One changelog for all your AI dev tools.

15 Sept

• API key creation governance controls at organization and project levels↗

  Organizations can now enforce API key policies at the org or project level—service-account-only, user-owned-only, or disabled—to lock down credential creation in production.

14 Sept

• Messages API can compact conversations on demand↗

  The API now lets you compress old conversation history into a signed block you can reuse, reducing token cost on long conversations while preserving recent context and thinking validity.

17 Sept

• Control MCP server startup wait time in headless mode↗

  Headless and CI runs can now control how long to wait for MCP servers to connect on the first turn, or skip waiting entirely with `0`.

• Sessions self-heal from corrupted transcripts↗

  Claude Code can now recover from corrupted session transcripts that caused endless retry loops, either by healing them automatically or offering a clear `/rewind` command.

• Slash commands work in Claude Desktop error hints↗

  Claude Desktop sessions now show appropriate help—slash commands instead of CLI flags that don't work in the desktop app.

• Memory pressure warning with recovery steps↗

  Claude Code now warns you when memory usage becomes critical and suggests how to recover, preventing silent failures during long agent runs.

15 Sept

• Permission checker now prompts for Bash it cannot fully analyze↗

  Fixed Bash commands the permission checker cannot fully analyze skipping the prompt under `permissions.blockReadsOutsideWorkingDirectories`, and a subshell hiding a dangerous `rm` in bypass mode.

• Context meter and auto-compact now count turns accurately↗

  The context meter and auto-compact were counting advisor-tool turns at roughly twice their real context size, causing auto-compact to fire at about half the real window. This is now fixed.

• Session forking from remote control now runs as background task↗

  You can now fork a session started with `claude --remote-control` or `/remote-control` from the Claude app; the fork runs as a background session on your computer, freeing up the app.

14 Sept

• Subagents can skip loading user and project CLAUDE.md files↗

  Subagents can now opt out of loading your CLAUDE.md files with `omitClaudeMd`, useful when a subagent should ignore your local config but still respect org policy.

• Fast mode now works in Claude Code Remote sessions↗

  Claude Code Remote sessions can now use fast mode if your organization permits it, applying the host's setting or the `/fast` command you type in the session.

• Fixed org policy cache persisting across account switches↗

  Switching accounts, organizations, or API keys mid-session now immediately refreshes the org policy instead of keeping the old one until the next hourly check.

• Sandboxed Bash commands now get per-command domain allowlists↗

  When sandboxing is on, Claude Code now analyzes each command to see which hosts it needs, opens access to only those, and blocks the rest—tighter than a blanket allowlist.

16 Sept

• Evaluation model IDs resolve through AI Gateway without default provider↗

  Evaluation can now resolve model identifiers through AI Gateway without requiring a default provider configured, and custom providers can use string aliases for models.

• Evaluation models can now be configured and resolved by ID↗

  You can now configure evaluation models in your provider and reference them by ID in experimental_evaluate, with explicit provider configuration required and no implicit fallback to Gateway.

• Tool search now works with direct tool calling↗

  You can now use tool search in models that support direct tool calling, expanding when you can filter large tool sets.

• Realtime WebSocket errors now report through onError callback↗

  When a realtime connection closes unexpectedly, you now receive diagnostics through your onError handler rather than losing visibility.

• Mid-conversation tool discovery and updates in code mode↗

  Code mode now supports discovering and updating tools during an active conversation, enabling dynamic tool availability without restarting.

• Deprecate rawInput in output-error UI message parts↗

  `rawInput` in output-error UI message parts is now deprecated and will be removed in the next major version.

• Experimental evaluation framework for structured outputs↗

  Experimental evaluation model lets you grade AI outputs as Choice, Score, or Boolean with typed answers, optional distributions, required probabilities, validation, and automatic retries.

15 Sept

• Realtime WebSocket runtime with continuous sessions and PCM playback↗

  The realtime WebSocket layer now supports continuous sessions with full audio playback, transcript tracking, and context management, replacing the previous turn-based model and removing autoContinueTools and managed Responses.

• OpenAI Live provider with server WebSocket and client delegation↗

  OpenAI's Live API is now available through a unified factory that handles server-side WebSocket connections with client delegation, letting you receive audio and transcript streams while your app controls agents and tools.

• Browser WebRTC option for Live conversations with client delegation↗

  Live conversations can now route through WebRTC directly from the browser instead of only WebSocket, letting you delegate audio handling to the client while the server manages permissions and session state.

• Tool output JSON values now serialize correctly↗

  Fixed serialization of JSON values in tool outputs, preventing data corruption when tools return structured data.

14 Sept

• Typed errors for AI SDK failures now exposed↗

  The SDK exports typed error classes for transport and completion failures, letting you catch and handle specific failure modes in your UI layer without string parsing.

16 Sept

• Hobby tier now auto-deletes old deployments more aggressively↗

  Hobby tier deployments now auto-delete more aggressively to stay within the 10GB storage limit. Only your 3 most recent production and 3 most recent deployments of any type are protected; preview deployments are no longer exempt. If you exceed the quota, old deployments delete immediately instead of after 30 days.

• Mem0 memory integration available in Vercel Marketplace↗

  Mem0 is now a native Vercel integration that automatically provisions scoped API keys and handles billing on your invoice, letting AI apps retain user context across sessions without separate account setup.

• Secure Compute and Static IP builds start 64% faster↗

  Builds using Secure Compute or Static IPs now reuse prewarmed containers instead of booting fresh ones, cutting build-start time by 64% automatically.

• Jev probabilistic decision model available on AI Gateway↗

  Jev is a decision model that answers typed questions about structured state in parallel, returning probabilities instead of generating text. It's available through AI Gateway and AI SDK 7.0.105+. Use it to automate clear cases and route uncertain ones for review without parsing LLM output.

• Is Agentic audit reports now filter by site type↗

  Is Agentic reports let you filter audit checks by your site type—Commerce highlights payment standards, App highlights APIs and auth—without rescanning or changing your score.

15 Sept

• Sandbox gives agents an isolated file system to check out code and run it↗

  Sandbox gives agents a sandboxed file system where they can clone your repo, run it, and pull in data needed to answer questions. Delphi uses it for an internal agent that helps customer success find root causes in Slack.

• Agents can read logs, inspect deployments, and change environment variables through SDK, MCP, or CLI↗

  Vercel exposed data and controls agents need—logs, deployment details, environment variables—across SDK, MCP, and CLI. Agents are now first-class users of these tools, not an afterthought.

• Python backends now run on Vercel with Workflows and Queues↗

  Vercel added Workflows for durable long-running work and Queues for background jobs, both provisioned automatically. This lets teams move Python backends from AWS without managing infrastructure themselves.

• AI Gateway routes requests to different models per customer and handles failover automatically↗

  AI Gateway lets you pick different models for different customers and automatically fails over to backups if a provider goes down. Delphi routes each digital mind to whichever model performs best for it.

• Gemini 3.8 Live models available through AI Gateway↗

  Google's Gemini 3.8 Live models—including Extended Thinking variant—are now available on Vercel's AI Gateway, enabling real-time voice interactions via the AI SDK's realtime API.

14 Sept

• Harness layer authenticates through native agent subscriptions↗

  The harness layer can now use subscription credentials built into coding agents like Cursor and Cline, keeping secrets on the host and injecting tokens into requests rather than passing them to the sandbox.