Your own Claude can do the work in this workspace: read a company, write the pitch, prep the call, map who you know, say what to do next, record what happened.
Do you need a token?
| What you are connecting | How | Token? |
|---|---|---|
| claude.ai | Add a custom connector | No |
| ChatGPT | Add a custom connector | No |
| Claude Code | claude mcp add | Yes |
| Claude Desktop | mcp-remote in the config file | Yes |
| Cursor | ~/.cursor/mcp.json | Yes |
claude.ai and ChatGPT connect by their own connector screen. They mint the token themselves and put it on your Claude tile; there is nothing to copy and nothing to paste. Start at From claude.ai or ChatGPT below.
Everything else pastes a token:
- Workspace settings → Connect your apps → Your own agent → New token. Label it "Claude", tick what it may do, press Mint.
- Press Copy setup for your agent.
- Paste it into Claude Code, Cursor or Claude Desktop. It asks which tool you use and wires it up.
The paste carries the server address, your token and the config for each client, so the agent does the rest.
What it can do
| Tool | What it does | Costs |
|---|---|---|
scan_company | Reads a company from its website. A reading that failed says why, and is asked again an hour on. | A scan from the token's daily budget. |
build_pitch | Researches a company and writes the pitch in your voice, deliverables included. Several minutes. | A run. |
prep_for_call | A one-page sheet for a call at a named stage. About a minute. | A run. |
find_warm_paths | Who in your imported network could open the door on a finished pitch. | A run. |
push_to_crm | Sends a finished pitch to the CRM your workspace has connected: the company, its people, a deal and a note that cites the sources. Pushing it again updates. With dryRun: true it writes nothing and answers every write the push would make, with the exact payload, so Claude can show you what lands before it lands. | A run. |
next_step | Paste their reply or your meeting notes. Back comes where the deal stands, the one next step, the message it needs in your voice, and a checklist. About a minute. | A run. |
tick_step | Ticks a line on that checklist. | Nothing, ever. |
get_pitch | One saved pitch, by id or by company. Says of each deliverable whether it was written, failed (and why) or was never asked for. | Nothing. |
list_library | The index of what is saved: company, score, verdict. Never the pitch itself. | Nothing. |
job_status | Where a running job has got to. Done means the whole result, sources included, and, per deliverable, written, failed with the reason, pending or not requested. | Nothing. |
record_outcome | What happened with a company. Every outcome sharpens the next pitch. | Nothing, ever. |
whoami | What the token holds and how much day is left. | Nothing, ever. |
Long jobs answer with a job id. Your assistant polls job_status until the status is done. A pitch takes several minutes, so it will wait between polls.
next_step reads what is on the company's outcome record (the same notes you see under Outcomes) with what you just pasted on top. Ask it again with only the company a minute later and it answers with the step. Paste what changed to plan afresh; the newest material decides.
push_to_crm needs a CRM connected first; the workspace's owner does that once under Workspace settings → Connect your apps. With none connected the tool says so rather than guessing one.
Every answer carries its sources. A pitch says which pages it read, who it decided the company is, and what compliance found. Nothing is held back that the studio itself would show you.
Which scopes to tick
The paste works with whatever the token holds. For the whole set:
| Tick | To get |
|---|---|
scan:run | scan_company |
work:run | build_pitch, prep_for_call, find_warm_paths, push_to_crm, next_step, job_status |
library:read | get_pitch, list_library |
context:write | record_outcome, tick_step |
A token for a colleague's Claude that should read but never spend gets library:read alone.
Wiring it by hand
The server is https://studio.prospektor.ai/api/mcp, streamable HTTP. Send the token as a bearer header.
