postgres-mcp-hardened

🚧 Version 0.1.10 — the rest of an outside review, and a fix that was too blunt

Published: binaries for five platforms with checksums, Sigstore signatures and build provenance; .mcpb bundles for one-click install; an image on ghcr.io for amd64 and arm64; a package on npm; and an entry in the official MCP registry.

0.1.8 and 0.1.9 closed six bypasses that four independent reviewers found in 0.1.7, none of them found by us. Two days of our own adversarial work had come back mostly clean the day before. Passing the tests you thought to write is not the same as looking. The one that mattered most needs no privileges at all: with a column redacted, a join on it through USING answered whether a given value was present, which is a complete equality oracle against the least-privilege reader this project tells you to configure.

0.1.10 closes the three findings that were left open. With MCP_ALLOW_SCHEMAS=public, SELECT * FROM secret.salaries was refused while describe_table handed over every column, type and default of that same table. An ordinary connection string carrying no sslmode used the driver's prefer, which sends everything in the clear whenever the server declines TLS, and anyone on the wire can make it decline. And MCP_SSLROOTCERT added a private certificate authority to 242 public ones instead of replacing them, so an operator who believed they had pinned trust to their own issuer had not.

The first version of that TLS fix was wrong in a way worth reading about. It asked "is this loopback" and demanded TLS from everything else, which took down six PostgreSQL version jobs, the conformance run, the container check and the adversarial corpus in a single push. All of them reach the database the way ordinary deployments do: postgres://user:pass@postgres:5432/db, a service name on a private network. Docker Compose and Kubernetes are not the public internet, and a server that demands TLS from a container on a bridge network is a server people switch off entirely. The question is now whether somebody untrusted can sit on the wire, not whether the address is loopback, and a real socket to a PostgreSQL reachable only by service name is in the test suite, because a rule about networks should be tested over a network.

A resource limit is documented and not solved, in THREAT_MODEL.md: 49 bytes of SQL make PostgreSQL fold a constant into 5.9 GB of backend memory during planning, and a five second statement_timeout does not stop it. The obvious shapes are refused; the general problem is upstream of anything this server can do.

Every change here was reproduced against a running server before being fixed, and each is in CHANGELOG.md with the query.

Everything here is 0.1.x because nobody outside this project has run it against their own data.

The official Postgres MCP server was deprecated in 2024 and still gets 391k downloads a month. Its entire defence is one database-level read-only transaction — and that alone does not stop every write. This is a maintained Rust replacement with defence in depth.

A drop-in Model Context Protocol server that lets an AI agent query PostgreSQL — read-only, enforced at the database level, with real SQL validation, timeouts, cost limits, OAuth 2.1, and an audit trail. Speaks Streamable HTTP and stdio, and negotiates the MCP revision: 2026-07-28 (current, and the default since upstream released it on 2026-08-03), 2025-11-25, and 2025-06-18 — what most shipping clients still speak today. A client asks for what it knows; it is not negotiated down.

Try to break it — one command, no database

The read-only guard has an offline mode. Hand it a statement and it says what it decided: no database, no configuration, nothing installed permanently.

npx postgres-mcp-hardened --validate "/* comment */ DROP TABLE users"
# REJECT: non-read-only statement: Drop

npx postgres-mcp-hardened --validate "SELECT 1; DROP TABLE users"
# REJECT: multiple statements are forbidden

npx postgres-mcp-hardened --validate "WITH d AS (DELETE FROM t RETURNING *) SELECT * FROM d"
# REJECT: non-read-only statement: non-read-only query (CTE / SELECT INTO / FOR UPDATE)

npx postgres-mcp-hardened --validate "SELECT * FROM orders WHERE id = 1"
# ALLOW

If something that writes comes back ALLOW, that is the most valuable thing anyone can send us. It needs no working exploit and no write-up — one line of SQL and "this should not be allowed" is a complete report. Anything that gets past the guard goes through SECURITY.md; everything else is an ordinary issue, and the bar for opening one is this looks wrong to me, not I am certain.

The fuzzer is deterministic and prints its seed, so whatever it finds reproduces on a machine that has never seen yours — a million mutations take about a minute:

npx postgres-mcp-hardened --fuzz 1000000
# fuzz: 1000000 iterations, seed 1592594996, slowest validation 8 ms
# RESULT: 0 invariant violations

For the whole thing against a real database, docker compose -f examples/docker-compose.yml up -d brings up PostgreSQL with sample data and the server in front of it, connecting as a role that holds SELECT and nothing else.

Every bypass found so far lives in the MUST_REJECT corpus in src/validate.rs and runs on every commit, recorded with what it cost rather than tidied away. Yours would join them.

Why

@modelcontextprotocol/server-postgres is deprecated on npm (last publish December 2024) and still sees 475,790 downloads in the 30 days to 9 August 2026. Credit where it is due: its approach is not naive — it wraps each query in BEGIN TRANSACTION READ ONLY and always ROLLBACKs, which is a real defence and one this server now adopts as well.

The problem is that it is the only defence, and it is not complete:

• A read-only transaction does not block every write, and a rollback does not undo everything it lets through. Two separate facts, and the second is the one that matters.

  gin_clean_pending_list() runs inside SET TRANSACTION READ ONLY and its work survives the rollback: an index with 25 pending pages has 0 after the transaction is rolled back. pg_backup_start() puts the session into backup state, survives DISCARD ALL, and with the default fast => false waits for a spread checkpoint while forcing full_page_writes on, which is a real cost on a busy server. pg_import_system_collations() also executes without raising SQLSTATE 25006, but be careful how much weight you put on it: that one IS undone by a rollback, so against a server that always rolls back it is a curiosity rather than a bypass.

