Skip to content
IsoKron

New: MCP Write-Safety Requirements 0.9 and two studies

When AI agents write through your MCP server, changes can silently disappear. We find and fix that.

We audit MCP servers for security and write correctness, then rebuild their write paths or build new ones in your repository.

19 of 30pairs lost a change through the MCP server

0 of 28lost through GitHub’s own per-field API

Figure 1. GitHub MCP server v2.1.0, issue #3751: two agents edit different fields of one issue at once, and both calls report success. Three runs.

You ship an MCP server

We find the writes that silently go wrong before your customers’ agents do.

You’re adopting one

Before agents write through someone else’s server, find out what it gets wrong.

What we measured on three public MCP servers

19 of 30

pairs of concurrent issue_write calls on different fields of one issue lost one change while both calls returned success (GitHub MCP server v2.1.0). GitHub’s per-field update, the control, lost 0 of 28 pairs.

GitHub audit

30 of 30

pairs of simultaneous update_dashboard patches lost an edit while both calls returned success (Grafana MCP server v2.2.0). The control is not clean: two simultaneous saves straight to Grafana’s API also lost an edit, in 30 of 30 pairs on Grafana 13 and 20 of 30 on Grafana 11.

Grafana audit

10 of 10

trials in which Payload’s MCP update tool wrote through another user’s document lock (@payloadcms/plugin-mcp 3.93.0). Payload’s REST API refused all 10 with HTTP 423.

Payload audit

Each figure is from the linked audit, which gives its versions, its n and what was measured. MWS is an open spec, draft 0.9, in public comment. All five audits, with upstream status

What our public research found, as of

7 of 100

popular MCP servers with write tools, sampled from the most-downloaded packaged servers in the official registry, can return a machine-readable “unknown” when a write’s outcome isn’t known. Read, not run.

100-server study

Merged

PostHog’s fix for concurrent insight updates that lost a write in 7 of 11 pairs, co-authored by IsoKron

Pull request #131620

0 of 60

pairs lost an edit with our proposed Grafana fix on Grafana 13, against 30 of 30 unpatched

Pull request #1518 (open)

Research on public MCP servers, not client engagements. Every issue and pull request we filed, with its status

Public research on the GitHub, PostHog, Grafana, Kubernetes and Payload MCP servers

We test whether your MCP server’s writes actually land. We test the same way in public: independent audits of GitHub, Grafana, PostHog, Kubernetes and Payload, with findings, reproductions and upstream status published.

Why it matters

Three things we found in public MCP servers, and what each means for a team whose agents write through one.
  • Lost updates

    Two agents edit one record. Both are told “success”. One change is gone.

    In GitHub’s, Grafana’s and PostHog’s MCP servers, one of two simultaneous writes to one object could be lost while both calls reported success.
  • Safety switches

    A read-only switch that silently stays off.

    In GitHub’s MCP server, GITHUB_READ_ONLY=yes is read as false, so the server starts with write tools on and no warning.
  • Partial writes

    An apply that writes one document and reports success.

    In the Kubernetes MCP server, a multi-document apply with a ---  or --- # comment separator wrote only the first document and reported success, in 3 of 3 runs for each shape. kubectl applied every document.

Write-safe MCP servers, at stated prices

Three steps: test, fix, keep

Test what your server’s writes do, fix what the tests find, and keep it checked as the server changes. The listed price is the price.
  1. Step 1

    Test

    Find out what your server’s writes do.

    • MCP Server CheckerFree

      Paste your server’s tools/list answer and get a grade for each area. It runs in your browser; nothing is stored or sent.

    • Writecheck readout$450

      A written reading of every verdict in your mcp-writecheck run and the top three fixes in order. No call needed.

    • Writecheck setup$950

      We add the mcp-writecheck GitHub Action to your repository as a pull request, with the first signed full-mode record. No call needed.

    • Write-Safety Audit$2,500

      We test your server’s write paths on a test tenant, with a security review, fix pull requests for up to five of the highest-severity findings and one retest.

  2. Step 2

    Fix

    Repair the write paths, or build the server new.

    • Write-path rebuild$6,000

      We rebuild the write paths and tool design of a server you already have, in its own repository, with a before/after benchmark.

    • New MCP serverfrom $7,500

      A new server for one API, one sign-in method and up to 10 tools. The price is fixed in the blueprint, which is part of it.

  3. Step 3

    Keep

    Keep it checked as the server changes.

    • Per-release regression check$750 per month per server

      We re-run the tests against up to four releases a month and report what changed. Three-month minimum.

    • Advisory retainer$2,000 per month

      Two 60-minute sessions a month, written answers in between, and reviews of your designs and pull requests.

On request, or on its own page: Design-partner auditDesign reviewMCP blueprintMore tools or a second serverDirectory submission packLarger audits, rebuilds and buildsSecurity firms and agencies

The Write-Safety Audit leads with the write paths and also covers a security review, the tool catalog and its token cost, and fixes, a retest and an Evidence Pack. The security review is not a penetration test of the platform behind the server or of its OAuth provider. Servers behind OAuth sign-in are in the fixed price; testing two accounts or tenants against each other is not available yet. A second server in the same engagement is +$1,750, the per-task agent benchmark is quoted on request, and testing never touches production data. What the audit covers

How we compare

How every finding in a client report is verified

  1. Step 1: Checked against a control

    Where the vendor has its own API, the same operation runs through it, so we know whether the bug is in the MCP server or in the platform behind it. For the lost-update finding, GitHub’s API lost 0 of 28.
  2. Step 2: Challenged by a separate AI refuter

    A separate AI, in a clean context, tries to explain the finding away. If it can, we drop it.
  3. Step 3: Reproduced and reviewed by a person

    A person reruns the reproduction and reads the evidence before it reaches the report.

Evidence: A finding we dropped

search_code found nothing on a private repo, and neither did GitHub’s own API, so we never reported it.

Our public audits say which of these checks their findings went through, and where a check fell short. More on the method

What you receive

Sample report

One entry per finding

Impact, reproduction, control, refuter result, confidence and a suggested fix.

Evidence Pack

Check the results yourself

An IsoKron Evidence Pack is the signed, timestamped record of an audit. You can check it on your own machine with one command, without trusting IsoKron’s servers.

Request an audit or a build

Tell us which server and what you need. We reply to confirm the scope, the price and the dates.