff - the async tasks SEP, elicitation, sampling. Host support for those is a patchwork that changes without warning. The lowest common denominator across every client is plain tools + polling, so we built on plain tools + polling. Less elegant, but runs everywhere. Additionally, some clients support scheduling, so polling becomes even more convenient for communication.

2. The agent talks to OUR agent (A2A).
`create_task` doesn't dump a ticket into a queue. It opens a conversation with our orchestrator, which asks clarifying questions like a decent PM would (including structured ones that are nicely displayed in the chat UI). The host agent answers from context on its own and only interrupts the user when money, scope, or a deadline is actually on the line. You can even ask "can you do this, and roughly how much?" before spending a cent. That capability-discovery step is the most underrated thing we built - knowing whether the work fits before actually paying.

3. Assume the model will retry, stall, and lose its place.
Models re-issue tool calls, so every mutation needs an idempotency key implemented in one way or another. Chat uses an offset cursor: send a message with a stale cursor and the server refuses to post it, handing back what you missed - so the agent re-reads and re-decides instead of talking over a reply it never saw. Race conditions on the agentic level :)

4. Spoon-feed the next step.
Never assume the model will derive state/next actions from a status enum. Enums describe "where the workflow is", but it's hard to tell "what to do next" from them alone. Every response carries next_action + when to poll + when to give up + a plain-language "guidance" string. The dumbest host and the smartest host follow the same breadcrumbs - they know exactly what's going on, what to do next, and when to escalate.

5. Money has no home in MCP.
There is no payment primitive in the protocol (some people would disagree, but it's true; there's no widely adopted one, yet). So approval and balance live in tools, the actual payment lives on our hosted page, and "insufficient balance" is NOT an error - it's a normal branch that returns a link which auto-approves the task once paid. Modeling it as an error would break every host that treats errors as dead ends.

6. The store review reshaped the product, not just the README.
To get into the Anthropic and OpenAI directories you don't just upload a manifest. You need public docs that match the connector exactly - every tool name, 3+ real example prompts, connection steps, troubleshooting, supported platforms, privacy policies, test accounts, and whatnot. Zero marketing language (literally: "describe what it does, not how great it is"). And the requirement that changed our positioning: you can't sell AI-generated output as the deliverable on these marketplaces (which is a bit questionable given the apps present in the ChatGPT marketplace). The human expert is the product; the AI only scopes and routes. We rewrote a lot of copy around that single line.

Bytes, by the way, never touch MCP - files move over pre-signed URLs, the server stays stateless, and "recovery" after a client restart is just "list my tasks," not "resume my socket." Uploading files from your agent for analysis is a bit harder than "downloading results," but it's already supported here as well.

I may be biased but.. I was really impressed by the simplicity of submitting my tasks to Tendem from Claude Cowork. I was literally just working in my session, typical workflow (needed to build a google ads campaign), got the Claude-generated results, and then when I needed to validate and refine them, I just asked Tendem to do it. It picked up my task, and what I got back a few hours later was better than I would have done it by myself. I felt confident that another human expert reviewed it and proposed structural improvements to the ad strategy. I saved hours of my time and it felt good.

Guys, I'm telling you, once you wire this into your Claude, Cursor or ChatGPT you don't go back. Becoming more and more AI-native and shipping autonomous agents is the craze that's humming along, and "human in the loop" has basically become the base principle by now. Personally I want a future where AI gets things right — with a human QA layer at its center

congrats on the launch Team Toloka 👏

MCP is a big value add for me, thank you! Having the human part join me where I'm already doing the AI part in my default place of work is super convenient!

Hi everyone! I'm Ksenia, Head of Tendem by Toloka 👋

To be honest Claude, Cowork and Claude code are my main job helpers nowadays. No kidding, most of work starts in Claude connected to office tools, our BI system and other products.

I bet many of you work the same way and sometimes you need an extra skill or opinion to enhance or simply verify the AI result. With Tendem MCP connector you can seamlessly delegate this job to a skilled expert and never wonder whether there is some AI slop left. AI taking care of the scoping and routing while a vetted expert does the more complicated and nuanced work. And it all comes back into the same conversation inside your favourite agentic interface.

What's the one task you'd never trust AI to finish on its own?