Mistral Vibe changelog: what's new each month
Every stable Mistral Vibe release summarised by month: the highlights, new features, improvements, fixes and anything you need to act on. 10 months covered; the current month updates daily.
November 2026 · so far
5 releases: 3.1.0 → 3.1.4
November brought significant improvements to Mistral Vibe's interface and extensibility. The release added Vim keybindings, persistent goals across turns, and enhanced configuration options for managing plugins and skills.
- enable_file_autocomplete now controls @ file completion and background file indexing.
- enable_sticky_prompt defaults to true; set to false to restore previous scrolling behavior.
- disabled_plugins and enabled_plugins settings control plugin state; previously installed plugins may need configuration.
Highlights
- Vim Normal and Insert mode support in the prompt input via vim_keybindings config option.
- New /goal command lets you set objectives that Vibe works on across multiple turns until completion.
- /skills command displays installed skills with search and toggles, and manages shared registry skills with version pinning.
- New context window visualization showing system prompt, tools, skills, and conversation space allocation.
- Plugin and skill management now supports disabling installed plugins and enabling skills by default.
New
- Stash draft messages with Ctrl+S to retrieve after sending another message.
- Pre-agent and post-agent hooks let you block, rewrite or enhance messages before the model sees them.
- Session hooks run on start, resume, clear, fork or end to add context or perform cleanup.
- Pre-LLM and post-LLM hooks for intercepting model calls and requesting retries within budget.
- Notification hooks run in the background for approvals and turn completion events.
- Goal management with /goal edit, /goal pause, /goal resume and /goal clear commands.
- Rail items open in place of input box for subagents, background tasks, pull requests, goals and todo list.
- Image file display support when using read_file on models that read images.
Improved
- PR/MR list now responds to Ctrl+C, Home, End, PageUp and PageDown navigation.
- Jump to bottom button appears when conversation is scrolled up.
- Prompt stays pinned at top of conversation while scrolling with optional disable via enable_sticky_prompt.
- Model picker now has fuzzy search with / prefix to filter by name.
- Footer shows Vibe Code usage percentage with color indicators and PAYG status.
- Tool group headers count running calls separately from completed calls.
- Status line displays what background work the agent is waiting for.
- Hovering over local file links shows code preview in a popup.
Fixed
- vibe-acp now properly notifies clients when administrator changes Vibe Code configuration.
- Shared skill versions unused for 30 days are deleted when registry can still serve them.
- Project-scoped skills, files and pins now correctly override global ones of the same name.
- Collapsed turn details expand with click or Ctrl+O for access to intermediate messages and reasoning.
- Emacs Ctrl+P and Ctrl+N keybindings now correctly navigate prompt history without edge cases.
October 2026
7 releases: 2.26.0 → 3.0.4
October brought major updates across Mistral Vibe's terminal and desktop interfaces, with version 3.0 introducing improved session management, expanded agent capabilities, and finer control over model behavior and costs.
- outside_workdir = "deny" (global or per tool) now refuses file and shell access outside working directory with a tool error instead of a prompt.
Highlights
- Rust CLI now supports voice mode with narration, branching conversations, and clickable interactive elements like tool approvals and model selection.
- Sessions can now fork into branches, and the app tracks subagent execution with individual cost pricing and status monitoring.
- Desktop can import projects and sessions from other coding agents during onboarding, and experimental model catalog shows Mistral's curated models.
- Images sent to models that don't support them no longer fail the turn; they're replaced with placeholder text and the session continues.
- New settings give admins fine-grained control over subagent limits, compaction behavior, and auto-approve restrictions.
New
- Rust CLI /teleport command sends sessions to Vibe Code Web with push approval and cancellation.
- Rust CLI /loop schedules recurring prompts in natural language with model-managed cron tools.
- Rust CLI /branch forks the current conversation into a new resumable session.
- Desktop import wizard brings projects and sessions from other coding agents during setup.
- Experimental model catalog with Mistral's curated models, overridable per-session.
- Rust CLI /release-notes command browses searchable full-screen release notes for all versions.
- Session subagent API methods allow clients to wait for, interrupt, and read subagent status.
- Experimental vibe run-server command runs multiple sessions from one process on shared app-servers.
Improved
- Vibe Code now warns when the selected model cannot view attached images.