  Reproduce it, but read the two preconditions first, because without them you will see a zero or an error and conclude we made this up. All three need superuser or ownership of the object. And the import only restores collations that are missing, so something has to be removed first:

  Dockerfile

  Copy

  -- as superuser, and note these are three separate transactions: a statement that errors
  -- inside a block aborts the whole block, so they cannot be run as one.
  DELETE FROM pg_collation WHERE oid IN (SELECT oid FROM pg_collation ORDER BY oid DESC LIMIT 200);
  
  BEGIN READ ONLY;
    DELETE FROM pg_collation WHERE collname LIKE 'zu%';  -- ERROR: cannot execute DELETE ...
  ROLLBACK;
  
  BEGIN READ ONLY;
    SELECT pg_import_system_collations('pg_catalog');    -- 200, no error
  COMMIT;                                                -- and now the rows are there
  

  Both are writes, both are inside a read-only transaction, and one is refused while the other is not. That asymmetry is why this server does not treat the transaction as its only defence. It is also why the role matters more than any of this: every example above needs privileges a least-privilege reader does not have, and this server refuses to start as a network listener when the role it was given can write. What it cannot control is which connection string somebody pastes into a client config, and the usual answer is whichever one they already had.

• No statement timeout, no cost guard, no row limit — one query can run until the server gives up.

• No authentication, no audit trail, no handling of prompt injection through returned row data.

• One source file of 143 lines, unmaintained since December 2024, no test suite.

This server keeps the rollback, adds AST validation in front of it, and adds the operational layers the original never had.

postgres-mcp-hardened vs the archived original

Install

Five ways in, in the order most people want them.

One click, for a client that accepts .mcpb bundles: download postgres-mcp-hardened-<your-platform>.mcpb from the latest release and open it. The bundle asks for the connection string and stores it in the OS keychain rather than in a plain-text config file. Nothing to install, nothing to edit.

Through npm — shortest, and the one your MCP client config can point at directly. There is no Node runtime involved at run time: the package is a launcher that fetches the native binary for your platform and verifies its checksum before running it.

{
  "mcpServers": {
    "postgres": {
      "command": "npx",
      "args": ["-y", "postgres-mcp-hardened", "--stdio"],
      "env": { "DATABASE_URL": "postgres://readonly_user:PASSWORD@localhost:5432/mydb" }
    }
  }
}

The connection string goes in env, not in args, on purpose: arguments show up in ps output and in shell history on a shared machine, and a database password does not belong there.

A binary from the releases page — one file, nothing to keep up to date, and the option to take if your machine has no Node at all. (Not a static binary, as this page claimed until 0.1.7: the -gnu and macOS targets link the system C library like any other native program. There is simply nothing to install alongside it.) Every release carries builds for Linux, macOS and Windows on x86-64 and arm64, each with a checksum and a signature; verifying them is the next section.

As a container, if that is how you run things. The image is distroless and runs as a non-root user, and the same signatures cover it as cover the binaries.

docker run --rm -p 127.0.0.1:8080:8080 --memory=512m \
  -e DATABASE_URL="postgres://readonly_user:PASSWORD@db-host:5432/mydb" \
  -e MCP_ADDR=0.0.0.0:8080 \
  -e MCP_BEARER_TOKEN="$(openssl rand -hex 32)" \
  ghcr.io/eszetael/postgres-mcp-hardened:latest

--memory is not decoration. The server idles at 7.7 MB and a normal request costs single-digit megabytes, but a caller can write SELECT repeat('x', 100000000) and drive peak memory to 400 MB — not through the result, which stays bounded at 300 bytes, but through the cost guard's own EXPLAIN, which PostgreSQL fills with the constant it folded while planning. That is a named residual risk in THREAT_MODEL.md, with the three repairs that were tried and what each one broke. Until it is closed, the memory limit is the thing that holds, so set one: --memory here, MemoryMax= under systemd.

MCP_ADDR must bind 0.0.0.0 and not 127.0.0.1, or the server listens on an interface that only exists inside the container and the published port answers nothing. The other easy one: localhost in DATABASE_URL means the container, not your machine, so a PostgreSQL running on the host needs host.docker.internal (Docker Desktop) or the host's address on the bridge (172.17.0.1 by default on Linux). Both of these were walked end to end against the published image before being written here, including that a read returns rows and DROP TABLE comes back as -32602 non-read-only statement: Drop.

From source — cargo build --release in a clone. Not cargo install: this crate is not on crates.io, and an instruction that fails is worse than one that is missing.

Checking what you downloaded

Every released binary is signed with Sigstore keyless signing — there is no private key for us to lose, and the certificate names the workflow, repository and tag that produced the file. Each artefact ships with a .sig and a .pem beside it:

F=postgres-mcp-hardened-x86_64-unknown-linux-gnu.tar.gz
cosign verify-blob "$F" --bundle "$F.bundle" \
  --certificate-identity-regexp '^https://github.com/Eszetael/postgres-mcp-hardened/' \
  --certificate-oidc-issuer https://token.actions.githubusercontent.com

Pin the identity, not just the signature. Without --certificate-identity-regexp and --certificate-oidc-issuer the check answers "somebody signed this", which is not the question. A verified certificate names the workflow, the repository and the tag that built the file — you can read it with base64 -d "$F.pem" | openssl x509 -noout -text (cosign writes the certificate base64-encoded, which surprises people who try openssl on it directly).

