Grok audio on Vercel, Go 1.26 green GC, voice APIs — Dev Signal
Grok audio on Vercel, Go 1.26 green GC, voice APIs
Share:
Tool of the Week
xAI Grok audio models ship on Vercel AI Gateway
Realtime voice, TTS, and STT from xAI route through Vercel's gateway with token-based client auth and unified observability—integrate via AI SDK 7.
Eliminates API key exposure on client and consolidates audio infrastructure into existing gateway spend controls. Realtime voice agents now require only server token endpoint and browser hook.
Replaces direct xAI audio integration with routed, metered access. Requires AI SDK 7, a token endpoint, and `useRealtime` hook. Ready now—playground test available before committing.
“xAI's audio models are now live on AI Gateway”
“Realtime voice, text to speech, and speech to text are all available through the AI SDK with the same routing, observability, and spend controls as your other models”
“These capabilities are available on the AI SDK 7 release”
“A voice agent has two pieces: a server route that mints a short-lived token, so your API key never reaches the client, and a browser component that connects with it”
Get issues like this in your inbox — free, every weekday.
Quick Signals
AI Gateway adds realtime voice and speech support
Realtime voice agents skip the speech-to-text→LLM→text-to-speech pipeline; single model handles bidirectional audio with server-side VAD and mid-conversation tool calls.
Developers building voice interfaces can now route audio through the same Gateway infrastructure as text/images, with unified routing, spend controls, and observability. Eliminates latency overhead of chaining three separate models for conversation.
Ready now in AI SDK 7 beta. Replaces custom WebSocket + three-model chains with `useRealtime` hook and token-based auth. Requires microphone access in browser; production-ready for OpenAI GPT-4 Realtime and xAI Grok TTS. Worth trying if already using AI Gateway.
“Audio launches with models from OpenAI and xAI”
“These capabilities are in beta and available in AI SDK 7”
“What sets it apart from chaining models together is that a single realtime model hears audio and produces audio directly, instead of running a speech-to-text, then language model, then text-to-speech pipeline”
“turnDetection: { type: 'server-vad' } lets the server decide when the user has stopped speaking, and lets the user talk over the model to cut a reply short (barge-in), with no client-side silence timers”
JetBrains Air lands Windows agent-first development
Dedicated agent orchestration environment now runs on Windows x64/ARM64, letting you parallelize agent tasks in isolated Git worktrees and review cross-agent diffs without CLI context-switching.
Data Point
LLM security scans repeat inconsistently without reference baseline
Claude finds 75.4% of Snyk Code vulnerabilities with high variance on extra findings; nearly 50% of LLM-only reports appear once in five identical runs, but reference-matched findings stay stable at 80%+ consistency.
If your CI/CD uses LLMs for security review before human code inspection, unrepeatable findings create noise in diffs and false confidence in coverage. You need to know whether the agent will flag the same issue twice or miss it on the next scan.
Do not replace SAST with LLM security review. Combine them. LLMs reach 75.4% F1 against Snyk Code's reference set with 24.6-point gap to deterministic baseline; unmatched findings are too noisy for solo use. Keep SAST deterministic, use Claude for unfamiliar patterns and prose risk explanation. Benchmark requires small Express apps—test your own codebase before committing LLM review to gate checks.
“The highest-recall LLM configuration found only 81% of Snyk Code reference vulnerabilities”
“Nearly 50% of LLM-only vulnerability reports appeared in just 1 of 5 identical scans”
“80 of 161 unique-unmatched findings appeared in only one of five identical repetitions, while only 22 appeared in all five”
“134 of 158 unique reference-matched findings appeared in all five repetitions”
“The best-scoring LLM configuration reached 75.4% Snyk-reference F1, leaving a 24.6-point gap against deterministic SAST reference reproduction”
3 issues a week · Free forever · 4,200+ developers
Eliminates friction between terminal-based agent execution and full IDE workflows. Native Windows support expands agentic dev patterns—parallel agent runs, multi-agent review chains—beyond macOS-only users, directly into daily coding loops.
