Kitsuno speaks MCP: connect Claude, ChatGPT, or any AI assistant to your own career record

Since this week, both of our platforms speak MCP. We started with Yako, our B2B tender-intelligence platform: an organisation’s own AI assistant can read the opportunity terrain Yako watches for it. One week later, Kitsuno followed with the bigger half of the idea: your assistant — Claude, ChatGPT, any client that speaks the protocol — can now connect to your own career record.

One URL, one consent screen, and the AI you already think with becomes a colleague that actually knows your evidence.

This article is the full tour: what MCP is, everything a connected assistant can do today, the prompts worth trying, what it costs, and the consent machinery underneath — including the wall our founder hit an hour after launch, because you deserve to know how this is really built.

What MCP is, in one paragraph

MCP — the Model Context Protocol — is an open standard that lets AI assistants connect to tools and data their makers never met. A platform runs one MCP server; any compliant assistant can discover it, authenticate through standard OAuth, and use its tools. No per-chatbot plugins, no captive widget, no copy-pasting your CV into a chat window ever again. You bring your own model. We bring your record and the switches.

To our knowledge, Kitsuno is the first consent-first career platform to ship this natively. We would honestly be glad to be wrong — the more records that belong to their people, the better.

What your assistant can do now

Seventeen tools. Here they are as things you can actually say.

Know your record

“What evidence do I have for stakeholder management?” — “Show me my testimonials from 2024.” — “Which of my projects mention learning design?”

Your assistant searches everything you flagged — Library evidence, experience, education, skills — and opens single items in full. It sees exactly what you shared and nothing else; more on the switches below.

Take everything with you

“Export my full Kitsuno record.”

One tool returns your complete record as a single document, and it deliberately ignores every share switch. Portability is your right under GDPR Article 20, and a right that costs money or requires the correct toggles is not a right. Free, for everyone, forever. This is also the shortest possible summary of how we think about your data.

Write from evidence, not imagination

“Generate a CV targeted at this Head of Learning role.” — “Draft a motivation letter that leads with the platform work.” — “Make the summary shorter and render me the PDF.”

Two lanes, and the first one is unlimited. Your assistant can read the same Writer brief our own engine uses — your strengths, your gaps, what evidence is thin — pull your shared items, and draft with its own model: ten versions of the cover letter, five angles on the summary, as many as the conversation needs. No quota, no cost on our side, ever. You already pay for your model; we are not going to meter your own intelligence against your own record.

The second lane is our Make engine, run on demand: targeted CV, motivation letter, speaker bio, project case study, reference request — grounded strictly in your verified record, inventing nothing, using your normal draft quota exactly as in the app. Either way, the assistant can refine in one-line corrections, render CVs to PDF and read the result back to catch what a score cannot (the three-page sprawl, the claim that drifted), and save its work into your Library, visibly marked, undoable.

Be ready for Thursday

“Brief me for the Deloitte interview.” — “Rehearse me. Ask the hard questions about the gap year.” — “Who do I know at this company?”

The genuinely new part. Your assistant fetches the same deterministic briefing prompt our Prepare room uses, generates your interview briefing with its own model from your shared evidence and the real role, and saves it back through our validation. Then it rehearses you from it — not from a generic question list, from your actual application. We deliberately kept our in-app rehearsal chat out of the protocol: your assistant is already a capable interviewer, and one model per conversation is the honest architecture.

Meet, the relationship web, answers what surrounds the interview: upcoming rounds with readiness at a glance, the dossier on a company including your own pipeline history there, who you actually know.

Read the pipeline honestly

“What’s the state of my search?” — “What arrived this week, and what have I let sit?”

Through the record and Meet tools your assistant sees what our app sees: what came in, what you drafted and never sent, what has been quiet too long. An assistant that reads your real record keeps everyone honest, including us.

Ask how anything works

“Ask Kitsuno why my applications feel generic.” — “How do I link evidence to a role?”

The assistant searches our entire Help book — the same canonical answers the in-app Help serves, in your language, read live so they are never stale. And then, unlike a help page, it can look at your actual brief and show you your version of the answer.

The switches underneath

Everything above obeys one model, and it starts at zero.

Off by default, item by item. Nothing is shared until you flag it. The “Share with AI assistant” view in your Library sweeps through everything — experience, education, skills, every evidence type — with per-item checkboxes, section toggles, and a live counter. Every item editor carries the same switch. Flag three projects and your references, keep the rest dark: that is the whole of what any assistant sees.

The person grants, not the client. The consent screen shows every permission with an honest plan label, regardless of what the connecting assistant happened to request. What a client asks for is a hint about ordering; your choice is the only authority. The OAuth standard has always permitted this reading — the authorization server issues scopes according to the resource owner — and it is the only reading of consent we find acceptable.

Revocation is total. One click under Account → Connections kills a connection completely — access token and the refresh tokens clients use to quietly stay signed in. Every access any assistant ever made is in a log you can read.

The export is the floor. The one tool no switch touches, because portability is a right. Everything else is a choice.

What it costs, plainly

Connecting: free. Record tools, export, Writer brief, Help book: every plan. Your assistant drafting with its own model from your brief and evidence: unlimited, no quota, no cost on our side — that lane is the point of the whole design. Running our Make engine: your normal draft quota. Prepare: Scout or Pro. Meet: Pro. Identical to the app, enforced in exactly one place — a free user’s assistant gets a clear, honest message naming the plan, not a silent failure. And grants are durable: tick the premium permissions today, and they light up the moment your plan covers them, no reconnection.

The spine, and why it gets stronger

Here is the quiet consequence of all this, and honestly the reason we are more excited about MCP than about any single tool: the Library just became the most valuable thing we build.

Models are getting better everywhere at once, and increasingly interchangeable. What no model can conjure — however capable — is a verified record of your working life: the projects with the reference documents attached, the skills linked to the evidence that proves them, the testimonials, the thesis, the case studies, structured and consented, item by item. And on the other side, no chat window can see the job market: Kitsuno’s scanners work through hundreds of boards and sources daily, deduplicate, geo-filter, extract, and score, turning the chaotic job web into clean rows a model can actually reason over.

Those two things — your evidence and the living terrain — are the spine. Any good LLM connected to it becomes immediately better at career work than the same LLM without it, because grounding beats brilliance every time someone asks “what should I say about the gap year” or “is this role actually a fit.” Every item you add to your Library now strengthens every assistant you will ever connect, present and future. We do the crawling, the cleaning, the structure, and the consent machinery; your model of choice does the thinking. That division of labour is the bet, and it means we get more useful as models improve, not less.

Why open, and why now

The assistant you trust was never going to be the one we built for you. You have a model you pay for, a tone tuned over months, a context that already knows how you think. Forcing you into our chat window to use your own career data would be capture dressed as convenience.

So both platforms took the other path. Yako first: the organisation’s terrain, readable by the organisation’s agent. Kitsuno one week later: the person’s record, readable by the person’s agent. One protocol, both directions of the same handshake, and nobody’s model in the middle except the one the user chose.

The agent web is arriving either way. The platforms worth trusting in it will be the ones whose consent survives contact with someone else’s agent.

Try it now: claude.ai → Settings → Connectors → Add custom connector → https://app.kitsuno.ai/mcp → approve what you choose. Then open your Library, tap Share with AI assistant, and start from nothing.

The longer story — twenty years of it, from a 2005 essay on constructivist pedagogy to this week’s launch — is on our founder’s Substack.

Published 23 August 2026. No affiliate links. No sponsored placements.