Older cosign builds predate --bundle; separate .sig and .pem files are published alongside for them, used as --signature "$F.sig" --certificate "$F.pem". Current cosign marks those flags deprecated, so prefer the bundle.

Public releases additionally carry SLSA build provenance, verifiable with gh attestation verify <file> --repo Eszetael/postgres-mcp-hardened.

Use it in Claude Desktop / Cursor (stdio)

{
  "mcpServers": {
    "postgres": {
      "command": "postgres-mcp-hardened",
      "args": ["--stdio"],
      "env": { "DATABASE_URL": "postgres://readonly_user:YOUR_PASSWORD@localhost:5432/mydb" }
    }
  }
}

Or run it as a remote server (Streamable HTTP)

DATABASE_URL="postgres://readonly_user:YOUR_PASSWORD@host:5432/mydb" \
MCP_ADDR="0.0.0.0:8080" \
postgres-mcp-hardened
# POST /mcp   ·   GET /health   ·   GET /ready   ·   GET /metrics

TLS: connections to PostgreSQL are encrypted whenever the server supports it, and sslmode=require, verify-ca and verify-full are all accepted (the certificate chain and the hostname are always verified, so require behaves like verify-full) — so managed Postgres (RDS, Supabase, Neon, Render) works out of the box. Certificates and host names are always verified — verified (acceptance: "a certificate naming another host is refused, by name"); for a private CA, point MCP_SSLROOTCERT at the PEM bundle. There is no "trust anything" switch.

Tip: point DATABASE_URL at a least-privilege read-only role. The server enforces read-only itself, but a scoped DB role is defense-in-depth.

Or run it on a container platform (Apify Standby)

The server needs no code changes to run as an Apify Actor in Standby mode. It reads the port the platform assigns from ACTOR_WEB_SERVER_PORT and binds 0.0.0.0 there — that port wins over MCP_ADDR, loudly, on stderr, because binding anywhere else means the run is never marked ready and the failure looks like a mysterious timeout. GET / answers the platform's readiness probe (x-apify-container-server-readiness-probe) without touching the database: container readiness is not database readiness, and a probe that waits on a busy pool turns a slow database into a container that never starts.

Input is a JSON-RPC request in the POST body — initialize, tools/list, tools/call, resources/list, resources/read, server/discover. Output is a JSON-RPC response; from 2025-11-25 a refused statement comes back as a tool execution error (isError: true) with the reason in the content, so the model can rewrite the query. tools/list is the authoritative description of every argument.

Authentication there is the platform's, not ours. Apify checks the caller's token before routing to the container, so the server does not additionally demand MCP_BEARER_TOKEN — requiring a second secret would mean an agent that finds this server cannot call it. That exemption is narrow: it needs both APIFY_IS_AT_HOME and ACTOR_WEB_SERVER_PORT, one alone changes nothing, and the server card then reports "type": "apify-platform" rather than claiming a lock we do not hold. Everywhere else the server still refuses to start on a network address with no authentication. Set MCP_BEARER_TOKEN as well if you want a second lock on the same door.

The other start gate is unchanged and matters more here: a role that can write is refused a network listener. Point DATABASE_URL at a read-only role — --print-setup-sql writes the statements.

Migrating from the deprecated server

The most-discussed problems reported against @modelcontextprotocol/server-postgres were reproduced against this server; here is how each behaves:

Testing

Beyond unit tests, the repository carries two harnesses that run in CI on every change:

• --fuzz — a deterministic fuzzer that mutates a corpus of known writes with transformations that do not change SQL meaning (comments, case, dollar-quoting, invisible Unicode, parentheses) and asserts that none of them ever becomes an allowed statement.
• tests/acceptance.sh — an end-to-end suite that starts its own PostgreSQL and checks 317 behaviours: every write-bypass reported against the deprecated server (including the COMMIT/END injection), truthful results, schema introspection, protocol conformance, configuration mistakes failing loudly, audit tamper detection, fair use under load, and multi-database deployments.

Every reported problem, answered

docs/COMMUNITY_ISSUES.md is the complete ledger: every problem reported against the deprecated server and every open issue against the maintained alternatives, each with what happens here — including the handful we could not fix in code, said plainly.

What we learned from the alternatives

Every server in this space has an issue tracker, and those trackers are a map of what goes wrong. The ones we deliberately built against:

• A published image that lags the code. The most-supported open complaint against the leading alternative. Our container is built and pushed from the same tag that produces the binaries, so it cannot drift.
• A hardcoded query timeout. Also among their most requested settings. MCP_STATEMENT_TIMEOUT is configurable and validated at startup.
• Unrestricted access by default. Some servers default to read/write and rely on the operator to restrict it. This one has no write path at all.
• Credentials in the client configuration. MCP_PASSWORD_FILE keeps the password out of it.
• Tables in a non-default schema silently not found. MCP_SEARCH_PATH fixes the lookup, and the tools take an explicit schema anyway.
• Deprecated transport. HTTP+SSE was replaced by Streamable HTTP in 2025-03-26, three revisions ago (this page said 2025-06-18 until 0.1.7, which was wrong by one revision; the specification's own changelog for 2025-03-26 records the replacement). We speak the current transport.

