Skip to content

Releases

v0.1.1

Choose a tag to compare

@github-actions github-actions released this 13 Nov 19:29

mssql-readonly-mcp is an MCP server that lets an AI assistant read one on-premises SQL Server database. Its SQL tool accepts only text that passes an allowlist for one plain SELECT, and it will not start with a login that fails its permission check (what that check covers and what it does not). It keeps a tamper-evident record (a hash-chained audit file) of every question asked. This release closes, best effort, a session that a Timeout on a slow link could leave open in v0.1.0 (U-1R; retested on SQL Server 2022 on Windows only), adds an icon to the bundles and corrects the docs. The gate, the login check, the tools and the audit format are unchanged.

Changes

  • U-1R: a Timeout on a slow link could leave its session open in v0.1.0. In attack test run 5, on a link throttled to 5 KB/s, 8 of 27 single calls (limit 3 s) and 21 of 450 calls repeated 50 in a row (limit 2 s), each answered Timeout, left the reader's session at SQL Server with an open read transaction and a lock (no data leak, no audit gap). The cause, diagnosed after v0.1.0 with a build that logs the client library's state (SQL Server 2022 image, on Windows): the client library's command timeout fires in the middle of a row while the server's reader close drains the throttled result; the library breaks the connection, and its parser becomes Closed, but the same read goes on parsing, reads row data as a token and breaks the connection a second time, which leaves the parser Broken. v0.1.0 closed only a connection whose parser was Closed, so it left this one open. The server now also closes a connection whose parser is Broken, when the client library has doomed it, the call's own SqlConnection is closed and no other SqlConnection holds it, with the earlier checks (the network handle is still there, no network callback is pending). The close is still best effort, and it fails closed: when an internal member of the client library that it reads is missing, it closes nothing, and the Gate tests fail. New unit tests (LeftOpenConnectionsDecisionTests) check each condition without SQL Server.
  • Bundle icon. The .mcpb bundles carry the project logo as their icon (icon.png); the v0.1.0 bundles had none. A win-x64 bundle built locally with the release script held the icon unchanged, and its manifest named it; a bundle built by the release workflow was not checked, nor how a host shows the icon.
  • README images. The banner and the two diagrams are one self-contained dark card each, instead of a light and a dark variant: with GitHub's theme set to follow the system, a dark GitHub theme on a light system showed the light images on a dark page.
  • Registry file. server.json at the repository root is the entry published to the official MCP registry; Releasing has the step that updates it after a release. At this tag it still names v0.1.0; it is updated after this release is published.
  • Docs. The sentences written before the v0.1.0 release now say what was checked after it; stale sentences found in a full audit are corrected; the issue forms carry their labels and ask for the release version; the U-1R sentences in the README, SECURITY.md and the docs describe the fix.

Measured for the U-1R fix

Retest on 2026-11-13, on Windows, against the SQL Server 2022 image (2022-CU27-ubuntu-22.04), one call per run in a new server process: a Timeout on a link throttled to 5 KB/s from SQL Server to the client, limit 3 s, the throttle released 10 s after the limit. In 100 of 100 runs the call was answered Timeout (3.5 to 3.8 s after it started), the client closed the connection 6.0 to 6.4 s after the start, while the link was still throttled, the call's session was gone 6.0 to 6.6 s after the start, the next call worked and the server exited with code 0; no run left the session. A diagnostic build of the same code, which also logs the client library's state and the close decision (this instrumentation can change timing), 25 runs of the same case: none left the session; the parser was Broken in 2 of them (in both the connection was doomed and had no owner) and Closed in 23, and the server closed the connection in all 25. In the diagnosis, the same case with the throttle held 60 s, on the code before the fix, left the session in 5 of 12 runs. Unit tests on Windows on 2026-11-13: Gate 416, Audit 90, all passed. The code of this release differs from the retested build only in source comments.

Not shown for the fix: SQL Server 2019 and 2025 in the throttled case; the 50 calls in a row of run 5; a link throttled for 60 s as in run 5; a real slow network (the retest throttles a local link); Linux and macOS. The attack tests were not run again (Security testing, U-1R).

Files

File Content
mssql-readonly-mcp-0.1.1-win-x64.zip Windows x64: one folder with the build, LICENSE and README.txt
mssql-readonly-mcp-0.1.1-linux-x64.tar.gz The same for Linux x64
mssql-readonly-mcp-0.1.1-osx-arm64.tar.gz The same for macOS on arm64
mssql-readonly-mcp-0.1.1-win-x64.mcpb MCPB bundle for Windows x64: manifest.json, icon.png, LICENSE, the build under server/
mssql-readonly-mcp-0.1.1-linux-x64.mcpb The same bundle for Linux x64
mssql-readonly-mcp-0.1.1-osx-arm64.mcpb The same bundle for macOS on arm64
SHA256SUMS SHA-256 of the six files above

Each build is self-contained, as in v0.1.0: the .NET runtime is included (not trimmed, not Native AOT). The program is mssql-readonly-mcp (mssql-readonly-mcp.exe on Windows); on Linux it needs the ICU libraries (Prerequisites). No code signing: the Windows executable is not Authenticode-signed, the macOS executable is not notarized, and the bundles are not signed with mcpb sign. No builds for other platforms (details). Still not shown: the Linux and macOS builds started on Linux and on a Mac, a start on a machine with no .NET, the executable bit in the osx-arm64 archive and in the bundles, and a v0.1.1 bundle installed in any host (details).