Replaces ad-hoc terminal agent calls and manual worktree management. Requires Windows x64 or ARM64; agent API keys (Claude, Codex, etc.); Git repo. Ready now—JetBrains shipped stability hardening specifically for Windows day-one reliability.
“Air is built for agent-agnostic development – you can use leading AI coding agents to implement features, fix bugs, investigate code, generate changes, and review results in a dedicated agent-first development environment”
“The Air desktop app fills the gap between running agents in the terminal and working in a full IDE”
“Run agents simultaneously in separate Git worktrees. Each agent gets its own working directory and branch, so their changes stay independent and don't conflict”
“Over the past few months, we invested heavily in app stability”
AI Gateway adds realtime voice and transcription support
Voice agents, text-to-speech, and transcription now available in AI Gateway with the same observability and spend controls as text models, via AI SDK 7.
Eliminates the need to chain separate models for voice interactions—a single realtime model handles audio in and out with low-latency response, reducing implementation complexity and latency for conversational agents. Observability and spend controls apply uniformly across modalities.
Replaces custom WebSocket plumbing and separate TTS/STT service integrations. Requires AI SDK 7 and a token route for client-side security. Beta status means API stability isn't guaranteed. Worth prototyping now if you need low-latency voice—the playground lets you test without code first.
“Realtime support, a single model takes audio in and audio out, so a user can talk and hear a reply back in near real time”
“same observability, spend controls, and bring-your-own-key support as text, image, and video models in AI Gateway, with no markup or platform fees”
“available via AI SDK 7”
“The useRealtime hook handles microphone capture and playback”
GitLab Orbit MCP brings lifecycle graph to Antigravity
Query your GitLab instance's dependency graph, vulnerabilities, and ownership directly from Antigravity agents via Model Context Protocol—no copy-pasting between tools.
Agents gain structured access to DevSecOps context (dependencies, vulnerabilities, past reviews, owners) instead of working blind on files and terminals alone. Early tests show 11x faster responses, 4.5x fewer tokens, 45x fewer hallucinations.
Replaces manual context gathering and custom scripts for agents querying your software lifecycle. Requires GitLab Premium/Ultimate tier and Antigravity access; install directly from MCP Store in settings with one click. Ready now—no config files needed. Queries consume GitLab Credits.
“agents grounded with GitLab Orbit responded up to 11 times faster, used up to 4.5 times fewer tokens, and produced up to 45 times fewer hallucinations”
“GitLab Orbit indexes your GitLab instance and builds a knowledge graph of relationships between groups, projects, users, work items, merge requests, pipelines, vulnerabilities, and source code”
“Antigravity agents, without GitLab Orbit, can see the files and reach the terminal. They do not understand the broader system”
“Open the MCP Store panel from the settings. Within the customization tab, find the MCP section. Click "Add MCP" and add GitLab Orbit”
mcpagentsdevops-contextgitlabantigravity
Go 1.26 enables green tea GC by default
New `new()` syntax accepts expressions for initialization, generic types reference themselves, and Green Tea garbage collector replaces the previous default with ~30% cgo overhead reduction.
Self-referential generics simplify complex data structure implementations, stack-allocated slice backing stores reduce heap pressure, and the rewritten `go fix` command with two dozen modernizers automates adoption of language features—direct quality-of-life wins for daily development.
Green Tea GC is opt-out (enabled by default), so upgrade immediately to benefit. New syntax is backward-compatible. Experimental packages (simd/archsimd, runtime/secret, goroutineleak profiling) require explicit opt-in—worth testing now if you use SIMD or cryptographic code, but not production-ready. Worth upgrading for the tooling and perf wins alone.
“the built-in new function, which creates a new variable, now allows its operand to be an expression, specifying the initial value of the variable”
“generic types may now refer to themselves in their own type parameter list”
“The previously experimental Green Tea garbage collector is now enabled by default”
“The baseline cgo overhead has been reduced by approximately 30%”
“The compiler can now allocate the backing store for slices on the stack in more situations”