Troubleshooting

Answers to the questions people actually asked about the deprecated server, so nobody has to open an issue to find them.

spawn npx ENOENT / "which Node version do I need?" — none. This is a single native binary, with no runtime to install beside it. Download it from the releases page and point your client at the file. There is no node_modules, no npx, nothing to keep up to date.

"The server starts but nothing is listening on a port." — that is stdio mode, which is correct for Claude Desktop and Cursor: the client talks to the process over its standard input and output, not over a socket. If you want a network endpoint, start it without --stdio; it then prints MCP HTTP listening on http://… and speaks Streamable HTTP.

"Can my client on another machine reach the database?" — yes: run the server next to the database in HTTP mode, expose it, and enable OAuth (JWT_PUBKEY_PEM, JWT_AUD, JWT_ISS). The database credentials then never leave the host the server runs on.

"Could not attach to MCP server." — the process exited before the handshake. Run the same command in a terminal: a configuration mistake prints its reason and exits with status 2 rather than dying quietly, and a connection problem is reported on the first query with the cause.

self-signed certificate in certificate chain / unable to verify the first certificate — your provider uses a private CA (Supabase, GCP and RDS all do). Download their CA bundle and set MCP_SSLROOTCERT to it. The error message names the step for your provider. We do not offer a "trust anything" switch.

Managed providers

This server always verifies the database certificate, including with sslmode=require. That is a deliberate deviation from libpq, where require encrypts without verifying and a machine in the middle can therefore read and rewrite every query and result without anyone noticing. The cost of being strict is that a provider with a private CA needs one extra step; the cost of being lax is that you never find out. If you disagree with the trade-off, verify-full with the bundle below is the same amount of work and leaves no doubt either way.

Not sure which case you are in? Ask the server itself, before configuring anything:

echo | openssl s_client -starttls postgres -connect YOUR_HOST:5432 2>/dev/null \
  | openssl x509 -noout -issuer

A well-known issuer (DigiCert, ISRG, Google Trust Services) means it will just work; anything naming your provider means you need their bundle.

With a connection pooler (Supavisor, PgBouncer) in transaction mode, note that this server sets statement_timeout and idle_in_transaction_session_timeout per session and runs every query in an explicit read-only transaction. Both are compatible with transaction pooling; session-level SET outside a transaction is not, which is why we do neither.

no pg_hba.conf entry … no encryption — the server accepts only TLS connections for that host and user. Add ?sslmode=require to the connection string.

INVALID_URL / invalid connection string — a password containing @, :, /, # or ? must be percent-encoded (@ → %40, : → %3A, / → %2F, # → %23).

"My table has hundreds of partitions and the list is unusable." — partition children are hidden by default; the parent is listed. Set MCP_SHOW_PARTITIONS=1 if you need them.

"I need production and staging at the same time." — either run one server per database (they are distinguishable: set MCP_SERVER_LABEL), or configure both in one server with MCP_DATABASE_URLS and pass database in the tool arguments.

Resources

Every table and view is exposed as an MCP resource (postgres:///<schema>/<table>/schema), so a client can browse the schema without issuing a query — the same capability the deprecated server offered, plus column comments, primary keys and foreign keys in the payload.

When the database has not answered, resources/list returns an empty list with the reason in _meta rather than a protocol error. A catalogue inspecting this server starts it with no database at all and calls resources/list straight after initialize; answering that with an error reads as a server that does not work. The empty list is not a claim that there are no tables — initialize says the database has not answered, security_posture gives the detail, and the reason travels with the list itself. A database that does answer and refuses is still an error, because reporting "no resources" for a missing privilege is the silent failure this server exists to avoid. verified (acceptance: "a host can inspect the server with no database and mcp-proxy in front")

Tools

• explain_query — the execution plan; with analyze it runs the statement and reports real timings and buffer usage, which is safe here because the statement is validated read-only and runs inside a transaction that is always rolled back. The plan comes with a summary: which node spent the time (self time, not inclusive), and where the planner's row estimate was furthest from reality — because a bad estimate is usually why the plan is bad.
• database_health — cache hit ratio, connections (this database and the cluster), the longest running statement and the longest abandoned transaction as separate figures, vacuum backlog, invalid indexes, sequences near their ceiling, replication lag, tables that have never been analysed (no planner statistics — the usual reason a database looks healthy and runs slowly), and the window the counters cover. Anything the role cannot see is declared rather than returned as a confident zero.
• analyze_indexes — unused indexes, duplicates, and tables scanned sequentially often enough that an index would pay off.
• top_queries — the heaviest statements, from pg_stat_statements.
• security_posture — what this deployment is actually able to do to your database, asked of PostgreSQL rather than assumed: whether the role can write, bypass row-level security or reach server files; whether the transport is authenticated; whether the audit chain is keyed; whether the connection is encrypted. Returns a grade — the worst finding, never an average — and, for anything wrong, the command that fixes it. The same summary reaches the model through initialize, because under stdio nobody sees stderr and the agent is the only messenger the operator has.
• query — run a read-only SQL query (validated, auto-LIMIT, cost-guarded). The response states what it did: returnedRows, appliedLimit, truncated, plus requestedLimit when a larger request was capped at the 10000-row maximum, offset when paging, and redactedColumns when masking is configured — so an agent never has to guess whether it received the whole answer.
• list_schemas, list_tables, describe_table — progressive schema discovery (parameterized, injection-safe). describe_table returns the schema comments (COMMENT ON TABLE/COLUMN), primary keys, foreign keys and defaults, so the agent reads what a column means and what it points at instead of guessing from its name — and a missing table is an error, not an empty column list.

