Uber cut AI costs with MCP servers

Hi there, this is your daily ☕️ Devshot.

In today's Devshot:

🔧 Uber cut AI costs with MCP servers

🌐 Microsoft open-sources TauGrid

🔀 Zed launches Delta

📊 TanStack Charts comes to TypeScript

⚡ C++23's flat_map beats std map on lookups

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

🔧 Uber cut AI costs with MCP servers LINK

Uber cut its per-session AI tool-schema costs to near-zero by routing over 1,000 MCP server interactions through a single gateway, dropping cost per 1,000 requests ~34% and cost per session ~52% from peak while keeping spend flat.

The problem was schema overhead: with ~100 tools installed, definitions added 50,000-70,000 tokens to every prompt and got re-sent on each turn, since vendors bundle huge tool sets, one suite ships 49 tools needing ~22K tokens.

Three fixes attacked it: CLI resolution lets the model invoke tools via shell without loading schemas, on-demand tool search pulls only relevant definitions, and code-mode batches sequences into one subprocess script, packaged as ~25 default "skills".

🌐 Microsoft open-sources TauGrid LINK

Microsoft has open-sourced TauGrid, a cloud-native platform for managing, scheduling, and monitoring AI workloads on GPU-enabled Kubernetes clusters, letting researchers submit jobs without touching Kubernetes while platform teams handle queues, quotas, and observability.

Written in Go, it defines workloads in a YAML file submitted with `tau run`, which validates the config and creates a Kubernetes Job or KubeRay RayJob, queued via Kueue by remaining quota and priority.

TauGrid tracks status, logs, and checkpoints so failed jobs can resume, requires a cluster running 1.30+ with GPU nodes and Helm 3.0, and its roadmap includes PyTorch DDP/FSDP, DeepSpeed, LoRA/QLoRA, and multi-cloud execution.

🔀 Zed launches Delta LINK

Zed has launched the public beta of Delta, a collaborative environment where developers and coding agents work, review and revise code together in shared threads instead of pull requests, available on macOS, Linux, Windows and browser.

Delta's core unit is the thread, a running record linking the agent conversation to the files it edits; DeltaDB underneath captures individual edits and chat events as "deltas," adding history between Git commits rather than replacing branches.

Zed has already disabled pull requests on Delta's own repo, where 33 developers landed 570 changes to main through threads; it still works with Git repositories and plans repository-based access using GitHub permissions.

📊 TanStack Charts comes to TypeScript LINK

TanStack Charts has landed as a framework-agnostic, type-safe visualization library for TypeScript, already hitting roughly 160,000 weekly downloads despite still sitting officially in Alpha with no stable API promise.

Instead of fixed chart types, you compose marks, scales, channels, transforms, and layers via a grammar-of-graphics API, with inferred types staying tied to your source rows and adapters for React, Vue, Solid, Svelte, Angular, Lit, and more.

A basic React line chart weighs about 29 kB gzipped and the suite sits around 38-45 KiB, undercutting Chart.js, Observable Plot, and ECharts, though the docs warn it suits simpler charts over heavy real-time datasets.

⚡ C++23's flat_map beats std map on lookups LINK

C++23's `std::flat_map`, available in GCC 15's libstdc++ and LLVM 20's libc++, stores sorted key and value vectors queried by binary search, and benchmarks show it often beats `std::map` for lookups and bulk construction.

On random lookups over 1M elements it ran ~150ns versus ~390ns for `std::map`, and range construction hit ~72ns against ~316ns, since two contiguous arrays cut memory overhead and improve locality.

The catch is mutation: inserting millions of random keys one at a time scales quadratically and becomes far worse than `std::map`, though sorted-order inserts, `insert_range` bulk merges, and disk serialization via `memcpy` stay fast.

Other news you might like

• Migrating the GitHub Copilot runtime to Rust, using CopilotLINK

• Developer uses GPT-6 Astra to get CoD Black Ops 2 Hijacked map running natively inside Minecraft — achieves 45fps performance using Minecraft’s OpenGL contextLINK

• Google Home MCP lets Antigravity, Claude, OpenClaw, & more control your smart homeLINK

• Not just Proton: Getting to know Valve's new SteamOS compatibility layersLINK

• Dropbox Evolves Riviera Content Processing Platform to Support AI WorkloadsLINK

• Kubernetes v1.37: Hardening Container Storage with Bind Mount Options and EmptyDir PermissionsLINK

🧰 Trending tools

Prelint: reviews AI-generated pull requests against your ADRs, docs, and past decisions, catching product drift before it reaches productionLINK

Switch: open-source collaborative workspace uniting people and AI agents in one shared room across Slack, Teams, Discord, or Mattermost, preserving context through every handoff.LINK

Mastra Factory: a TypeScript framework for building AI agents and apps, with workflows, memory, streaming, evals, tracing, and an interactive Studio UI for testing.LINK

Anysite.io: pulls fresh web and business data for outbound, ABM, research, and monitoring via MCP or REST, with flat monthly plansLINK

Meridian: open-source, locally-run activity tracker that logs your work with context, revealing where your time went and the blockers behind it.LINK

AgentSky: runs managed AI coding agents with auto-recovery, controlling remote development from WhatsApp, Telegram, Slack, web, API, or CLILINK

📚 Trending papers & reports

Factory controller testing gets its first shared benchmark, 50 programs in 83 variants with verified pass or fail answers, so companies can finally prove their safety-checking tools actually catch dangerous flaws in industrial machine code.LINK

Code-checked reasoning training teaches AI to run and verify its own math and physics work, boosting accuracy up to ~17 points and cutting scientific-law errors by ~73%, beating leading models on physics tests.LINK

AgentPack is a cleaner training set of 1.8 million code edits jointly made by humans and coding assistants like Claude Code and Cursor, and it teaches AI code editors better than older human-only data.LINK

AI-built Linux utilities proved as reliable or more reliable than their human-written originals under standard crash testing, with fewer memory bugs but more freeze-ups, though skilled prompting and human oversight remained essential.LINK

Agent-friendly design system gives websites a shared layout that both people and software agents can read clearly, letting bots complete tasks more reliably while keeping the human interface visually flexible and familiar.LINK

See you tomorrow for a new dose of ☕️ Devshot!