- Sessions list their background processes in a side panel with live status and a Stop action.
- Model reasoning now shows by default as a collapsible Thought row.
- App-server reports context fullness and when sessions are open elsewhere as read-only.
- Rust CLI model and thinking pickers now apply immediately and can be selected per-session with 's' key.
- Free-text answers to agent questions accept multiple lines and multi-line pastes.
- Skills added or edited on disk are picked up by running sessions without manual reload.
- Footer shows active model and thinking level next to token counter and updates in real-time.
Fixed
- API key stored in ~/.vibe/.env fallback is now created owner-only so it's not readable by other accounts.
- Custom tools that fail to load now show a warning with the file name and error instead of disappearing silently.
- Connectors needing no login now show as Not connected in VS Code and connect in one click.
- Vibe warns when the model refuses a request as too long or compaction shortens inputs.
- Setup now warns when a shell-set API key overrides the saved key.
- Using --agent to pick a disabled mode now opens your default starting mode with a notice.
September 2026
8 releases: 2.25.0 → 2.25.8
September brought major improvements to session management, connector integration, and the experimental unified harness, alongside better control over background processes and tool execution. Desktop and IDE experiences were significantly enhanced with new status indicators and permission handling.
- SessionBackendKind literals renamed from "python"/"rust" to "legacy"/"unified" to match backend identity naming.
Highlights
- New /branch command lets you fork conversations into separate resumable sessions.
- Skills browser with search and filter for managing session-scoped skills, plus ability to toggle skills on and off without removal.
- Smart-approve now escalates risky actions to confirmation prompts instead of blocking them, with transparent reasoning shown to users.
- Connectors for Notion, Linear, Slack, GitHub, Gmail, Google Calendar, and Drive now available in the code editor.
- MCP servers declared by plugins now start automatically and are callable within sessions.
New
- Skills browser for session skills with search, filter, and shared skills syncing on session start (experimental, off by default).
- Plugin-declared MCP servers in /mcp with per-plugin tagging and login authentication.
- /branch command to fork the current conversation into a new resumable session.
- Built-in worktree plugin with skills for creating, reusing, and cleaning up git worktrees.
- Background process tools (process.start, process.stop, process.output, process.list, process.write) with improved visualization.
- Subagent conversation history readable through the owning parent session, with configurable live status and read-only transcript switching.
- Session todos tracked and shown as a pinned line under input, with /todo command for full view.
- Per-session scratchpad in experimental-harness sessions with notes restated to the agent.
Improved
- Tool call groups now fold into compact summary lines by default, expandable on click, with Ctrl+O to toggle all.
- @ file mentions now use Git-aware discovery and accept standalone pasted files and folders.
- Session-scoped model pinning for resumed conversations.
- Desktop keeps 15 recent managed Code worktrees by default, with configurable cleanup that skips active or dirty worktrees.
- Browser SSO now completes behind split-horizon reverse proxies with opt-in URL rewriting.
- Smart-approve shows approval reason as a note rather than a warning for auto-approved calls.
- Session titles on custom Mistral endpoints now use the active model.
- File completions in Rust CLI now respect caret position and allow directory descent.
Fixed
- Invalid API key errors now show where the key came from, revealing shadowed environment variables or keyring conflicts.
- Single invalid connector no longer prevents the rest of your connectors from loading; individual tools with oversized schemas are dropped independently.
- Sessions no longer drop from the picker when touched while paging through the list.
- On Windows, opening a session already open in another process now reports a clear error instead of a raw permission error.
- On Windows, idle-session cleanup no longer mistakes sessions held by other processes for idle ones.
- Configuration warnings now properly report when managed configuration contains removed Vibe Code settings.
August 2026
7 releases: 2.23.3 → 2.24.5
August brought significant improvements to session management, configuration handling, and performance across Mistral Vibe. The month focused on smoother workflows, better model routing, and faster startup times.
- Default agent renamed from "ask" and now defaults to accept-edits mode.
Highlights
- Admin config layer lets teams enforce shared configuration over user settings.
- Session picker now shows live preview in-app for easier navigation.
- New /retry command and improved retry handling with adaptive backoff and connection pooling.
- Faster startup: TUI renders before session initialization and imports are deferred.