Protocol revisions

The server answers initialize with the revision the client asked for when it implements it, and with its newest otherwise. Over HTTP the revision comes from the MCP-Protocol-Version header, per request — one client's negotiation cannot change the contract another client is served under.

If a request carries no header, the server does not fall straight back to the oldest contract. It reads the revision this session agreed on at initialize, which is what the transport specification asks for: the default applies only "if the server does not receive an MCP-Protocol-Version header, and has no other way to identify the version — for example, by relying on the protocol version negotiated during initialization". A session is that other way, so a client that negotiated 2025-11-25 and then omitted the header keeps the contract it agreed to rather than being silently demoted.

A header we cannot parse is a different matter from a header that is absent, and the specification is explicit about it: "If the server receives a request with an invalid or unsupported MCP-Protocol-Version, it MUST respond with 400 Bad Request." A version we do not implement — 2025-03-26 or not-a-date — is refused with 400 and the list of revisions we do speak, rather than served under a contract the client never agreed to. Falling back is for silence, not for disagreement.

Only a request with neither a header nor a session falls back, and it falls back to 2025-06-18 — the oldest revision this server implements — rather than the 2025-03-26 the specification names. That revision is not implemented here, and answering under a contract the server cannot honour would be worse than answering under the oldest one it can.

The difference that matters is where a refusal goes. Under 2025-06-18 "this statement is not read-only" was a JSON-RPC error: the client saw a broken call and the model often never saw the reason. From 2025-11-25 (SEP-1303) it arrives as a tool execution error — isError: true with the reason in the content — so the model rewrites the query instead of handing the user a failure. What does not change is the audit: the refusal is recorded by the code that refuses, and the acceptance suite asserts both halves together, so friendlier errors can never quietly mean a quieter log.

Mcp-Method and Mcp-Name are held to agreement, not presence. The draft requires them; earlier revisions do not, and demanding them would break every client shipping today. But a gateway that routes or authorises on Mcp-Method while the server executes the body has decided about a different request than the one that runs — and that is true whatever revision is in force. So a header that is present must match the body under every revision, while a client that sends none is untouched.

Protocol failures stay protocol failures. A malformed envelope, an unknown method or a missing token is not something a model can fix by rewriting SQL, and a client's error handling expects those where they have always been.

The next revision, before it lands

2026-07-28 is the largest break MCP has had: no initialize, no session header, no ping. That identifier comes from LATEST_PROTOCOL_VERSION in the draft schema, and it is not a release date — MCP names a revision for the last date a backwards-incompatible change was made, so it describes the draft's history rather than a schedule. Every request carries its own protocol version in _meta, and a new server/discover replaces the handshake. We implemented it early behind a switch, because a draft moves and a server advertising support for a moving target will be wrong in public. Upstream cut schema/2026-07-28 on 2026-08-03 — the released schema differs from the draft we had verified against in four documentation URLs and nothing else — so the switch is gone and this is what the server speaks by default. Clients on 2025-11-25 and 2025-06-18 are answered as before.

server/discover answers under every revision, because the specification expects clients to use it as a backwards-compatibility probe — which only works if older servers answer it. Ours answers with the revisions we speak and, in _meta, the full security posture. That is deliberate: a client can learn it is talking to a server connected as a superuser before it sends a query, as structured data rather than prose a model has to notice.

Two of the draft's rules are security controls here, not formalities. Mcp-Method and Mcp-Name must agree with the request body, and we refuse the mismatch (-32020) — the headers exist so a gateway can route and authorise without parsing the body, and if header and body may disagree, then the thing that authorised and the thing that executes saw two different requests. And a client stating a version we do not implement is told so (-32022) rather than quietly served under a contract it never agreed to.

What the safety costs

Measured, not asserted: tests/bench/, against the pg driver running the same query on the same machine (PostgreSQL 18.6 in Docker, 50k-row table, 300 sequential requests, rate limit off). Re-measured 2026-08-17 on a shared VPS at load average 2.6 — median of two runs:

An earlier table here said +3.6/+5.2/+3.7 and "about 4 ms". Those came from a quieter machine, and the driver floor moved with them — 0.28 ms against today's 0.43 for the same lookup — so it was the hardware talking, not the code. Two things were checked before changing the number, because the obvious suspect was our own build: the 0.1.6 binary, built before lto was switched on, measures +5.4/+7.8/+5.9 on this machine within the same hour. Identical. The release profile halved the binary and cost nothing here.

Expect 5 to 8 ms per query, and treat any single figure on this page as a reading from one machine on one day. The shape matters more than the size: the overhead is nearly constant. If the AST validation were the cost, it would grow with the query. It does not. The time goes on round trips — the session is reset, the timeouts and read-only flag are set, a read-only transaction is opened, the cost guard plans the statement, then the query runs and the transaction is rolled back. Five or six exchanges where the driver has one.