Claude Code:
`` claude mcp add --transport http prospektor https://studio.prospektor.ai/api/mcp \ --header "Authorization: Bearer psk_…" ``
Cursor, in ~/.cursor/mcp.json:
`` {"mcpServers":{"prospektor":{"url":"https://studio.prospektor.ai/api/mcp","headers":{"Authorization":"Bearer psk_…"}}}} ``
A client that cannot send a header, Claude Desktop among them, goes through mcp-remote:
`` {"mcpServers":{"prospektor":{"command":"npx","args":["-y","mcp-remote","https://studio.prospektor.ai/api/mcp","--header","Authorization: Bearer psk_…"]}}} ``
From claude.ai or ChatGPT
Their connector screens have no box for a token, and need none.
claude.ai:
- Settings, then Connectors in the left sidebar, under Customize. On a Team or Enterprise plan an owner adds it once under Organization settings → Connectors, and everyone else presses Connect afterwards.
- Press Add at the top right of the Connectors panel and pick the custom-connector option. Prospektor is not in the directory below; that button is the way in.
- Paste
https://studio.prospektor.ai/api/mcpand press Add. Leave Advanced settings alone; there is no client id or secret to fill in. - Prospektor opens. Signed out, it asks you to sign in and comes back here.
- Press Connect.
- In a chat, press +, then Connectors, and switch Prospektor on.
ChatGPT: Settings → Apps → Advanced settings, switch on Developer mode, then Add custom connector and the same address. On Plus and Pro, ChatGPT's custom connectors are read-only: it will scan and read your library and refuse to spend a run. Business, Enterprise and Edu get the whole set.
The screen it opens
One screen, one button. It is headed Connect Claude, says It will work in your workspace as you, and shows five boxes with four ticked: read a company, start work that spends a run, read a saved pitch, record an outcome. Take a copy of this workspace out is left unticked; tick it only if you want the app to be able to export. Then Connect, or Not now.
Only the workspace's owner can connect an app; a member gets one sentence naming the owner to ask. A paused workspace says so and connects nothing.
Check it worked
Ask your assistant what is this workspace wired to. It answers with your workspace's name, the token's label and how much of the day is left; the name is the proof it is pointed at the right workspace.
The connection is a token row under Workspace settings → Connect your apps → Claude, labelled with the app's name; the row says when it was last used and how many calls it has made. Revoke stops the app at once. It renews itself every thirty days for as long as it is used.
The same verbs over plain HTTP
Every tool is also a route on /api/v1, for a script or a CRM:
| Route | Tool |
|---|---|
POST /api/v1/pitch · GET /api/v1/pitch?id=… | build_pitch · job_status |
POST /api/v1/prep · GET /api/v1/prep?id=… | prep_for_call · job_status |
POST /api/v1/paths | find_warm_paths |
POST /api/v1/crm | push_to_crm |
POST /api/v1/next · GET /api/v1/next?company=… · PATCH /api/v1/next | next_step · tick_step |
GET /api/v1/library?id=… or ?company=… | get_pitch |
GET /api/v1/library[?q=…] | list_library |
Scans, outcomes, whoami and exports are on the same surface. See API tokens under the workspace help for the day cap and how to revoke.
What it costs
A run from your Claude is a run. A free workspace has its one full pitch, and your assistant meets the same wall you would. Runs past that need a subscription. The token's day cap is 1,000 calls; whoami, record_outcome and tick_step never count.
When something goes wrong
- "This token does not hold work:run": mint one that does. Tokens never gain a scope after minting.
- "Free workspaces run one full pitch, and this one has run it": the free run is spent. Subscribe to run the rest.
- "Nothing to map yet": import who you know first, under Your network.
- "Nothing on record for …":
next_stepplans from the company's outcome record. Paste what they said innote, or record an outcome first. - Still queued after ten minutes: a run hit its limit. Start it again.
- 401 with a date: the token expired. Mint another.
- The connector adds but no Prospektor screen appears: the address is wrong. It is
/api/mcp, not/mcpor/api/v1. - "Only the workspace's owner can connect an app": you are signed in as a member; the screen names who to ask.
- "This page had expired. Connect the app again.": the consent screen sat open too long. Add the connector again.
- ChatGPT lists the tools and refuses to run a pitch: the Plus and Pro read-only limit above, not a fault.
Revoke a token from Workspace settings → Connect your apps → Revoke. Everything holding it stops at once.