The best fix of the week is stuck in one person’s chat.
Your team learns useful things with AI every day: a caveat in a vendor’s data, the query that finally worked, the approach that wasted a morning. Those stay in individual chat histories, so the next colleague starts from zero. You can’t just share chats either, because they’re full of client names and working detail.
Crewknow keeps the learning, not the whole chat.
Drafted by your assistant. Signed by you. Found by your team.
-
save this: reconcile vendor segment tables to the annual-report note before counting subsidiaries
Saved to your Crewknow Drafts. Colleagues can’t see it until you sign it.
Title: Vendor segment tables mislabel some associates as subsidiaries
Scrubbed on arrival: 2 client names replaced with a placeholder.
Review and sign in your browser: app.harbourpine.example/drafts/d_4821
Capture.
Your assistant drafts it.
Crewknow doesn’t upload your full chat history. With Claude Code, your computer picks out the few passages of a session that look like a learning and sends only those short excerpts, as drafts, to your firm’s Crewknow instance. When you tell your assistant “save this”, the assistant writes the draft. Either way, your firm’s instance scrubs each draft’s title and text when it arrives, and you see it before it’s shared with colleagues. It’s saved to your Drafts, and colleagues can’t see it until you sign it.
Automatic drafting at session end works with Claude Code today. In Cursor, say “save this”.
-
Vendor segment tables mislabel some associates as subsidiaries
On [Client A] diligence, the segment table counted two associates as subsidiaries. Reconcile it to the annual-report note before counting.
2 client names replaced. Scrubbing only catches what it recognises, so read it before you sign.
Who can see this once you sign it?
- Only me Kept in My entries
- Project Kestrel team Inside the engagement’s wall
- Whole firm
Sign as Wei Chen (unavailable until you choose who can see it)EditDiscard
Nothing is preselected. Choose who can see it, then sign.
Sign.
You check it and sign.
Your drafts wait for you in Drafts in the web app. Before you review a draft, Crewknow replaces client names and codenames from your firm’s list, plus emails, international phone numbers, API keys and similar secrets, in the draft’s title and text. It’s best-effort and misses things it doesn’t recognise, including some ID-number and phone formats, which is why your draft isn’t shared with colleagues until you sign it. You check it, edit if needed, and sign it in the web app.
-
How many subsidiaries does the vendor have?
The segment table lists 14, but reconcile it to the annual-report note first: segment tables often count associates as subsidiaries.
Source: “Vendor segment tables mislabel some associates as subsidiaries”, signed by Wei Chen · version 1 · Active · re-confirm by 14 Jan 2027 · Open
Retrieve.
Your team finds it.
Before a colleague starts a similar task, their assistant searches signed entries through a single MCP connection. Each result shows who wrote it and its re-confirm date.
Illustrative examples. Harbour & Pine, its people and entries are fictional. Screens show the target design.
| Entry | Written by | Source |
|---|---|---|
| Check a company’s registry status date, not just its status | Ana Lim | Kickoff session |
| Vendor DD: ask for the group structure chart first | Wei Chen | Kickoff session |
| Market sizing: state the year of every source figure | Kai Ng | Kickoff session |
Illustrative example. Harbour & Pine Advisory, its people and its entries are fictional.
Included in every pilot
Day one isn’t empty.
Before your team’s first day, we run a 45-minute kickoff session with them. Together you pick the lessons people keep re-explaining, such as data caveats, fixes and dead ends, and turn about ten of them into entries. Each person saves their own as a draft, checks it after scrubbing, and signs it in the web app, so every kickoff entry has a named author from your team.
| Assistant | What works |
|---|---|
| Claude Code | Today: search, plus automatic drafts at session end |
| Cursor | Today: search, plus “save this” |
| claude.ai | Connector coming |
| ChatGPT | Connector coming |
MCP link https://mcp.harbourpine.example
Works with the assistants your team uses.
Claude Code and Cursor connect through MCP today. Connectors for claude.ai and ChatGPT are coming. Tell us which assistants your team uses on the form.
Your name on what you figured out.
Every entry carries its author’s name. It’s signed with the key of the person who signed it, or of whoever last edited it, and countersigned by the firm. Keys are held encrypted on your firm’s instance, and a person’s key is used when they sign or edit an entry in the web app. Your instance can check an entry’s signatures on request.
Walled by engagement, not by goodwill.
Your admin sets up walls per engagement or deal team, usually from the list you already keep for conflicts.