That is a deliberate trade and you can see exactly what it buys. For an agent making tens of calls it is invisible; if you are putting this in front of a latency-critical serving path, you are using the wrong tool, and it is not one.

Under concurrency the interesting number is not throughput but what happens past the limits: at 8 concurrent clients it served 373 requests/second and turned away 192 more with "too many requests in flight", which is the in-flight cap doing its job rather than a queue growing until something falls over.

Would this index help? — answered without creating one

The one capability the leading alternative is genuinely known for is index tuning: it can tell you an index would pay off before you build it. It gets there by defaulting to a connection that can create real indexes — safe only if you remembered to restrict it.

simulate_index answers the same question from a connection that cannot write anything. hypopg registers a hypothetical index in backend memory: the planner sees it, storage never does, and it is gone when the call returns. You get the plan and cost with and without, and — separately — whether the planner actually reached for it, because a cost that barely moves and an index the planner ignored are different answers.

The tool takes a table and a list of columns. Not a CREATE INDEX statement. The definition is assembled server-side from identifiers the catalogue confirmed exist, quoted by PostgreSQL itself, so there is no path from a tool argument to arbitrary DDL — a column name carrying SQL dies on the catalogue lookup, and there is a test that fires exactly that. The numbers are planner estimates: treat a large improvement as a reason to test the index, not as proof.

Conformance is checked by somebody else's client

Every other test here is our harness talking to our server. If we misread the specification, we misread it the same way in both halves and everything passes. So CI also drives the server with the official MCP SDK — the client library the ecosystem uses — over stdio and Streamable HTTP: handshake, tool listing and schemas, a read, a refused write arriving as a tool execution error rather than a protocol one, resource listing and reading. A protocol mistake shows up as a client that cannot talk to us. tests/conformance/.

Working on this

git config core.hooksPath .githooks   # once, per clone

.githooks/pre-push runs format, clippy, the unit tests and the documentation-claim checks before anything leaves your machine. It exists because of a specific mistake: a commit went out with a failing clippy lint, and the first anyone knew of it was a failure email. Note that cargo test passes on that code — clippy lints are not compiler errors — so "it builds locally" is not the same answer as "CI will be green".

It deliberately skips the acceptance suite and the PostgreSQL matrix: those need Docker and about twenty minutes, and a hook people cannot afford to run is a hook people bypass. CI runs everything. git push --no-verify when you want to see something fail in CI on purpose.

Setting up the role

DATABASE_URL=postgres://admin@host/mydb postgres-mcp-hardened \
  --print-setup-sql --role mcp_reader --schemas public --redact ssn,email > setup.sql
# read it, then:
psql -v pw="$(openssl rand -base64 24)" -f setup.sql mydb

Run with a connection string and the table and column lists come from the catalogue; without one you get the same document with placeholders. The difference matters most for redaction: the columns to grant back have to be read from the database, because writing them from memory is how a column meant to stay hidden gets handed back.

The output ends with checks that return no rows when it worked, and a reminder that the server itself will tell you what the role can do the moment you point it at the database.

Limiting what the server can reach

MCP_ALLOW_SCHEMAS and MCP_ALLOW_TABLES restrict which relations a query may touch. Either one turns the allowlist on; schema.* and schema.table both work.

MCP_ALLOW_TABLES='public.customers,public.orders,analytics.*'

The check reads the query plan, not the SQL. That is the whole design: the planner has already applied search_path, resolved every alias, expanded views to base tables, and knows that a CTE named customers is not the table customers — so WITH customers AS (SELECT 1) SELECT * FROM customers runs and touches nothing, while WITH x AS (SELECT * FROM salaries) SELECT * FROM x is refused. Reading the statement instead is what lost three rounds of adversarial review.

Two consequences worth knowing before you turn it on:

• A partition rides on its parent. You allow events; PostgreSQL decides which children to read.
• A view needs its base tables allowed too, because the plan names those. Allow both, and let the database privileges keep the base table unreachable directly — that is the boundary in any case.

pg_catalog and information_schema are outside the surface unless MCP_ALLOW_CATALOG=1: an agent that can still read the catalog can enumerate exactly what the allowlist was meant to hide. The schema tools keep working, because they run fixed queries rather than caller SQL.

That covers three routes to the same facts, because for a while it covered only one. A catalogue view plans to scans over real relations and the plan names them — but pg_settings plans to a single Function Scan on pg_show_all_settings, naming no relation at all, and current_setting() is a scalar call that never appears as a scan. Both used to return the server's configuration under an active allowlist. Functions whose name begins with pg_, plus current_setting, inet_server_addr and inet_server_port, are now refused alongside the catalogue relations, and MCP_ALLOW_CATALOG=1 opens all of them together. Ordinary set-returning functions — generate_series, jsonb_each, unnest, regexp_split_to_table — carry no such prefix and are unaffected. current_user, session_user, current_database and version stay readable on purpose: an agent already knows what it connected to and as whom.

The corpus of things that got through

Every shape that defeated a control during review lives in tests/adversarial/, with the round in which it stopped working. It runs on every build, and — because the cases are written against placeholders rather than our fixture — you can point it at your own database:

ADV_URL='postgres://…' \
  ADV_TABLE=people ADV_TABLE2=orders ADV_REDACT_COL=ssn \
  ./tests/adversarial/run.sh

ADV_TABLE needs the sensitive column, ADV_TABLE2 is any other readable table, ADV_REDACT_COL is the column to redact. All three matter: this example omitted ADV_TABLE2 until 0.1.7, so it kept its default of film — a table from the Pagila sample database — and following the instruction exactly produced three "mismatches" that were only a missing relation. The harness now checks the three up front and says which one is wrong, because a security corpus that reports a typo as a failed control teaches you to ignore the failures that matter.

