Research
What we learn auditing and building MCP servers. We publish a piece only when it carries original testing, a public finding or a working implementation, and each one states what it was tested against.
Start here
Start here if you build MCP servers
Start here if you are adopting one
More research, newest first
RSS feedStudyWrite correctness
Conflict errors and an AI agent: what a version check stops, and when the error text matters
Told not to re-read, the agent lost the other writer's change in 38 of 48 runs when the conflict error listed the state unquoted, and in 0 of 48 when it was quoted.
29 min read
StudyWrite correctness
Do popular packaged MCP servers say whether a write happened?
Of 100 popular packaged MCP servers, read from source, 7 can say in a machine-readable form that a write's outcome is unknown.
25 min read
Public auditWrite correctness
Payload MCP plugin 3.93.0: updates and deletes write through another editor's document lock
Payload's MCP update tool wrote through another editor's document lock in 10 of 10 trials; Payload's REST API refused all 10 with HTTP 423.
43 min read
StudyWrite correctness
Send the change, not the object
Changing one issue field through GitHub's MCP server lost a concurrent change in 19 of 30 pairs; GitHub's per-field update lost 0 of 28.
16 min read
Public auditWrite correctness
Kubernetes MCP server v0.1.2: Helm installs ignore the namespace argument, and some multi-document applies drop documents
Helm installs through the server put the chart’s objects in the wrong namespace in 3 of 3 runs, where the Helm CLI as the same identity got it right.
32 min read
StudyWrite correctness
Lost updates in four MCP servers: GitHub, Grafana, PostHog and Kubernetes
In all four MCP servers we audited, one write could undo another writer's change, each for a different reason.
8 min read
Public auditWrite correctness
PostHog MCP server: concurrent updates could silently lose a write in PostHog's own API
Concurrent updates to one insight lost a write in 7 of 11 pairs, and the bug was in PostHog’s API, which PostHog has fixed.
42 min read
Public auditWrite correctness
GitHub MCP server v2.1.0: concurrent issue-field updates can silently lose changes
Two agents editing different fields of one issue lost a change in 19 of 30 pairs, while both calls reported success.
27 min read
GuideTesting and benchmarks
How to test an MCP server without fooling yourself
The rules our audits run under, what review caught in our own tooling, and a 10-step checklist for testing your own server.
11 min read
GuideTesting and benchmarks
Tamper-evident audit evidence: how to verify an MCP audit report yourself
Check an IsoKron Evidence Pack on your own machine with one command, against timestamps from two independent authorities.
18 min read
EssayTool design and efficiency
Why MCP tool design matters for agent reliability
In MCP-Universe, connecting 7 servers instead of only the ones a task needed halved one model’s success on location tasks.
10 min read
Sample reportTesting and benchmarks
GitHub MCP server v2.1.0: sample audit report
Three of the 17 findings from our public GitHub audit, laid out the way a client report is.
GuideProduction operations
Shipping a production MCP server on the 2026-07-28 spec: auth, rate limits, and observability
Once a second person calls your server, identity, per-user limits and tracing must come from each request, on the 2026-07-28 spec.
9 min read
Every piece is by Joshua Bryan, IsoKron and ends with how it was tested and which checks its findings passed. What we file publicly and what we report privately follows our disclosure policy. Each article is also available as Markdown for agents: add .md to its address.