Check a download

Compare the file's SHA-256 with its line in SHA256SUMS (on Linux: sha256sum --check --ignore-missing SHA256SUMS in the download folder):

(Get-FileHash .\mssql-readonly-mcp-0.1.1-win-x64.zip -Algorithm SHA256).Hash.ToLowerInvariant()

The release workflow attests every file in SHA256SUMS (build provenance). The check needs the GitHub CLI (details):

gh attestation verify .\mssql-readonly-mcp-0.1.1-win-x64.zip --repo oguzhanpisgin/mssql-readonly-mcp

Tested on

SQL Server 2019, 2022 and 2025, as Linux container images pinned by exact tag: mcr.microsoft.com/mssql/server:2019-CU32-GDR13-ubuntu-20.04, 2022-CU27-ubuntu-22.04 and 2025-CU9-GDR1-ubuntu-22.04. Other versions and Windows installations were not tested. In the last full run before this release (2026-11-13, on Windows, one run per image) all tests passed: Gate 416, Audit 90, and Integration 259 of 259 on each of the three images (Building and testing). One integration test, Truncated_result_whose_cancel_is_held_until_the_network_works_again_holds_its_locks_until_then_only, is still not stable: it passed in those three runs, which does not show it stable, and failed in 4 of 19 runs on the 2022 image, on Windows, on earlier code; no failure message was captured then, so the cause of those failures is not known. The failures whose message was captured after v0.1.0 (one in 40 local runs on the 2022 image, one CI run on the 2019 image) had a cause other than U-1R: the request had ended while the cancel was held; in the local one, SQL Server ended it on its own because the whole result fit into the network buffers between SQL Server and the test's proxy. The fix does not change that; it can fail in a CI run or a local run. CI runs on every push and pull request: the build and unit tests, then, when they pass, the integration tests on each of the three images.

What it does not solve, and known limits

Unchanged from v0.1.0 except U-1R above. Read What it does not solve and SECURITY.md, "Known limits" before you install it. The rows the assistant reads still go to the AI vendor that runs the model; read-only does not change that.

Security testing

The record of the security testing is in docs/security-testing.md; it lists what differs from the code the attack tests ran on. Report a vulnerability privately through GitHub private vulnerability reporting, not in a public issue (SECURITY.md).

Start

Follow the five-minute start in the README. Which of its blocks were run: [Installing](https://github.com/oguzhanpisgin/mssql-readonly-mcp/blob/v0.1.1/docs/install.md#what-was-t...

Read more

v0.1.0

Choose a tag to compare

@github-actions github-actions released this 13 Nov 14:40

mssql-readonly-mcp is an MCP server that lets an AI assistant read one on-premises SQL Server database. Its SQL tool accepts only text that passes an allowlist for one plain SELECT, and it will not start with a login that fails its permission check (what that check covers and what it does not). It keeps a tamper-evident record (a hash-chained audit file) of every question asked. This is the first release.

Files

File Content
mssql-readonly-mcp-0.1.0-win-x64.zip Windows x64: one folder with the build, LICENSE and README.txt
mssql-readonly-mcp-0.1.0-linux-x64.tar.gz The same for Linux x64
mssql-readonly-mcp-0.1.0-osx-arm64.tar.gz The same for macOS on arm64
mssql-readonly-mcp-0.1.0-win-x64.mcpb MCPB bundle for Windows x64: manifest.json, LICENSE, the build under server/
mssql-readonly-mcp-0.1.0-linux-x64.mcpb The same bundle for Linux x64
mssql-readonly-mcp-0.1.0-osx-arm64.mcpb The same bundle for macOS on arm64
SHA256SUMS SHA-256 of the six files above

Each build is self-contained: the .NET runtime is included (not trimmed, not Native AOT). The program is mssql-readonly-mcp (mssql-readonly-mcp.exe on Windows). On Linux the README lists the ICU libraries (for example the libicu package) as a prerequisite: the server turns globalization on because Microsoft.Data.SqlClient needs it (Prerequisites). No code signing: the Windows executable is not Authenticode-signed, the macOS executable is not notarized, and the bundles are not signed with mcpb sign. No builds for other platforms (details).

When these notes were written, the packaging had been tried with local builds on a Windows machine with the .NET SDK installed and with one hand run of the release workflow on an earlier commit of main (version 0.0.0-dev), not with a tag run. Not yet shown then (details): starting the Linux build on Linux and the macOS build on a Mac, including whether Gatekeeper starts an executable that is not notarized; starting a build on a machine with no .NET installed at all; that the executable bit is kept in the .tar.gz archives and .mcpb bundles; installing a .mcpb bundle in any MCPB host other than the Microsoft Store build of Claude Desktop on Windows, where the win-x64 bundle was installed but a tool call typed in its chat was not shown.

