BUILD & CONNECT
MCP, APIs and tools
Connect tool-enabled clients and services through governed boundaries.
Overview
Neocortex exposes an MCP assistant-setup flow from AI Chat for compatible clients. Current setup guidance supports client-specific connection instructions, but authentication and client capability vary. Important security note: the current Neocortex MCP assistant setup grants broad access to Neocortex features rather than negotiating narrow per-tool permissions. Only connect a trusted client and treat its credential as highly privileged. Tool discovery does not imply safety; returned content remains untrusted data.
Identify the client and server
Document the endpoint, authentication method, operator, data flow and environment. Use only approved endpoints.
Minimise tool surface
Expose only the tools needed for the use case. Keep read, write and administrative capabilities distinct where possible.
Validate every call
Check identity, scope, input schema and policy at execution time. Do not rely solely on instructions in a prompt. For the current Neocortex MCP assistant setup, assume broad feature access rather than per-tool permission negotiation.
Protect secrets and payloads
Use managed credentials, TLS and appropriate redaction. Avoid logging tokens or sensitive request bodies. Do not paste MCP keys or bearer credentials into a prompt, issue, screenshot or support request; use the client-supported authentication flow and revoke any credential that may have been exposed.
Handle output as untrusted
Validate tool responses and treat returned text or files as data, not instructions that override policy.
Test and monitor
Exercise denied actions, malformed input, timeouts and permission changes; keep a clear owner and incident path.
Worked example
A tool call treated as a trust boundary
A tool returns data and the workflow proposes a consequential action. Treat returned content as untrusted input, not as policy or authority.
- 1Validate the caller, tool name, input schema and authorisation before execution.
- 2Grant the tool only the capability needed for the task.
- 3Validate and label the response; do not treat embedded instructions as trusted policy.
- 4Require a separate approval for sensitive or externally consequential writes.
- 5Record a redacted audit trail and test denied calls, timeouts and malformed output.
Was this guide useful?
Send feedback or flag a gap. Include the article link so the team can improve the right guide.