MCP endpoint
Axiomony serves stateless Streamable HTTP MCP at POST http://127.0.0.1:3000/mcp. Each request is authenticated as a real PostgreSQL user; the PostgreSQL server validates the username and password itself.
| Setting | Value |
|---|---|
| Transport | Streamable HTTP (stateless POST) |
| Local URL | http://127.0.0.1:3000/mcp |
| Authentication | HTTP Basic: PostgreSQL username and password |
| Required request header | Content-Type: application/json |
| Recommended Accept header | application/json, text/event-stream |
| Access control | PostgreSQL roles, grants, RLS, constraints and triggers |
Test the connection
Use curl --user USERNAME (without typing the password in the command) to have curl request the password interactively. Use a PostgreSQL user allowed to connect to the configured database.
curl --user platform_admin \
-H "Accept: application/json, text/event-stream" \
-H "Content-Type: application/json" \
-X POST http://127.0.0.1:3000/mcp \
-d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'The response lists the tools visible to that user. Seeing a tool does not imply unrestricted access to any table. PostgreSQL checks privileges again during data operations.
curl --user platform_admin \
-H "Accept: application/json, text/event-stream" \
-H "Content-Type: application/json" \
-X POST http://127.0.0.1:3000/mcp \
-d '{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"whoami","arguments":{}}}'Add Axiomony to an MCP-capable AI client
Select Streamable HTTP
Choose an MCP connection option that supports a remote/local HTTP endpoint with custom request headers. The exact UI depends on your client.
Enter your MCP endpoint
Use http://127.0.0.1:3000/mcp for clients running on the same host. Containerized or remotely hosted clients need a different reachable hostname and a correctly configured trusted transport.
Provide HTTP Basic credentials securely
Use your own PostgreSQL username and password through the client’s secure credential configuration. Axiomony does not accept a shared administrator password for everyone.
Discover your allowed tools
Use the client’s tool discovery (or the tools/list request above). Ask the AI to run list_tables or whoami as a first test.
ChatGPT and Claude cloud connection is not plug-and-play yet. Many hosted AI clients require an OAuth/OIDC authorization flow, not arbitrary Basic headers. The current version does not provide that flow. Do not put PostgreSQL passwords in chat prompts, generated JavaScript or published files.
Administrate using natural language
Administrators (members of mcp_admin or PostgreSQL superusers) can propose changes using manage_database. The external AI does not need to know PostgreSQL syntax; the internal LLM prepares SQL from a natural-language instruction. All changes require deliberate confirmation.
{
"instruction": "Only the assigned approver may approve an invoice below EUR 1000 after validation is complete."
}manage_database only creates a temporary plan held in server memory. Use its returned plan_id with review_plan, then explicitly approve using implement_plan with confirm: true. Plans expire after 20 minutes by default. The executor still needs the real PostgreSQL DDL privileges.
How user-created applications use data
For task-specific pages and reports, the preferred data path is trusted AI/MCP runtime → Data MCP → PostgreSQL. Generated static browser code must not store or forward reusable PostgreSQL credentials. The platform does not yet include an authenticated browser-to-MCP bridge, so publishing an HTML file alone does not make it a secure database app.
Browse all MCP tools or read the current security boundaries.