Checked after the tag run (2026-11-13): the tag run passed the main check, attested the six files and created the draft release with seven files; a hand run of the release workflow on the tagged commit, just before the tag, had built the three archives and the three bundles (version 0.0.0-dev). Before the release was published, SHA256SUMS matched all six files, gh attestation verify exited 0 for each of them, the program of the win-x64 archive, started without configuration, exited 2 with its configuration error line, the linux-x64 archive lists the program with the executable bit, and the win-x64 bundle's manifest says version 0.1.0. Still not shown: the executable bit in the osx-arm64 archive and in the bundles, the Linux and macOS builds started on Linux and on a Mac, a start on a machine with no .NET, and a v0.1.0 bundle installed in any host (details).

The bundles are packed with the MCPB CLI pinned in packaging/mcpb/package-lock.json; their attestation shows that the release workflow built them, not that this tool left the server files unchanged. packaging/mcpb/package.json overrides two packages under the CLI, so the CLI runs with dependency versions its author did not pin: tmp 0.2.7 where its tree asks for 0.0.33, and zod 4.6.7 where it asks for 3.x, because no version in the ranges it asks for has the fix for their known vulnerabilities. With them, npm audit --package-lock-only on 2026-11-13 reported 0 vulnerabilities, and bundles packed on Windows with the old and the new lock file held the same files with the same content (details).

Check a download

Compare the file's SHA-256 with its line in SHA256SUMS (on Linux: sha256sum --check --ignore-missing SHA256SUMS in the download folder):

(Get-FileHash .\mssql-readonly-mcp-0.1.0-win-x64.zip -Algorithm SHA256).Hash.ToLowerInvariant()

The release workflow attests every file in SHA256SUMS (build provenance). The check below is meant for that attestation; when these notes were written, it had not been run against a real release. It needs the GitHub CLI (details):

gh attestation verify .\mssql-readonly-mcp-0.1.0-win-x64.zip --repo oguzhanpisgin/mssql-readonly-mcp

What it does

Promise, in short Full text
A write or DDL statement is refused before it reaches SQL Server, by an allowlist over the ScriptDom syntax tree, not a keyword filter. Writes refused before SQL Server
In every case tested, SQL Server ran the allowed text or refused it before running. Allowed text on a real SQL Server
The server refuses to start with a login that fails its permission check: exit code 3, no override. Login check at startup
No call returns data before its hash-chained audit line is written. If the line cannot be written, the query has still run and the call returns an error and no data. Audit line before data
Results are capped by rows, characters and time. Result caps
A query waits a bounded time for locks and loses deadlocks against default-priority sessions. Its own shared locks can still make a writer wait, until the call ends, and for a call ended as Timeout or Cancelled at its cutoff until its clean-up at SQL Server ends, which can be after the answer, and on a slow link a Timeout can leave them longer, with its session (U-1R). Lock timeout and deadlock priority

Three tools: list_tables, describe_table and run_query. Only run_query takes SQL, and that SQL goes through the gate (Tools).

Tested on

SQL Server 2019, 2022 and 2025, as Linux container images pinned by exact tag: mcr.microsoft.com/mssql/server:2019-CU32-GDR13-ubuntu-20.04, 2022-CU27-ubuntu-22.04 and 2025-CU9-GDR1-ubuntu-22.04. Other versions and Windows installations were not tested. In the last full run (2026-11-13, one run) all tests passed: Gate 387, Audit 86, and Integration 259 of 259 on each of the three images. One integration test, Truncated_result_whose_cancel_is_held_until_the_network_works_again_holds_its_locks_until_then_only, is not stable: it failed once in the full run of the earlier code, on the 2022 image (258 of 259), and in 4 of 19 runs on that image on earlier code, that run included; it is not fixed and can still fail in a CI run or a local run, and that it has the same cause as U-1R below is reasoned, not proven (Building and testing). CI runs on every push and pull request: the build and unit tests, then, when they pass, the integration tests on each of the three images.

What it does not solve, and known limits

Read What it does not solve before you install it. Five limits to know first:

Limit Where
The rows the assistant reads leave your building. They go to the AI vendor that runs the model, with the table and column names and any SQL Server error text. Read-only does not change that. README
The login check runs once, at start. Rights granted afterwards to the login, a role it is in or public are not seen until the server restarts. What check-login refuses
A call that has not ended 0.5 s after its time limit, its clean-up included, is answered as Timeout with no data (Cancelled, with no answer, if the client cancelled it first); in attack test run 4 every such Timeout of a single call came at most 0.56 s past the limit. A call that ends before its cutoff keeps its own outcome, so an ok answer with data can arrive up to 0.5 s after the limit, or later when the answer itself is delayed. Many calls frozen at once in the middle of a row delay the answers of every call, healthy ones too: in attack test ru...
Read more