- Slash commands and setting pickers work while the agent is running.
New
- /retry command to retry an interrupted response.
- Admin config layer for shared and enforced configuration.
- "Default" unpinned model option to track recommended models without pinning specific aliases.
- Inline ghost-text completion for mid-prompt skills.
- /clear and /new commands accept optional first prompt and show resume ID.
- In-app session picker with live preview.
- /log-level command for runtime log-level control.
- LLM-generated session titles and terminal title.
Improved
- /rewind now lets you choose between forking and rewinding in place.
- LLM retries surfaced in UI, logs, and ACP/VS Code conversations.
- /status shows provider prompt-cache usage and discounts cached tokens from session cost.
- User and project config layers compose with trusted project config overlaying user config.
- Local sessions start in the selected worktree.
- Faster session resume by deferring sub-agent instantiation and reusing model connections.
- Edit and read tool headers show paths relative to current directory.
- Models displayed by name instead of raw config alias.
Fixed
- Refused config change no longer wipes the picker.
- Connector bootstrap failures now surfaced instead of reporting no connectors.
- Links normalized before opening to prevent broken clicks.
- Themes apply correctly at startup.
- Loading nix-managed skills through symlinked SKILL.md no longer freezes.
- Python dependencies updated to clear known security advisories.
July 2026
9 releases: 2.18.4 → 2.23.2
July brought major configuration improvements, shell tool enhancements, and a significant hooks system update. The month featured both experimental A/B testing of OS-native shells and graduation of previously experimental features to stable.
- Hook types renamed in hooks.toml: postagentturn → postagent, beforetool → pretool, aftertool → posttool.
- Models field migrated from list to dict format in configuration.
Highlights
- Hooks system graduated from experimental with renamed hook types: postagentturn, beforetool, and aftertool are now postagent, pretool, and posttool.
- New managed bash tool for running and tracking shell sessions, with OS-native shell tools available as A/B testing.
- Registry skills system for discovering and managing skills, plus a skill-creator built-in skill for authoring your own.
- Redesigned /config command now offers a searchable, full-screen settings browser with typed edit modals.
- ACP review support lets you approve or revert agent changes directly from ACP clients with live refresh.
New
- worktree NAME option to create or reuse a git worktree and run inside it.
- disabled-tools CLI flag to turn off specific tools for a run.
- Per-tool description overrides via markdown files in the tools directory.
- vibe mcp add and vibe mcp remove commands to manage remote MCP servers without editing TOML.
- @file mentions now inject as readfile tool calls.
- Auto theme that follows terminal or OS appearance.
- Skill-creator built-in skill for authoring, updating, and deleting your own skills.
- Browser sign-in can target custom login domains for self-hosted deployments.
Improved
- Whole-line content now shown in the edit diff.
- Long tool call titles wrap instead of being cropped.
- Session resume with --continue now matches the resolved working directory.
- Leanstral context limit raised.
- Tool descriptions compressed to reduce context bloat.
- Scrolling in long conversations is smoother and faster.
- Dense, grouped tool-call rendering in the TUI.
- Trust dialog simplified and now defaults to "Trust folder".
Fixed
- MarkupError crash when a tool error contained square brackets.
- StopIteration crash in QuestionApp when submitting an empty multi-select.
- Crash when @-mentioning a file, which aborted the turn.
- Exception on unsafe configuration read.
- grep result chips no longer leak absolute host paths into context on ACP.
- Duplicate mcpservers names in config are now rejected.
June 2026
11 releases: 2.14.0 → 2.18.3
June brought significant usability improvements to Mistral Vibe, with new hooks for tool customization, better file handling, and enhanced IDE integration. The month focused on making the interface more responsive, securing credentials, and expanding MCP server support.
- writefile tool is now create-only and refuses to overwrite existing files.
- Built-in tool prompts no longer reference searchreplace; use edit instead.
Highlights
- Message queue lets you type while the agent is working, with Esc to pause and Ctrl+C to drop messages.
- Experimental beforetool and aftertool hooks let you intercept and modify every tool call via shell scripts.
- API keys now store securely in the OS keyring instead of plain text.
- Clickable URLs in web tool output and image paste support in the TUI (macOS).
- Context window token usage now displays at the bottom of the TUI.
