Your assistant knows the task. Now it can know the people.
Throughline connects your team’s consented profiles to Claude, ChatGPT, Cursor and any MCP-compatible client. Ask "how should I open this one-to-one?" where you already write, and the answer comes from the real profiles of the people involved.
- Claude & ChatGPT connectors
- Open MCP server
- OAuth, scoped per workspace
- Every answer cites its sources
Watch it fetch the real profile.
An assistant without Throughline will answer this question. It will sound confident and it will be generic, because it is reasoning about a name, not a person.
With Throughline connected, the assistant stops and fetches a real profile first. You can see it happen — the call, the profile it resolved, and the citation on the answer. If the person hasn’t opted in, there is no profile to fetch, and it says so instead of improvising.
throughline · get_profile
{ person: "maya.okonkwo" }
resolved · Steadiness (S) · shared with you
Send the change before the meeting rather than opening with it. Maya reads a surprise as the team having moved without her.
Evidence
- Steadiness (S) · warning before a change lands
- Context flex · agrees in the room, raises it after
Source
Dana
Profile shared with you
One intent. Two colleagues. Two different messages.
This is the test we hold ourselves to: swap one colleague’s name into the other’s guidance and it should visibly stop making sense. Advice that survives the swap wasn’t personal enough to ship.
To Dana
Dominance (D) · wants the headline first
“Deadline's moved to the 14th. I've cut the analytics work to make it fit — that's the trade I'd make, tell me if you'd cut something else instead. Everything else holds.”
To Maya
Steadiness (S) · warning before a change lands
“I want to flag something before it lands on you as a surprise. The deadline has moved to the 14th. Here's what I've already taken off your plate to protect it, and here's what I'd like your read on before I commit to anything else.”
Connect it once. It’s everywhere you already work.
Throughline speaks the Model Context Protocol, so anything that speaks MCP can use it. These three are the ones we document and test on every release.
Claude
Desktop, web and Claude Code. Remote MCP connector with OAuth.
ChatGPT
Connectors in Plus, Pro and Team workspaces.
Cursor
MCP server configured per workspace, for people who live in the editor.
See how you work in six minutes
Free, no sign-up. Then bring your team in, free for five people.
Five things your assistant gains.
Deliberately few. Each one answers a question a manager actually asks, and none of them will answer a question we’ve decided shouldn’t be asked.
get_profileRead one colleague's working style
Big Five scores, working-style scores and Enneagram type, plus how the person takes feedback, what drains them, and where their behaviour shifts with context.
get_motivationRead what someone works for
The shape of reward they respond to — autonomy, recognition, mastery, security or progression — what energises and drains them, and how they want to be recognised. For designing an offer, an onboarding plan or a stretch assignment. It says what a person values, never what they are worth.
get_pairRead the relationship between two
The layer conventional reports skip entirely: where these two specific people will misread each other, what each needs from the other, and which friction is structural, not personal.
get_teamRead the shape of the whole team
Composition, the styles that are over- and under-represented, and the failure modes a team with this mix tends toward. Useful before a reorganisation, an offsite or a hire.
get_playbookPull the situation's playbook
Critical feedback, delegation, running a disagreement, onboarding, breaking bad news. The playbook supplies the structure; the profiles supply the specifics.
list_teamSee who has opted in
Who has a profile, who is mid-assessment, and who has chosen not to take part. The assistant needs this to know when to stop rather than guess.
And what there is no tool for
There is no endpoint that ranks colleagues, scores potential, supports a hiring or promotion decision, or builds a profile of someone who hasn’t consented. Not gated — absent.
The first MCP server that knows a team, not just a person.
Looking one person up over MCP isn’t new — a couple of tools do it, and we’d rather say so than have you find out. What no one else has is the layer underneath: consented profiles for a whole team, and measured guidance for any two of them together.
| Personality database with a connector | Throughline | |
|---|---|---|
| Where the profile comes from | Predicted from public data — LinkedIn, published writing, inferred signals. The person may never know a profile of them exists | An assessment the person chose to take. No consent, no profile, no exceptions |
| Who it's about | Mostly people outside your company — prospects, candidates, contacts. Built for selling to someone | The colleagues you work with every day. Built for working with someone |
| Whether it's mutual | One-way. You look them up; they learn nothing about you — an asymmetry the other person didn't agree to | Mutual by construction. You see them only if they can see you |
| What the model receives | A type label and generic tips for that label. Four people with the same code get one answer | Continuous scores, intensity and context-flex. Same style, different person, different answer |
| The relationship layer | Absent — you look up two people and compare by eye. The synthesis is left to you | Measured separately, with its own guidance. The pair is the unit, not the person |
| What it refuses | Usually nothing. Ask it to rank your team and it will try | Ranking, hiring, promotion, anything unmeasured. Declines and explains why |
The short version — Everyone else connected an existing prospecting database to an assistant. We built for the assistant from the start, which is why the pair layer exists at all. Nobody adds consent and mutuality to a database of people who never agreed to be in it.
About four minutes, once, per person.
Your workspace admin turns it on
One toggle in the Throughline console issues the workspace's OAuth client. Nothing is exposed until a colleague connects their own account to it.
Each person connects their assistant
Paste the server URL into Claude or ChatGPT and sign in. The token is theirs, scoped to what they're allowed to see, and revocable by either side.
Ask it something difficult
No new habit to build and no app to live in. The next time you draft a hard message in the assistant, ask who you're writing to.
What people ask about the integration.
Mostly about privacy, which is the right instinct. The full answer lives on the security and privacy page.
No. Profiles are sent to your assistant as context at the moment you ask a question, in the same way any MCP tool returns data. Anthropic and OpenAI both offer commercial terms that exclude API and connector traffic from training, and those are the terms we operate under.
We do not train anything of our own on your team’s data either — not a shared model, not an aggregate benchmark, not an "anonymised" corpus.
Bring the hardest working relationship you manage.
Twenty minutes, live, in a real assistant. If the answer isn't visibly better than what you'd get by asking without us, don't buy it.
See pricingTell us what you’re working on.