Can AI agents or LLM tools access my documents securely?
The current FormKiQ MCP Server is a Python application you install and run on your workstation or a server you manage. It connects an MCP-compatible AI client to an existing FormKiQ API using a JWT. Tools are read-only by default; writes and purge require explicit opt-in. Caller permissions and the AI client's handling of retrieved content still matter.
Where the MCP server runs
The code is published in formkiq-mcp-server. Release v0.2.0 requires Python 3.11+ and can run through pipx or from a checkout using uv and server.py.
- On a workstation: configure your AI client to launch the Python MCP process and communicate with it over standard input/output (
stdio). The process calls your deployed FormKiQ JWT API over HTTPS. - On a server you manage: run the Python application there. It supports Streamable HTTP; configure caller authentication, network access, and process management before exposing it to remote clients.
This MCP process is separate from the FormKiQ backend deployed in AWS. Installing the backend does not automatically install the Python application on a workstation or provision a hosted MCP endpoint. Supply the existing API endpoint and a current JWT when configuring the process. See the installation guide for the launch configuration.
Separate discovery from authorization
The MCP server exposes operation discovery and focused request schemas. Its API groups and module settings filter which tools are offered; they do not install backend features or grant permissions. The backend still evaluates the JWT and site access for each call.
Start with the pinned release and the smallest useful read-only API group. Supply a current Cognito AccessToken through protected local configuration. The documented server release requires manual token refresh and restart; it does not gain IAM signing by changing the endpoint URL.
Review the complete data path
A local MCP process can return documents to an AI client that uses an external model. Review that client's processing and retention settings separately from FormKiQ's AWS deployment. Treat document text and tool results as untrusted evidence, not instructions authorizing further tool calls.
Test allowed and denied documents with an ordinary user. Before enabling mutations, check the client's approval controls and the separate write/purge flags. If deploying a shared remote MCP service, authenticate its callers as well as protecting its backend token.
Edition and configuration
The public MCP server and backend have independent versions. Installed editions/modules and caller permissions determine which discovered operations actually work.