A security claim you can only check by reading our source is a claim you have to take on trust, and this project's own history is the argument against that.

Security model

The full statement of what this server guarantees, what it does not, and which control enforces which promise is in THREAT_MODEL.md — including the controls that have been defeated in review and are therefore described as depth rather than as boundaries.

• Encrypted transport: TLS to PostgreSQL via rustls (no OpenSSL in the image), certificate verification always on, private CAs via MCP_SSLROOTCERT.

• Read-only, two ways: every statement is parsed with sqlparser and rejected unless it's a SELECT/WITH/EXPLAIN/SHOW; the DB session is additionally set default_transaction_read_only = on.

• Anti-DoS: enforced statement_timeout, auto-injected LIMIT, and an EXPLAIN-based cost guard that rejects expensive plans before they run.

• Sensitive columns — defence in depth, and honest about it: MCP_REDACT_COLUMNS masks values at every depth and refuses to run a query that references those columns, including the ways round it that an adversarial panel actually found — renaming (SELECT password AS pw), wrapping (md5(password)), serialising the whole row (row_to_json(t), t::text, json_agg(t)), and naming the column as a string rather than an identifier (to_jsonb(t) ->> 'password', #>> '{password}', $.password), whole-row wildcards (ROW(t.*)::text) and positional renaming ((SELECT * FROM staff) AS x(c1, …, c9)). It is still name-based filtering, and name-based filtering cannot be a boundary against the whole SQL language — four adversarial rounds each got past it through a shape nobody had listed.

  The fourth, in 0.1.7, did not find a new shape of name. It went around names altogether: SELECT get_raw_page('people', 0) hands back 8192 bytes of the table as the disk holds it, and every value on that page is in there, including the redacted one. Demonstrated, not theorised — with MCP_REDACT_COLUMNS=ssn, SELECT ssn FROM people was refused while the raw page came back with the social security numbers in plain ASCII. pageinspect and its relatives are now refused as a category of their own, because "this returns storage rather than columns" is a different problem from "this writes", and telling an operator which one they hit is worth a separate message.

  The same round found the quieter version of it. PostgreSQL's planner keeps a sample of each column's real values, and pg_stats publishes them: with 3,000 rows, SELECT * FROM pg_stats WHERE tablename='people' returned {123-45-6789,555-00-1111,987-65-4321} while SELECT ssn FROM people was refused. That query never names the redacted column, so a name-based rule has nothing to act on. The value-bearing statistics columns — most_common_vals, histogram_bounds, most_common_elems, stavalues1…stavalues5 — now join whatever you configure, whenever you configure anything. The rest of the view is untouched: n_distinct, null_frac and correlation are what the index advice below is built from and they carry no values, so removing the whole relation would have broken ten columns to fix four.

  Two more doors turned out to open on the same room. A value too long for its row is stored in a TOAST table, and that table is readable by name: SELECT chunk_data FROM pg_toast.pg_toast_16384 returned the redacted text in the clear. pg_largeobject is the same thing for large objects — the bytes underneath lo_get. Both are refused now, as relations rather than functions, with a message that says why: they hold physical storage rather than columns. The catalogue that merely describes the database is untouched — pg_tables, pg_stat_activity, pg_largeobject_metadata, and the statistics columns index advice needs.

  All four say the same thing, and it describes this feature's limit better than any list of patches: a column filter protects columns, so anything that reads underneath columns is outside what it can promise. Raw pages, planner samples, TOAST chunks and large-object bytes are four doors into that space, all four found in a single afternoon by looking for the shape rather than the names. The honest assumption is that there are more, which is why the database role and the read-only transaction are the real boundary and this stays what it says it is: defence in depth.

  So the server stops asserting and asks the database: at startup it reports every table where the connected role can still read a redacted column, with the exact statements that fix it, and MCP_REDACT_REQUIRE_REVOKE=1 turns that report into a refusal to run. Note the fix is a table-level REVOKE followed by a GRANT of the columns that stay — a bare REVOKE SELECT (password) ON staff is silently a no-op while the role holds SELECT on the whole table. With column-level grants PostgreSQL then refuses SELECT * on that table, so callers name columns instead; describe_table lists them and marks the redacted one.

• Prompt-injection aware: row data is returned inside a trusted="false" provenance block with delimiters escaped, so a malicious cell can't hijack the agent.

• It generates the role you should be running as: --print-setup-sql writes the DDL for a role that inherits nothing, bypasses nothing, creates nothing, reads only the relations you name, and — where you have named sensitive columns — has them revoked in the order that actually works. It prints; it never executes. Applying this needs administrative rights, and a tool whose whole identity is "read-only" has no business holding an administrator's password.

