Midplane

Safe-by-default SQL guardrails for AI agents. Midplane sits in the query path between an AI agent (Claude, Cursor, any MCP client) and your Postgres database. It parses every statement with a real SQL AST — not a regex blocklist — enforces a declarative per-table access policy, blocks destructive DML/DDL, and writes an event-sourced audit log of which agent ran what, before the query executes.

📖 Full documentation lives at midplane.ai/docs — agent setup, the policy reference, self-hosting, deployment, and the threat model. This README is just the orientation.

Why this exists

AI coding agents are being plugged into production Postgres without an audit trail or a safety layer. The deprecated Anthropic reference Postgres MCP shipped a stacked-statement injection vector (Datadog Security Labs, 2025); the common service-role setup hands an agent a connection that can read and write every table. Midplane parses every query as an AST, denies the dangerous shapes, and writes a durable audit row before the query reaches your database.

What it blocks

• Destructive writes by default — DELETE FROM users is denied even with a WHERE, until you opt the table into read_write.
• Whole-table wipes and schema destruction — no-WHERE DELETE / UPDATE and all DROP / TRUNCATE / ALTER, regardless of table policy.
• Stacked-statement injection — SELECT 1; DROP TABLE users denied at parse time.
• Writes hidden inside a read — WITH x AS (DELETE FROM users RETURNING *) SELECT * FROM x is denied at the inner DELETE, not the outer SELECT. The same recursive walk covers subqueries, UNION arms, and JOINs.

The full policy model, the parse → policy → audit pipeline, and the adversarial corpus that pins it are documented at midplane.ai/docs.

In a real chat

A delete all users prompt to Claude Code, against a Midplane-fronted DB:

⏺ Hold on — that's destructive and irreversible. Before I run DELETE FROM public.users, please confirm.

user: confirm delete all users

⏺ Midplane blocked it: "writes to table users are not allowed by the table-access policy." The DELETE was audited but not executed.

Quick start

Three ways in — same open-core codebase behind all of them. Step-by-step guides are at midplane.ai/docs.

Managed cloud

The fastest way to try Midplane: sign up at app.midplane.ai and go from zero to your first guarded query in a couple of minutes. Dashboard, policy editor, hosted audit log, agent-token issuance. Nothing to install, multi-region, fully supported.

Guard one database yourself

Put the MIT engine in front of a Postgres database and point an agent at it. Nothing to install — npx ships with Node and fetches the midplane package on first run (needs Node 22.16+; on anything older it says so and exits). Add this to your MCP client's config (Claude Code, Claude Desktop, Cursor — they all take this shape):

{
  "mcpServers": {
    "midplane": {
      "command": "npx",
      "args": ["-y", "midplane", "server", "--stdio"],
      "env": { "DATABASE_URL": "postgres://user:pass@host:5432/db" }
    }
  }
}

Keep the connection string in that env block rather than on a command line, where it would leak to ps aux and your shell history. The block still lands in a plaintext config file, so give Midplane its own least-privilege Postgres role: it governs which SQL runs, not what the role underneath it can reach.

That config is already the safe default: reads allowed, writes and DDL denied, every query audited to ~/.midplane/audit.db. Read the log back with npx midplane audit denies. To open specific tables up, generate a policy with npx midplane init — it introspects your schema over a read-only connection, suggests a tenant column, and writes a validated midplane.policy.yaml.

For a CI pipeline or a long-lived sidecar, the same engine ships as a self-contained image with no Node in it — midplane/midplane:0.20.0, serving Streamable HTTP instead of stdio. Setup · engine/README.md.

Self-host the whole app

The complete single-tenant product — dashboard, policy editor, audit log, agent-token issuance — keyless and uncapped, on your own Postgres. Docker is the only prerequisite:

git clone https://github.com/midplaneai/midplane && cd midplane
./bin/self-host up                               # → http://localhost:3000

That generates secrets into .env.self-host, brings up Postgres + the web app, applies migrations on boot, and prints the dashboard URL — the first email+password signup becomes the owner.

Running from source, the single-image deploy, the engine-spawn topology, and the full walkthrough: midplane.ai/docs (in-repo: SELF_HOST.md).

Open core

Midplane is open core, MIT, and self-hostable. Everything outside apps/web/src/ee/ is the Community Edition — the whole single-tenant product, uncapped when self-hosted. apps/web/src/ee/ is the commercial Enterprise Edition (SSO/SAML today; the governance band over time); deleting it leaves a working MIT build. The managed cloud is the same codebase and the supported, paid path. See LICENSE for the MIT terms and NOTICE for the ee/ carve-out.

Architecture

One codebase, two deployables:

• Control plane (repo root) — dashboard, policy management, audit views, agent-token issuance, hosted MCP proxy. MIT except apps/web/src/ee/.
• Engine (engine/) — the MIT query-path engine, compiled to a self-contained binary. The control plane spawns it per project and never reimplements it, so hosted and self-host run the exact same engine — only the packaging differs.

apps/web              Next.js dashboard + Better Auth + projects API
packages/db           Drizzle schema (customers, projects, audit index)
packages/kms          encryptDsn / decryptDsn (env-mode dev, AWS KMS prod)
packages/router       Hosted MCP request handler — token → project → engine
engine/               The MIT query-path engine
infra/telemetry-proxy Cloudflare Worker for anonymized OSS install telemetry

Operating the managed multi-region cloud (Fly + Neon + KMS) is in docs/deploy.md.

Contributing

Issues and PRs welcome — start with CONTRIBUTING.md. The single highest-leverage contribution is a new entry in the adversarial SQL corpus: a bypass attempt and the policy fix that defeats it. Commits are DCO-signed (git commit -s). For security issues, follow SECURITY.md — don't open a public issue.

License

MIT — see LICENSE. No copyleft, no BSL, no source-available rug-pull. The one carve-out is apps/web/src/ee/ (the commercial Enterprise Edition, governed by apps/web/src/ee/LICENSE and recorded in NOTICE); deleting it leaves a fully working MIT build.

More: Docs · Pricing · Support · Design system · Code of Conduct