Skip to content
Agents tracked: 258 Downloads (7d): 219M up 6.1% GitHub stars: 5.5M VS Code installs: 148M Releases (7d): 327 Agent status: 1 with issues Updated Oct 7, 2026
AgentGid

CopilotKit changelog: what's new each month

Every stable CopilotKit release summarised by month: the highlights, new features, improvements, fixes and anything you need to act on. 3 months covered; the current month updates daily.

October 2026 · so far

2 releases: 1.76.0 → 1.77.0

October brought significant UI improvements and stability enhancements to CopilotKit. The month focused on refining the inspector experience, improving tool handling, and expanding configurability.

Highlights

  • AG-UI 1.0 released for CopilotKit, marking a major version milestone for the UI system.
  • Inspector now scales properly on small screens and offers an option to hide it indefinitely.
  • Intelligence indicator can now be configured to auto-mount or not, giving more control over default behavior.
  • MCP tools with colliding names are now prefixed instead of dropped, preventing tool loss.
  • Browser activity is now captured as AG-UI events for better learning and debugging.

New

  • Option to hide the web inspector indefinitely.
  • Configurability for intelligence indicator auto-mounting.
  • Raw browser activity capture as AG-UI events for the learning system.
  • Consolidated Vue and Angular hosts onto the shared MCP apps package.

Improved

  • Web inspector now fits and scales correctly on small screens.
  • MCP tools with name collisions are prefixed rather than dropped.
  • LangGraph runs now properly cancel when a stop signal is sent.
  • Generated controls no longer accidentally submit host forms.
  • Popup rounded corners are now preserved in React and Vue.
  • Malformed component values are now contained instead of breaking rendering.

Fixed

  • Provider IDs are no longer reused as message IDs in per-response contexts.
  • LangGraph stop cancellation now works correctly after dependency update.
  • Generated form controls no longer trigger unwanted form submissions.
  • Malformed component values are now handled gracefully.

September 2026

25 releases: 0.5.0 → 1.75.2

September brought significant improvements to agent integration, runtime stability, and learning skill delivery across CopilotKit. The month featured new framework support, fixes to longstanding runtime issues, and enhanced developer tooling.

Action needed
  • Inspector Learning option removed from Fetch, Express, and Hono handlers; requires Intelligence runtime with configured Learning container.
  • Angular A2UI rendering no longer requires Lit dependency; web component catalog support added.
  • Channels packages no longer ship zod range from channels-discord and channels-telegram.
  • @copilotkit/react-core now requires @tanstack/react-virtual ^3.14.13.

Highlights

  • Angular components can now be registered directly as tools that agents call, expanding framework integration capabilities.
  • Native LangGraph and Mastra support for verified learned skill delivery enables agents to use dynamically learned capabilities.
  • Fixed multiple v1 runtime surfaces silently broken since v1.50.0, restoring critical functionality.
  • Automatic Learning skill delivery now available in BuiltInAgent with more resilient SSE streaming.
  • Improved runtime telemetry and enhanced Inspector onboarding experience for better visibility and usability.

New

  • CopilotActivity component for rendering individual activity messages outside the full chat view.
  • Initial releases of intelligence-langgraph and intelligence-mastra packages for native agent framework integration.
  • Provider-level agent configuration support in runtime surfaces.
  • Support for consuming learned skills from multiple containers across SDKs.
  • Targeted notifications in Inspector's What's New with launcher HUD impression tracking.
  • transformMessages option added to CopilotChatMessageView for message filtering and transformation.

Improved

  • Human-in-the-loop tools now wait for respond, keep parallel calls independent, and report errors properly on abort.
  • Home and launcher Learning status now sourced from Learning's own endpoint rather than Memory availability.
  • Managed Intelligence entitlement authority exposed through new readonly signals from runtime info.
  • MCP tool visibility filtering and unambiguous account/proxy selection now enforced.
  • React Native render-tool hooks converged onto react-core for consistency.
  • Inspector no longer treats in-memory agent runners as lacking durable Threads support.
  • More resilient SSE streaming for improved reliability.
  • Express is now an optional peer dependency, allowing Express 5 apps to compile correctly.

Fixed

  • Slot rendering with reactive component input and output bindings, template context, and projected fallback content.
  • Express bridge now preserves unread request payloads when body-parser sets an empty body.
  • Per-message state cloning and virtual-scroll performance issues in react-core.
  • Reconnect cursor preservation when control frame repeats old checkpoint.
  • A Memory policy that grants nothing no longer fails the run.
  • Socket reconnection no longer refreshes realtime credentials for sockets that never open.

August 2026

30 releases: 0.6.0 → 1.70.0

August brought major improvements to runtime reliability, security updates, and a redesigned web inspector with new navigation and thread management features. The month focused on stability, transport resilience, and better observability for deployed agents.

Action needed
  • Per-channel provider declaration removed from ChannelActivationConfig; one activation now declares all supported providers.
  • identifyUser is now optional when runtime declares at least one Channel with its own identity resolution.

Highlights

  • Security fix for XSS vulnerability in link opening; updated webcore to enforce strict URL allowlist.
  • New web inspector with grouped navigation, thread usage tracking, and inspector metadata for license visibility.
  • Improved runtime resilience with gateway handoff recovery and bounded in-memory agent runner to prevent out-of-memory errors.
  • Home dashboard and isolated agent Playground in the inspector for better project monitoring and agent testing.
  • Unified memory policy system for agent and browser access with per-channel grant scoping.

New

  • Inspector metadata system providing trusted identity, plan, and usage information end-to-end.
  • Web inspector Home page displaying project status, runtime health, available services, and Intelligence state.
  • Isolated agent Playground for testing conversations and forking saved threads without modifying originals.
  • Non-blocking product news notifications with unread signal on launcher.
  • Raw AG-UI event data in feedback callbacks for React implementations.
  • Automatic runtime transport selection for improved compatibility.
  • Channel status reporting that separates transport and provider connectivity legs.
  • Step lifecycle event recording in the inspector (STEPSTARTED and STEPFINISHED).

Improved

  • Runtime now accepts Intelligence runs that survive gateway handoffs with exponential backoff event queueing.
  • Agent runner events batch up to 32 per push with 5ms flush window when gateway supports it.
  • MCP SSE transport loads lazily to reduce bundle size for configurations not using it.
  • Memory realtime now works with project-only credentials without requiring user join tokens.
  • Assistant message rendering no longer misinterprets currency symbols as math expressions.
  • Correct live thread scope prevents duplicate threads in All Agents view.
  • Correct run IDs preserved after SSE replay so cumulative snapshots maintain message chronology.
  • HITL follow-ups now maintain single logical run ID across resolution steps.

Fixed

  • Channels recover from error-only websocket failures including transient 502 responses.
  • Late lock renewal failures no longer crash runs after thread lock is released.
  • AGUISendStateDelta now initializes arrays before append operations to prevent patch compaction failures.
  • Duplicate tool-call IDs no longer appear in message view when TOOLCALLSTART is applied multiple times.
  • Unenforceable mcpApps tool policies are now rejected instead of silently ignored.
  • Misconfigured memory agents now return 500 error instead of silently skipping configuration.

Summaries are written automatically from the official release notes (full changelog ↗); check the original notes before relying on a detail. CopilotKit: pricing, features and alternatives · All changelogs