New
- Image attachments via @-mentions in the TUI for vision-capable models.
- Session deletion from the resume picker and over ACP.
- Fallback URL for browser sign-in when the browser cannot open automatically.
- EnvironmentLayer configuration populated from VIBE-prefixed environment variables.
- /mcp login, /mcp logout, and /mcp status commands for OAuth-backed MCP server authentication.
- vibe --check-upgrade command to force an immediate update check.
- yolo as an alias for --auto-approve.
- j/k navigation in selectable lists across the TUI.
Improved
- Tool result output collapses by default to keep responses scannable.
- Common read-only shell commands (ls, cat, pwd, etc.) are now allowed without approval by default.
- Syntax-highlighted, line-numbered, theme-aware diffs for file edits with all replaced occurrences shown.
- Fuzzy search in slash command autocomplete.
- /resume now scopes the session picker to the current folder and opens much faster.
- Clipboard copy confirmations appear inline instead of stacking toast notifications.
- Completion popup uses a two-column layout.
- Chat messages have right padding so text no longer collapses into the scrollbar.
Fixed
- Toast notification now shows when a SKILL.md file fails to parse instead of silently dropping it.
- JetBrains IDEs no longer trigger a preemptive auth prompt and no longer drop terminal arguments.
- Context-too-long responses now classified correctly when wrapped as HTTP 422.
- SSE streams parsed on CR/LF boundaries only to avoid crashes on valid Unicode line separators in JSON.
- MarkupError crash when tool output contained square brackets.
- OpenTelemetry chat/LLM spans missing after Mistral Python SDK 2.4.10+ telemetry opt-in gate.
May 2026
9 releases: 2.9.4 → 2.13.0
May brought significant usability improvements across sessions, configuration, and connectors, along with new A/B testing infrastructure and a revamped onboarding experience. The month saw refinements to plan editing, bash execution, and SSL handling for enterprise environments.
- Programmatic mode (-p) no longer auto-approves tool calls by default.
- enableexperimentalbrowsersignin flag no longer required for Mistral providers; stale entries silently ignored.
Highlights
- New /loop command runs prompts or slash commands on a recurring interval.
- Plan mode now displays live-editable plans with Ctrl+G to modify them directly.
- Skills can be shared across agents by loading from ~/.agents/skills.
- Connectors migrated to public GA endpoints with improved authentication surfacing.
- GrowthBook A/B testing layer added for system prompt experiments.
New
- /rename command to rename the current session.
- /loop command for recurring prompt or slash command execution.
- add-dir flag to pull additional repository roots into a session.
- Load skills from ~/.agents/skills for cross-agent sharing.
- Textual theme selection system with onboarding theme picker.
- Custom compaction prompts via compactionpromptid config setting.
- max-tokens flag for programmatic mode.
- enablesystemtruststore config flag for OS trust store and corporate TLS setups.
Improved
- Bash bang commands now run via async subprocess for better latency.
- Tool permissions marked "Always allow" persist across sessions.
- Syntax-highlighted file diffs for writefile and searchreplace in the IDE agent webview.
- Plan mode readability improved with live-editable display.
- Parallel bash tool calls in ACP each render their own live terminal.
- Session titles now format @mention syntax for human-readable display.
- Connectors disabled by default when newly discovered; existing user choices preserved.
- MCP HTTP transport now uses shared SSL context honoring SSLCERTFILE and system trust store.
Fixed
- Non-retryable errors raised by sub-activities now correctly stop retry in Temporal activities.
- Compacted session IDs displayed correctly.
- Parallel instances no longer clobber each other's history files.
- Wildcard suffix stripped when persisting bash allowlist patterns.
- Original line endings preserved in ACP searchreplace.
- Crash during initialization.
April 2026
11 releases: 2.7.2 → 2.9.3
April brought major enhancements to Mistral Vibe's debugging, configuration, and integration capabilities. The month culminated in version 2.9.0 with new scratchpad support, an experimental hooks system, and expanded model options.
Highlights
- Builtin skills system including self-awareness capabilities for agents.
- Scratchpad directory for temporary working files shared between agents.
- Console View for enhanced debugging and monitoring of agent operations.
- Experimental hooks system with post-agent-turn lifecycle support.
- Support for Opus 4.7 model and OpenAI Responses API adapter.
