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

jcode changelog: what's new each month

Every stable jcode 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

3 releases: 0.90.0 → 0.91.0

October brought major improvements to web search capabilities, provider compatibility, and integrations. The month focused on making searches faster and cheaper while adding new Google Workspace tools and better multi-provider support.

Action needed
  • Set websearch.prefernative = false or JCODEWEBSEARCHPREFERNATIVE=0 to revert to local web search instead of provider-side search.

Highlights

  • Web search now uses your model provider's native search by default, making results faster and cheaper than local scraping.
  • Google login now includes a built-in calendar tool for viewing, creating, updating, and deleting events.
  • Installing skills mid-session no longer clears the prompt cache, keeping subsequent turns fast and affordable.
  • OpenAI-compatible providers now automatically retry on heavy usage errors instead of failing.
  • New /desktop command opens your current session in Jcode Desktop.

New

  • Google Calendar tool for viewing, creating, updating, and deleting events through integrated Google login.
  • Guided Google OAuth app setup that detects existing gcloud credentials and validates project IDs.
  • /desktop command to open the current session in Jcode Desktop.
  • Alt+Up/Down keyboard shortcuts to cycle speed tier between Standard, Fast, and Ultrafast.
  • Gmail draft management: update, list, and delete drafts.
  • Esc key support to stop the current turn and run a pending follow-up prompt.

Improved

  • Provider-side web search results now summarized clearly when switching providers so the model doesn't retract correct answers.
  • MCP tool results and mcpsearch output are fitted to token budgets instead of fixed caps.
  • The mcp list action now lists only tool names for cleaner output.
  • Mcpsearch ranks name matches first in results.
  • The client displays the context window the server actually resolved instead of a placeholder.
  • OpenAI login validation no longer hangs when account hits usage limits.
  • Saved API keys stay in their env files so editing takes effect without restart.
  • Todo quality gates share rows instead of taking one line each.

Fixed

  • Retrying a turn no longer leaves web search rows from the aborted attempt.
  • OpenAI-compatible providers receiving 5xx heavy usage errors are now retried automatically instead of failing.
  • Browser actions now target the right browser when multiple browsers run the bridge extension.
  • The jcode log now records which binary actually launched to diagnose stale copies on PATH.
  • Copilot now accepts reasoning effort for every model whose catalog advertises it on both APIs.

September 2026

15 releases: 0.81.5 → 0.89.3

September brought major enhancements to remote work capabilities, voice input, and model support, including native SSH sessions, Claude Opus 5.5 integration, and a redesigned status line that keeps critical information visible on narrow terminals.

Highlights

  • Run the local terminal UI against a remote Jcode server over SSH, keeping workspace and tools on the remote host.
  • Claude Opus 5.5 is now the default Anthropic model with support for adaptive thinking and conversation-bound controls.
  • Built-in voice input for the CLI with native microphone capture on macOS and Windows and coding-aware dictation.
  • The status line is now always visible, adapts to terminal width, and shows git status, branch, and context usage.
  • Cursor provider now works correctly with MCP tools, reasoning models, and fast/high/thinking model IDs.

New

  • Run the local terminal UI against a remote Jcode server over SSH with explicit local login imports.
  • Claude Opus 5.5 available directly through Anthropic with adaptive thinking and progress summaries.
  • Native voice APIs with bounded microphone capture and Nari streaming for coded conversation.
  • Jev browser handoff now handles complete multi-step tasks and returns control only when finished.
  • Agents can show interactive applets and swarms can message each other with unique labels.
  • MCP tools load through provider-native deferred loading without busting prompt cache.
  • Accounts have persistent auto-switch failover pools that rotate in your chosen order.
  • Move live sessions between cloud and local with /cloud and /local, preserving uncommitted code.

Improved

  • Session status line stays pinned below input showing directory, branch, git status, context, provider, and model.
  • Status line on narrow terminals compacts information step by step instead of disappearing.
  • Model pickers remember route usage and use shared usage-aware ranking across clients.
  • OpenAI tool details begin streaming earlier.
  • Voice input starts recording immediately without waiting for connection and keeps first words.
  • Memory recall uses Jev to select relevant memories without loading a local embedding model at startup.
  • Light terminal themes have clearer contrast with better-aligned resume previews and Markdown blockquote gutters.
  • Browser tool now detects default browser and supports Chrome, Edge, Brave, Chromium, and Safari in addition to Firefox.

Fixed

  • Cursor MCP tool calls now complete on Composer, Grok 4.7, Grok 4.6, and Claude models instead of hanging.
  • Switching a session to Cursor with /model no longer leaves MCP tool calls stuck.
  • Cursor model names resolve against live model list, fixing validation errors for Grok 4.6, Claude Opus 5/5.5, and GPT-5.6 Sol.
  • Pasted input while choosing a login stays in the active picker instead of leaking into the prompt.
  • Clear inline images when the session is cleared.
  • Remote login status is sanitized before display so host credential details are never echoed back.

August 2026

27 releases: 0.65.0 → 0.81.4

August brought significant expansions to jcode's SDK capabilities, desktop experience, and agent control options. The month saw production-ready TypeScript and Rust SDKs, a redesigned desktop interface with multi-session workspaces, and deeper integrations with AI providers including native authentication flows.

Action needed
  • Todo status values now strictly constrained to pending, inprogress, completed, and cancelled; unknown values rejected instead of silently stored.
  • Cloud relay and session settings use neutral names while accepting legacy configuration keys for backwards compatibility.

Highlights

  • Production-ready TypeScript and Rust SDKs can now launch isolated agents or connect to running instances with full streaming and structured output support.
  • Desktop app redesigned with multi-session workspace navigation, project file explorer, persistent plans, and native math typesetting.
  • Pinned todos stay visible at the top while scrolling, with semantic quality assessments and automatic iteration based on evidence of remaining work.
  • Integration discovery now investigates candidates' capabilities, compatibility, setup, pricing, and limitations before selection.
  • Native authentication flows added for Cursor, Grok Build, and Gmail with memory and context preservation.

New

  • TypeScript and Rust SDKs with schema-validated structured runs, file management, and lifecycle event subscriptions.
  • Desktop project file explorer and in-app help overlay.
  • Multi-session workspace navigation in desktop with spatial panels and Vim-style keyboard navigation.
  • Tool output limits to prevent oversized results from overwhelming agent context.
  • Headless embedders with external wake mode for typed wake requests.
  • Session search across late transcript content with preserved search prefixes.
  • Integration discovery capability investigation before selection.
  • Opt-in transcript telemetry with privacy-preserving automatic secret redaction.

Improved

  • Self-update now shows live progress bar and graceful in-place reload.
  • Ctrl+L is true terminal-style clear with history preserved above the prompt.
  • Desktop scrolling faster and smoother across all input methods.
  • Tool schemas adapt automatically to each provider and recover from incompatibilities.
  • Todo quality checks now shorter and clearer with better guidance.
  • Session context now reports local date, time, and timezone.
  • Large MCP tool catalogs automatically deferred to keep agent context compact.
  • Model thinking shown by default with full thought history display.

Fixed

  • Todo completion synonyms no longer trigger false auto-poke loops.
  • Desktop sessions connect reliably when startup is slow.
  • OpenAI model responses no longer duplicate thinking text.
  • Closed macOS terminal windows no longer leave jcode sessions running unexpectedly.
  • Session attachments correctly correlated when multiple clients connect concurrently.
  • Transient provider failures and MCP notifications handled more reliably.

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