The Crelate MCP server lets an AI assistant work with your Crelate data on behalf of one of your users. That assistant can be Claude, ChatGPT, Microsoft 365 Copilot, Gemini Enterprise, or a development tool like Cursor or VS Code. It is a connector rather than a second copy of your data. Every request it handles becomes an ordinary Crelate API call, made as that user.
The short version
- It acts as a specific person, never as your organization. A connected assistant sees exactly what that user sees when they sign in to Crelate, and nothing more. A recruiter's assistant cannot reach records the recruiter cannot reach.
- Each user opts in individually. There is no switch that connects your whole team at once. Every user connects their own assistant and approves it themselves.
- It is overwhelmingly read-only. Of 122 available tools, 5 can change anything at all. None can delete.
- Access can be withdrawn at any time, by the user or by Crelate support, and takes effect within about a minute.
What an assistant can do
The tool list is generated directly from Crelate's public API, so it cannot quietly drift away from what that API already allows. The counts on this page are read from the running server rather than written by hand.
Reading
Searching and retrieving contacts, jobs, companies, activities, notes, placements, applications, tags, users, and the picklists and workflow settings that give those records meaning. This is the bulk of what an assistant does: answering questions about your data.
Writing
5 of the 122 tools can change anything:
- Change Business Development workflow stage
- Change Employment Search workflow stage
- Change Recruiting workflow stage
- Change Sales workflow stage
- Create a Note activity
None of them deletes a record or empties a field.
No tool sends email on a user's behalf, and none changes your Crelate configuration.
Who controls access
Three separate approvals stand between an AI product and your data. All three have to happen, and any one of them can be withheld.
| Gate | Who decides | What it controls |
|---|---|---|
| Which AI products may ask at all | Crelate | Only AI platforms on an allowlist Crelate maintains can begin a connection. An arbitrary application cannot point itself at this server and start prompting your users to sign in. |
| Which assistant a user connects | The individual user | The user signs in with their normal Crelate account and sees a screen naming the exact product asking for access before anything is granted. Approving one assistant does not approve the next one. |
| What that assistant may do | The individual user | Permission is granted in named slices (read one record, read in bulk, change one record, change in bulk, delete) and the user chooses which to grant. An assistant approved only to read cannot write, whatever it asks for later. |
Underneath all three, the user's own Crelate permissions still apply to every individual request. Granting an assistant broad permission never widens what the person behind it could already do.
Microsoft 365 Copilot and Gemini Enterprise
These two platforms cannot identify themselves the way Claude and ChatGPT do, so they work differently. An administrator creates one connector for the organization from the AI Platform Connectors settings page, then hands the credentials to whoever builds the Copilot or Gemini agent. That is the one place an administrator, rather than an end user, holds the key. Each person using it still signs in as themselves, and is still bound by their own permissions.
Turning access off
- A user can disconnect any assistant from their own Crelate settings.
- You can have Crelate support revoke a specific user's access, during offboarding for instance.
- Removing or disabling the Crelate user account stops everything that account's assistants could do, because every request is made as that user.
Revocation is close to immediate, though not instantaneous. A revoked connection can keep working for up to roughly a minute before the change reaches every part of the network. Plan around that minute for a time-critical offboarding.
How your data is protected
| Concern | How it is handled |
|---|---|
| In transit | Every connection is HTTPS, in both directions. Unencrypted connections are not accepted at all. |
| At rest | Credentials are stored encrypted (AES-256-GCM) and decrypted only for the moment a request needs them. The encryption key lives in a managed secrets store, never in configuration or source code. |
| Your records | This server keeps no copy of your Crelate data. It passes each request through to the Crelate API and returns the answer. Nothing is retained afterward. |
| Sign-in | Standard OAuth 2.1 with PKCE on every attempt, so an intercepted sign-in cannot be completed by anyone else. |
| Logging | Operational logs record which tool ran, whether it succeeded, how long it took, and internal user and organization identifiers. They never contain a credential, and never contain your candidate, client, or job data. |
| Abuse control | Requests are rate limited per user and per tool (120 calls a minute), so a misbehaving assistant cannot hammer your data or your API quota. |
| Administration | The endpoints that create and revoke access are reachable only by Crelate's own systems, behind two independent checks. They are not exposed to the public internet. |
Worth knowing before you roll this out
- An assistant is only as careful as the person driving it. A user who can change a candidate's workflow stage in Crelate can ask an assistant to do it, and an assistant may act on an ambiguous instruction. The write surface is deliberately small for exactly this reason, though it is not zero.
- Your data leaves Crelate for the AI provider you connect. Records an assistant retrieves are sent to that provider, whether Anthropic, OpenAI, Microsoft, or Google, and are handled under their terms rather than Crelate's. If your organization has rules about which AI vendors may process candidate data, that decision belongs at the point where you choose which assistants to permit.
- Assistants can be confidently wrong. Treat what one reports the way you would treat a new colleague's summary: useful, and worth checking before it drives a decision that matters.
- Usage is metered. Tool calls are counted per organization.
Connecting
The server address is https://mcp.crelate.com/mcp. A user adds that as a remote MCP connector in
their assistant, signs in with their Crelate account, and approves the access described
above. There is no key to copy, and no configuration file to distribute.
For anything this page does not answer, such as a security questionnaire, a data processing agreement, or a rollout plan for a large team, talk to your Crelate account manager or email support@crelate.com. If a walkthrough would be more useful than a page, book a demo.