Database connectors
PostgreSQL, MySQL, SQLite. Schema introspection, foreign-key graph, sample rows. Read-only by default — Calame opens connections with SET TRANSACTION READ ONLY.
MCP / Database / Self-hosted
Calame is the control plane between your data and your AI clients. Connect Postgres, MySQL or SQLite, sculpt access profiles with PII masking and row-level scoping, then serve them to Claude, Cursor, VS Code — or your own LLM.
$ docker compose up
↳ Calame is running on http://localhost:4567What's inside
Calame is one binary that ships the schema introspector, the profile editor, the auth layer, the masking engine, and the MCP server runtime. No glue code.
PostgreSQL, MySQL, SQLite. Schema introspection, foreign-key graph, sample rows. Read-only by default — Calame opens connections with SET TRANSACTION READ ONLY.
Per-table & per-column visibility. Row-level scoping (e.g. restrict to client_email = X). One database, many published views.
Auto-detection of emails, phones, addresses, card numbers. Per-column or global rules: full redaction, hash, partial mask, or custom regex.
Open, Bearer token, password, OIDC SSO, OAuth 2.1 with Dynamic Client Registration, or external validation URL. Per-user revocation.
Query in natural language without leaving the UI. Anthropic, OpenRouter, or any OpenAI-compatible endpoint — including local Ollama, vLLM, LM Studio.
Every query, every actor, every scope — recorded and exportable. Write queue with approval workflow if you ever opt into mutating queries.
How it works
Plug Calame to a database, point it at a folder of PDFs and Markdown, and serve everything to your LLM through a single MCP endpoint. Same auth, same masking, same audit log — wherever the data lives.
Databases — Calame introspects schema, columns, foreign keys. Files — point at a folder, Calame parses PDFs, Markdown, DOCX, HTML, CSV and embeds them.
Pick tables and folders, mask PII, draw row-level scopes, choose an auth mode. Save a profile — it becomes a published MCP endpoint.
Drop the URL into Claude Desktop, Cursor, VS Code. They auto-discover both query and retrieval tools via OAuth 2.1 DCR.
Plays well with
Calame speaks the open Model Context Protocol over HTTP — so any compliant client works out of the box. Drop a URL, optionally an Authorization header, you're done.
New MCP clients ship every month — Calame stays compatible because it implements the spec, not a vendor SDK.
{
"mcpServers": {
"calame": {
"transport": {
"type": "http",
"url": "https://calame.acme.io/mcp/sales",
"headers": {
"Authorization": "Bearer ••••••"
}
}
}
}
}
Open at the core
Calame is dual-licensed. The core — connectors, profiles, MCP runtime, PII masking, chat, audit — is Apache 2.0. The ee/ modules (SSO + RAG) are BUSL 1.1: Pro unlocks them on one instance, Enterprise on as many as you need.
Self-host without limits.
Unlock SSO + RAG, scale to many instances, or get hosted.
From git clone to first MCP query
One Docker command. One profile. One MCP URL. Your database stays where it is — Calame just decides what your LLM is allowed to see.