New
- Alt+Left and Alt+Right keyboard shortcuts for word-wise cursor movement in chat input.
- /data-retention slash command to view Mistral AI's data retention notice and privacy settings.
- /mcp command to display MCP servers and their status.
- /connectors alias for /mcp with R refresh shortcut in MCP browser.
- /copy slash command for improved workflow efficiency.
- Connector OAuth authentication flow in /mcp menu.
- Thinking level picker in ACP CLI for model configuration.
- Teleport surface for GitHub connection status monitoring.
Improved
- Manual command output forwarding to agent context.
- Webfetch content truncation for better readability.
- Startup time improved through lazy loading of heavy dependencies and optimized folder parsing.
- Bash denylist matching and error messages.
- Encoding detection fallback in readsafe for non-UTF-8 files.
- MCP and git I/O deferred to background thread for faster CLI startup.
- Parallelized git subprocess calls during startup.
- Builtin skills system with enhanced agent self-awareness.
Fixed
- Broken build on Linux.
- Errored MCP servers now properly excluded from the banner count.
- Command messages are now skipped during rewind navigation.
- Stale configurations from subagent switch.
- Markdown fence context loss causing streaming rendering problems.
- Textual version compatibility issues.
March 2026
9 releases: 2.4.0 → 2.7.1
March brought significant expansions to Mistral Vibe's capabilities, including new AI modes like voice and theorem proving, better conversation navigation, and improved infrastructure for observability and tool execution. The month saw seven releases focused on both user-facing features and backend robustness.
Highlights
- Voice mode with real-time transcription support for hands-free interaction.
- Rewind mode lets you navigate and fork conversation history to explore different paths.
- Theorem proving agent powered by leanstral for mathematical problem solving.
- Parallel tool execution improves performance when running multiple operations.
- Text-to-speech functionality complements the new voice capabilities.
New
- Dedicated theorem proving agent accessible via /leanstall command.
- Voice mode with real-time transcription support.
- Text-to-speech (TTS) functionality.
- Rewind mode to navigate and fork conversation history.
- Skill tool for managing task lists and workflows.
- AGENTS.md support at user level (~/.vibe/) and folder hierarchies for customizing agent behaviour.
- Standalone --resume command for session picker.
- OpenTelemetry (OTEL) tracing support for observability.
Improved
- Parallel tool execution for better performance.
- Plan mode improvements.
- Skills now extract arguments when invoked for better parameter handling.
- Web search tool infers server URL from provider config instead of using hardcoded production API.
- BFS for vibe folders improves startup performance.
- List-based model picker for /model command.
- Update notification toast no longer times out, ensuring users see restart prompts.
- Smoother scrolling with scroll sensitivity reduced to 1 line per tick.
Fixed
- Space key functionality restored in VSCode 1.110+ and extended to all input widgets.
- Arrow-key history navigation now works correctly at wrapped-line boundaries.
- Ruff isort and formatter config conflict resolved.
- Text selection errors when copying from unmounting components.
- Non-standard HTTP 529 status codes now handled gracefully with retries.
February 2026
1 release: 2.3.0
February brought expanded web capabilities, improved session management, and streamlined tool interactions. The release focuses on better user control and real-time feedback across chat and API modes.
Highlights
- Web search and web fetch tools let agents retrieve and search live web content.
- Resume sessions by name with the new /resume command for easier conversation management.
- Tool call arguments now stream incrementally in the UI for faster visual feedback.
- MCP servers can request LLM completions via the sampling protocol, enabling more sophisticated agent interactions.
New
- /resume command to select which session to resume
- Web search tool for retrieving web content
- Web fetch tool for fetching web content
- MCP sampling support for LLM completions from MCP servers
- Chat mode for ACP via session/setconfigoptions
- Notification indicator in CLI with terminal bell and window title alerts
- MCP server discovery cache (MCPRegistry) persisting across agent switches
Improved
- Tool call arguments stream incrementally in the UI instead of appearing all at once
- ACP session/setconfigoptions now supports switching mode and model
- MCP server discovery survives agent switches without re-discovering unchanged servers
Summaries are written automatically from the official release notes (full changelog ↗); check the original notes before relying on a detail. Mistral Vibe: pricing, features and alternatives · All changelogs