Lost updates in four MCP servers: GitHub, Grafana, PostHog and Kubernetes
Miserable_Bed_219 · reddit · 2026-11-13
IsoKron audited write correctness in four widely used MCP servers, testing collisions against each platform's native API as a control:
- GitHub MCP v2.1.0: 19 of 30 concurrent field-pair edits lost; GitHub's own API lost 0 of 28. issuewrite merges and rewrites all fields with no version check; a fix PR is open.
- Grafana MCP v2.2.0: 30 of 30 lost on Grafana 13 and 11 — patch mode forces overwrite so the version check never runs. A PR saving via the Kubernetes-style API lost 0 of 60 on Grafana 13, still 14 of 30 on Grafana 11.
- PostHog MCP: 7 of 11 lost, but the root cause is PostHog's API itself (6 of 11 via direct API); a maintainer merged a fix 17 hours later. Cohorts and dashboards remain unfixed.
- Kubernetes MCP v0.1.2: server-side apply with force: true is an intentional upstream choice; stale re-applies overwrote changes 10 of 10 without resourceVersion, rejected with 409 10 of 10 with it.
The author shares a self-test recipe: run two processes editing different fields of one object 10+ times, read back, then repeat against the product API to isolate the layer. Limits: small counts, one machine — proves collisions lose writes, not how often agents collide.
More from coding & agent
- Sketchy Update: Claude Now Sketches Molecules and Circuits, Open Source — Nkgsporty · 2026-11-13
- Dev postmortem on Opus-reviewed fixes: ~1/3 accepted as-is, blind trust would've been horrible — Vjeux · 2026-11-13
- Descope's Envoy gives AI agents their own sign-in identities, compatible with Poppy — AiBreakfast · 2026-11-13
- Beaker makes production agents 37% faster and 35% cheaper, claims the model isn't the bottleneck — VictorSanh · 2026-11-13
- CMU trains LLMs to write memories with RL for continual learning across task streams — aviral_kumar2 · 2026-11-13
- Why AI agents need a new data stack: semantics, federation, governance — bigdata · 2026-11-13