Anthropic open-sources oncall-kit: Claude handles on-call duty, locating incidents in 4 minutes, with 80% of integration code AI-written

Forum topic · QianXun · 2026-08-23

Summary

On August 22, Anthropic engineer Sachin Malhotra detailed an internal system, described as a 'cyber on-call engineer' (Claude Tag), and open-sourced the methodology as the oncall-kit project. The system runs as a resident Slack session connected to alert sources from Grafana, Datadog, and PagerDuty. When a failure signal arrives, it reads the alert in natural language, performs cross-source correlation analysis to pinpoint the specific service, commit, and owner, and pushes an incident report with fix suggestions — compressing a traditional on-call process that typically takes two hours down to minutes: as fast as 4 minutes to locate a fault and 14 minutes for a full situational report. Notably, over 80% of the code merged into oncall-kit's main branch was written by Claude itself, with engineers only reviewing and approving. The post argues this marks a shift in AI coding competition from model capability toward deployment-stack depth, though the system remains limited on novel, unknown failure modes. It also situates oncall-kit amid rising Codex usage and suggests open-sourcing builds a data feedback loop from real production incidents.

Anthropic Open-Sources oncall-kit: When Claude Takes the On-Call Shift

On August 22, Anthropic engineer Sachin Malhotra publicly detailed an internal system called the "cyber on-call engineer" (Claude Tag), covered in a report by 新智元. The system can locate faults in as fast as 4 minutes and produce a situational report in 14 minutes. More than 80% of the connected code was written by Claude itself. The full methodology has been open-sourced as the oncall-kit project.

What the system does

This is not a single-point demo. Claude Tag lives in Slack as a persistent session, wired into alert sources from infrastructure tools like Grafana, Datadog, and PagerDuty. The workflow:

1. Reads and interprets incoming alerts in natural language 2. Performs cross-source correlation analysis 3. Pinpoints the specific service, specific commit, and responsible owner 4. Pushes a situational report with fix recommendations to the on-call engineer

This compresses a process that traditionally takes an on-call engineer two-plus hours down to minutes.

80% self-produced code

The more notable stat is that over 80% of code merged into oncall-kit's main branch was generated by Claude, with the engineering team only responsible for critique and sign-off. Anthropic isn't just using Claude to write new features — it has pushed Claude into the operations loop, letting it modify its own deployment stack. Combined with the Claude Code 2.1.234 refresh on August 21 (browser-use + Skills API + Files API), this forms a closed loop: the frontend "writes code," the backend "takes the on-call shift."

Why on-call is the new moat

On-call duty pushes AI coding from "personal productivity tool" to "enterprise infrastructure component." For two years, AI coding companies competed on generating more code at the PR stage — the "writing" track. But operations after production is a long tail: mid-sized internet companies typically spend 3–5x development costs on operations. Anthropic's move effectively declares: "Our model not only writes code, it can shoulder operational responsibility — and it wrote the ops tooling itself."

The competitive picture

The same day, Codex weekly active users broke 20 million (confirmed by Tibo on X, August 22). The Claude Code vs. Codex contest used to be about who writes better complex code; the framework has now shifted to who can embed AI deeper into enterprise production pipelines. Per Alphabet AI's August 22 report, Codex grew 20.8% over four weeks versus Claude Code's 5.2%, with the user-base gap narrowing from several-fold at the start of the year to 1.7x. Open-sourcing oncall-kit partly reflects this pressure: with user growth slowing, opening engineering culture and technical assets to the community is more realistic than guarding differentiation behind closed doors.

Don't over-romanticize

The 4-minute localization has boundaries. oncall-kit works well on known error patterns (CPU saturation, memory leaks, dependency timeouts), but for novel, unknown failures — an unseen kernel panic or cloud-provider infrastructure jitter — it still relies on a human to make the final call. Anthropic's own engineers stress that "Claude is a highly capable copilot, but does not vouch for reliability."

The deeper play: a data feedback loop

Traditional AI coding vendors lack real production failure samples — how code behaves in production, how it's fixed, whether rollbacks are clean — data locked inside enterprise monitoring systems. By open-sourcing, Anthropic trades for community-contributed incident samples: every enterprise adopting oncall-kit becomes its "training ground." It's an early sketch of a business model shifting from "selling model tokens" toward "selling an operations operating system."

The takeaway

The real AI coding news of August 22 isn't Codex's user count — it's Anthropic putting Claude on on-call duty. That same week, Google Antigravity Anywhere shipped remote agent capability (Aug 21) and OpenAI pledged not to retain enterprise data. All three top vendors simultaneously moved the battlefield from the model layer to the deployment layer — a marginal shift more significant than any single new model release.