• It will not expose a role that can write: when the listen address is reachable from the network, the server asks PostgreSQL what the connected role is actually allowed to do — superuser, BYPASSRLS, membership of pg_write_all_data and friends, and write privileges on a bounded sample of tables — and refuses to start if the answer is more than "reader", naming each reason and pointing at --print-setup-sql. It refuses an unauthenticated network listener for the same reason. Loopback and stdio are left alone: there the caller is the operator. The overrides (MCP_ALLOW_EXCESSIVE_ROLE, MCP_ALLOW_ANONYMOUS_NETWORK) take the literal value i-accept-the-risk so they cannot be switched on by a typo, and they are recorded in the audit log. This server enforces read-only itself, but that enforcement is code, and code has been wrong before; a role that cannot write is the part no bug of ours can undo.

• A browser cannot reach it: a request carrying an Origin is refused with 403 unless the operator listed that origin, and on a loopback listener a Host that is not localhost is refused too — the shape a DNS-rebinding attack takes when it aims at a database server on your laptop.

• The audit knows the configuration: the chain opens with a startup record naming the version, the transport and every setting in force, with connection passwords stripped and secrets reduced to fingerprints, plus a config_fp an operator can pin across restarts. A log that says what happened but not under which settings cannot answer the first question an incident asks.

• The audit notices being shortened: a hash chain proves entries were not altered, but a log with its tail cut off is internally consistent — recomputing it finds nothing wrong. Alongside MCP_AUDIT_LOG the server therefore keeps <log>.hwm, a one-line record of the last sequence number and hash it wrote, updated only after the entry is durably appended. On start the two are compared, and a disagreement is reported: entries missing from the end, a rewritten last entry, or a log that has gone away entirely. This is not proof of tampering — an unclean shutdown looks the same — but a tamper-evident trail owes you the question, not the verdict. Keep the sidecar with the log when you archive or move it; deleting it only loses the truncation check, never an entry. The offline verifier is unchanged and still needs an external anchor: --verify-audit <file> --expect-last <hash>. verified (acceptance: "a shortened log is noticed at startup, without any external anchor")

• A wrong setting is fatal, not merely wrong: an unparsable listen address, an audit file that cannot be written, sslmode=disable to a database on another machine, a metrics token that is also the database credential, a boolean spelt yes — each used to be accepted and quietly do something other than what was meant. Startup now stops and names the setting.

• A misspelt setting is fatal — somebody else's setting is not: MCP_REDACT_COLUMN (singular) used to start the server with redaction quietly switched off, so a near miss of a real setting still stops startup and names the intended spelling. A name that resembles nothing we define was set by another program sharing the environment: it is reported and ignored. mcp-proxy, which every catalogue puts in front of a server to inspect it, exports MCP_PROXY_DEBUG — until 0.1.6 that one variable made this server exit before reading a request. MCP_X_* remains reserved for the operator's own use. verified (acceptance: "a misspelling is still fatal")

• No schema leaks: database errors are mapped to structured, actionable messages that never echo table/column names.

• OAuth 2.1: optional RS256 bearer-token validation (signature, exp, aud, iss) with scope enforcement; disabled when unconfigured for local/self-host use.

• Audit: every tool decision is logged as a tamper-evident, hash-chained JSON line (no raw SQL).

• Supply chain: dependency licences, sources and advisories enforced in CI (cargo deny, cargo audit); a CycloneDX SBOM is attached to every release.

• Runtime: ships as a distroless, non-root container — 14.8 MB to download, 41 MB on disk for linux/amd64 at 0.1.6, built and smoke-tested in CI. Both numbers, because a single one is always the flattering one: docker images shows the second, your bandwidth pays the first.

Footprint

Measured on an ordinary VPS against a 16k-row sample database, so you can check the "written in Rust" claim rather than take it:

Two lines here were wrong until 0.1.7. Idle memory said 5.2 MB and measures 7.7 — five starts under identical conditions landed within 0.1 MB of one another, so the old figure is not noise, it is a different measurement whose method was not written down. The binary line said "11 MB, static", and the file people actually downloaded was 18.9 MB and dynamically linked. The size was never measured on a release build, because this crate had no [profile.release] at all, so more than five megabytes of debug symbols shipped to every user. Setting strip, lto and codegen-units = 1 took it to 9.2 MB. The word "static" was simply not true of any of the five targets we publish, none of which is a musl build. The image shrank with the binary, from 14.8 MB compressed in 0.1.6 to 12.6 MB — both figures read from the registry manifest of the published image rather than from a local build, because a local build is not what anybody pulls.

A twelve-minute soak of mixed traffic (reads, refusals, errors, aborted requests, session churn, unauthenticated requests) served 51,499 requests and ended with the same 15 open file descriptors it started with. Resident memory went from 8.1 MB to 11.2 MB, and the shape of that is the interesting part: 8.1 to 10.7 happened inside the first 400 requests, and the remaining 51,000 added 0.5 MB in a curve that flattened as it went. That is an allocator settling, not a leak. This page used to say memory stayed "flat", which was true of everything after the first few seconds and not true of the number, so here is the number.

Reproduce it with tests/soak.sh rather than believing the paragraph.

Configuration

License

MIT — see LICENSE. The core is MIT and stays that way.

Commercial and team use

Pointing this at a production database inside an organisation raises questions the MIT core does not answer: policy bound to an identity rather than to a scope, an audit log shipped somewhere it cannot be quietly edited, evidence a compliance reviewer will accept, a deployment somebody is on the hook for.

A team edition covering those is being scoped right now, and what goes into it is not decided. If that is what your organisation needs, write to eskulapstudio@gmail.com and say what would have to be in it. There is nothing to buy yet — the answers are what decides whether it gets built at all, and in what order.