For the pilot, we commit in the pilot letter to setting up a separate Crewknow instance for your firm. Authors, curators and admins can edit or retract entries, only in the web app while signed in. Your assistant can’t. Each edit and retraction is recorded in a hash-chained audit log that shows accidental or partial changes. It doesn’t protect against someone with direct database access.
| Engagement | On it | Blocked |
|---|---|---|
| Project Kestrel | 6 people | 2 people |
| Project Juniper | 4 people | Wei Chen |
Scrubbing misses things it doesn’t recognise, which is why you review every draft before it’s shared. If you retract an entry, assistants that already received it keep their copy. If they fetch it again, they’ll see it’s retracted.
How your firm’s data is handled during a pilot.
Only drafts leave your computer, not your chat history. A draft isn’t shared with colleagues until its author signs it, and your firm’s instance scrubs its title and text when it arrives.
Walls, scope and sensitivity are checked by the server on every search. Sharing confidential entries firm-wide needs a curator.
A separate instance for your firm, as a pilot-letter commitment.
Assistants can search and send drafts. Editing and retracting need a signed-in person in the web app.
What we don’t have yet: SOC 2 or ISO 27001 certification, or a third-party penetration test.
Try it with one team for three weeks.
A free three-week pilot for one team of 5–10 people. In the pilot letter we commit to setting up a separate Crewknow instance for your firm. We connect your team’s assistants and walls, and check in with you each week. The pilot letter states the price to continue before you start. At the end of the pilot we export your firm’s data for you, then delete your firm’s instance and its database.
What we ask: a sponsoring partner, permission to use MCP connectors, and 15 minutes a week of feedback.
Questions partners ask first.
What leaves my device?
Crewknow doesn’t upload your full chat history. With Claude Code, your computer picks out the few passages of a session that look like a learning and sends only those short excerpts, as drafts, to your firm’s Crewknow instance. When you say “save this”, your assistant writes the draft. Either way, your firm’s instance scrubs each draft’s title and text when it arrives, and you see it before it’s shared with colleagues. Drafts are kept on your firm’s instance. Your firm’s admins can export a member’s data, including unsigned drafts. Each export is logged. Your draft isn’t shared with colleagues until you sign it.
Where does our data live?
In a Crewknow instance we commit in the pilot letter to set up only for your firm. Other firms’ data isn’t in it. Data leaves your instance to the AI assistants your team connects (search results go to them, under your firm’s agreement with that provider), your Slack or Teams digest channel, and any model provider your firm enables. Our hosting provider and any other sub-processor are listed in the security note. If your firm enables a hosted model, it’s named in your pilot letter. A drafting model sees conversation excerpts before scrubbing; a search model sees your search text.
How does scrubbing work, and can it miss things?
Before you review a draft, Crewknow replaces client names and codenames from your firm’s list, plus emails, international phone numbers, API keys and similar secrets, in the draft’s title and text. It’s best-effort and misses things it doesn’t recognise, including some ID-number and phone formats, which is why your draft isn’t shared with colleagues until you sign it. Edits are checked again before they’re signed.
Who can see an entry?
Only people allowed by its scope and sensitivity. Walls, scope and sensitivity are checked by the server on every search, through the web app and through assistants. Sharing confidential entries firm-wide needs a curator.
Can my assistant change or delete entries?
No. Assistant connections can search, send drafts and report whether an entry helped. Authors, curators and admins can edit or retract entries, only in the web app while signed in. Curators and admins can do this only for entries they’re allowed to read. Each edit and retraction is recorded in a hash-chained audit log that shows accidental or partial changes. It doesn’t protect against someone with direct database access.
What happens if I retract an entry? Can I delete data?
Retract an entry and it stops appearing in new searches for everyone, while its history is kept. Assistants that already received it keep their copy, and Crewknow can’t take it back. If they fetch it again, they’ll see it’s retracted. To have specific data removed during the pilot, ask your admin to contact us and we’ll purge its content; backup copies remain until backups expire. At the end of the pilot we delete your firm’s instance and its database.
How long do you keep our data? Can we get it back?
For the pilot, until you ask us to delete it. Each person can get an export of their entries, drafts, searches and usage, and at the end of the pilot we export your firm’s data for you.
Do you train AI models on our data?
No. Crewknow doesn’t train models. Whether a provider does is set by its terms with your firm. With the pilot’s default settings, search runs on your firm’s instance.
Do you have SOC 2 or ISO 27001?
Not yet. We haven’t had a third-party penetration test either.
Which assistants work with it?
Today, Claude Code and Cursor connect through MCP. Automatic drafting at session end works with Claude Code; in Cursor, say “save this”. Claude (claude.ai) and ChatGPT connectors are coming. Tell us what your team uses on the form.
Will people actually bother?
That’s what the pilot measures. Drafts are written for you, and signing is a short review in the web app. If your team doesn’t use it after three weeks, you’ll know. The pilot is at no cost to your firm.