Uber runs Git as a service

Hi there, this is your daily ☕️ Devshot.

In today's Devshot:

🐙 Uber runs Git as a service

❌ OpenAI cuts off Cursor

📉 Anthropic's Claude Code raise is a cut

🦞 OpenClaw 2.0 adds multiplayer AI agents

🌐 Cloudflare Workers now accept inbound TCP

🔐 Kubernetes v1.37 adds pod certificates

Plus: 🎁 6 other news you might like, 🧰 6 tools, and 📚 5 papers.

🐙 Uber runs Git as a service LINK

Uber built GitFarm, a Git-as-a-service platform that runs repository operations for its large monorepos on a shared backend, eliminating local clones on client systems and cutting client-side resource use by more than 80%.

Instead of full checkouts that take ~15 minutes and six CPU cores per service, GitFarm exposes Git through a gRPC API, mounting pre-warmed checkouts into ephemeral sandboxes to deliver a ready checkout in under one second.

A code ownership service dropped from 70-plus cores to 16 and startup from ~15 minutes to under one, while a compliance auditor cut median latency from 110-160 seconds on Buildkite to 20-30 seconds; live since early 2025.

❌ OpenAI cuts off Cursor LINK

OpenAI will terminate its contract with Cursor effective November 12, 2026, after SpaceX's acquisition of the AI coding tool triggered a change-of-ownership clause, citing an inability to trust that SpaceX will honor the terms of service.

Developers using GPT models inside Cursor will keep access through their own OpenAI API keys, and OpenAI will continue supporting its IDE extensions for Cursor, so the move targets the ownership change rather than the tool itself.

Cursor co-founder Michael Truell said OpenAI models account for only about 5% of the tool's AI traffic and that talks with OpenAI are underway to resolve the situation.

📉 Anthropic's Claude Code raise is a cut LINK

Anthropic is trimming Claude Code's weekly usage caps by about 17% starting September 14, since the permanent 25% baseline bump lands below the temporary 50% boost users currently have active.

The 50% boost stays live only until September 14, after which Pro, Max, Team, and Enterprise plans settle at the raised baseline, leaving less capacity than developers get today despite the confirmed increase over original limits.

Recent Claude Code updates include performance optimizations and an automatic feedback tool that generates a reviewable bug report when something fails, optionally attaching the session's conversation log, with the feature toggleable off in settings.

🦞 OpenClaw 2.0 adds multiplayer AI agents LINK

OpenClaw 2.0 has shipped, the platform's largest release with over 16,000 pull requests, adding shared cloud sessions for team collaboration, a simplified first-run setup, and a browser app rebuilt from scratch.

The new onboarding detects existing resources on your machine, including ChatGPT or Claude subscriptions, API keys, and local models, cutting many configuration steps, while the rebuilt browser app adds a "Session Rail" status display and an integrated companion thread for questions.

Sessions run in three places: the local gateway by default, your own hardware via Paired Devices, or disposable rented machines through the Crabbox tool backed by AWS or Hetzner, with provider credentials always staying on the gateway.

🌐 Cloudflare Workers now accept inbound TCP LINK

Cloudflare Workers can now accept inbound TCP connections through a new connect(socket) handler, lifting an HTTP-only server restriction that stood since 2017 and adding gRPC support as the first feature built on top.

Raw sockets route in via Spectrum and can pass from Worker to Durable Object to Container through getTcpPort(), where full-duplex works for any TCP protocol, running unmodified Go and Python gRPC servers as examples.

Workers alone handle only unary and server-streaming gRPC by translating to gRPC-web, not bidirectional streaming, and the whole release ships as private beta behind a signup form since Cloudflare doesn't use gRPC internally.

🔐 Kubernetes v1.37 adds pod certificates LINK

Kubernetes 1.37 promoted Pod Certificates and Cluster Trust Bundles to GA, building X.509 certificate issuance for TLS and mTLS directly into core so workloads get proof-of-possession credentials instead of relying solely on bearer service account JWTs.

Kubelet generates a private key per pod, files a PodCertificateRequest to a named signer controller, then writes the key, certificate chain, and trust anchors to the container filesystem, with automatic rotation that apps must track via inotify or polling.

The mechanism is pluggable so a cluster can run many signer types at once; any core-shipped signers will cap certificate lifetime at 24 hours (91 days for others), but no signer ships yet, so testing requires a third-party one like Tinycert.

Other news you might like

• The new Go JSON API: twice as fast, or 1.5x slower?LINK

• Debian Linux developers just rejected a ban on artificial intelligenceLINK

• AWS Open Sources Kiro Crew for Asynchronous Coding AgentsLINK

• Donkey Kong 64 finally gets a fully native PC port written in C — DK64 ReKONGpiled brings ultrawide support, uncapped framerates, and zero AI codeLINK

• California lawmakers unanimously pass Linux exemption from age-verification law — software distributed under the GPL, MIT, BSD, and Apache licenses are exemptLINK

• OpenTelemetry has graduated… now what?LINK

🧰 Trending tools

AnySearch: queries multiple trusted sources in parallel for AI agents, returning filtered, de-duplicated results as structured data ready to useLINK

SKI: adds voice control to AI coding agents like Claude Code and Codex, letting you dictate commands and hear responses hands-free while codingLINK

ZooData: converts webpages into structured JSON to cut LLM token usage by 75%, with pre-analyzed Amazon and TikTok e-commerce data for AI agents.LINK

CartAI: automates checkout across merchant sites through one API, handling payments, shipping, and order tracking without building per-site integrations.LINK

ditto.site: converts any public URL into clean Next.js or Vite code, extracting components, design tokens, fonts, and hover states via a free API.LINK

Replay QA: records browser sessions with full devtools access, turning them into shareable bug reports so AI-built apps get QA coverage without writing test suites.LINK

📚 Trending papers & reports

Self-improving security fixer lets an AI repair software vulnerabilities by remembering lessons from past fixes, patching correctly ~90% of the time and beating the strongest prior tool by roughly 34 to 40 points.LINK

Agent grading replaces pass-fail scoring of AI agents with partial credit that captures how far a task actually got, matching human judgment better and telling successes from failures more reliably than holistic scoring.LINK

Image search stress-testing maps 50 image-tweaking tricks into ten categories and finds that weather effects and one realism-preserving method best expose when visual search engines return wrong matches, without looking fake.LINK

Prompt scoring shows that measurable writing traits like sentence length and structure can predict how well an AI will perform before you run it, letting teams pick better prompts without costly trial-and-error calls.LINK

AI compliance scorecard turns dense regulations like the EU AI Act and GDPR into a practical checklist that automatically rules out unsafe language models before teams adopt them, tested against 20 attack scenarios.LINK

See you tomorrow for a new dose of ☕️ Devshot!