What's New

MCP

• New connectors: added out-of-the-box Salesforce and Datadog MCP servers to the catalog.
• Live attach: add an MCP server to a running session without restarting the VM.
• Composer @-mentions: attach MCP servers by @-mention directly in the chat composer.
• Reliability: clearer connection status and friendly names for failed servers, plus fewer spurious OAuth "Reconnect" prompts.

Sessions

• Tunnels: open a port to a cloud environment to access running services, like your development server, from anywhere.
• Usage & cost: see per-turn and per-session usage and cost, with accumulated cost badges, in the session details.
• Session details: session creator, originating webhook, and linked environment now shown.
• Attachments: upload SVG files as chat attachments.

Admin, Configuration, and Settings

• Experts: view expert version history and roll back to previous version.
• Spaces: MCP servers and secrets can be assigned to multiple spaces.
• Settings: moved all personal and organization settings into a single settings navigation.

Application

• Sidebar: display settings (filters, sort, density), pinned-folder controls, empty-folder tidy-up, and "Archive all sessions" for folders.
• Mark unread: mark sessions — including active ones — as unread.

New Features

• Cloud Workflow Triggers: `auggie cloud` experts can now use Jira, Confluence, GitLab, and Microsoft Teams as persistent workflow trigger sources.
• Session Space Pinning: `auggie cloud session create` now accepts `--space-id` to pin a new session to a Space.
• VFS File Links: Added `auggie cloud vfs get-url` to generate compact, durable links to VFS files.

Improvements

• Daemon Worktree Directory: `auggie daemon` can now be configured with a custom worktree directory.

Bug Fixes

• Fixed `auggie cloud expert list --by-usage` to report weekly active users consistently with the webapp.
• Fixed owner parity for opted-in shared MCP servers.
• Fixed `${...}` variable expansion across MCP server and hook config files, including plugin-root variables.

New Features

• Cloud Projects: Added `auggie cloud project` commands to create and manage projects and their agents.
• Cloud Tunnels: Added `auggie cloud tunnel open/close/list` commands.
• Cloud Analytics: Added `auggie cloud analytics` commands to record custom events and view cost analytics.
• Sub-Agent Hooks: PreToolUse and PostToolUse hooks now run during sub-agent sessions.

Improvements

• Expert Discovery: `auggie cloud expert list` adds usage-ranked discovery (`--by-usage`), a `--space-id` filter, Space membership display, and improved discovery ergonomics.
• Secrets Spaces: `auggie cloud secret list` and `set` now accept a `--space` flag to scope secrets to a Space.
• Workspace Discovery: `auggie daemon` can auto-discover git workspaces under a non-git container (with optional nesting depth) and accepts `-w`/`--workspace-root`; cloud agents on custom environments also auto-discover workspaces.
• MCP Server Restart: Added targeted MCP server restart without restarting the cloud VM.
• MCP Diagnostics: Selected-but-dropped registry MCP servers are now reported so diagnostics resolve correctly.
• Context Compression: Restored context compression so long agent sessions handle large conversations more reliably.
• Syntax Highlighting: Added Scala and additional language grammars to the TUI syntax highlighter.
• Cloud Events: Added Teams as a cloud events source.

Bug Fixes

• Fixed the agent loop silently exiting on an empty model response (previously surfaced as "Worked for 0s"); it now retries and reports an error.
• Fixed the cloud message queue staying paused after an agent interrupt.
• Fixed `knowledgebase` path validation on Windows.
• Fixed git worktrees to resolve to their main repository for indexing permission checks.
• Fixed the expert capability map to include GitLab and align Jira/Confluence capabilities.
• Fixed an overly strict `$augment-*` secret-name validator in MCP bundles.

Claude Fable 5 (Mythos family) is now available in the Augment Code model picker. At roughly 2x the cost of Opus 4.8, Fable 5 is intended for long, multi-step tasks that require deep reasoning.

In early testing, it caught every known issue in our toughest internal code review benchmark and flagged additional potential problems we’re still validating. In practice, it has started to expose the limits of benchmarks we've relied on to differentiate frontier models.

If your day-to-day work is mostly coding, existing models continue to offer tremendous value. For longer running, more complex work involving deep reasoning and multi-turn workflows, Fable 5 offers an advantage in quality.

Fable 5 is at its best when it can operate with full context and follow-through, not just generate an answer. Augment’s unified agent platform, Cosmos, unlocks that.

With Cosmos, Fable 5 can pull the right context from across your tools, coordinate work across the full SDLC (triage → spec → implementation → review → test → deploy), and keep durable memory of decisions and patterns so progress compounds over time. Instead of a single prompt in a single surface, you get a coordinated system where agents run in the background, trigger off real workflow events, and bring your team in when judgment matters, so you can apply Fable 5’s premium capability to the hardest, highest-leverage engineering work.

One caveat: Early Mythos model family reports focused on the extensive cybersecurity capabilities of this model class. Anthropic applies hard restrictions on this use case, and will fall back to the Opus family for those tasks.

Fable 5 is powerful, and that capability comes at a premium. We’ll soon be adding it to our model router, Prism, to help teams access right-sized models for the tasks at hand, saving cost without sacrificing quality.

If you try Fable 5, let us know: we’re eager to hear how you use it and where you find the most value with this new model family.

New Features

• Auggie Guide: Added a built-in `auggie-guide` subagent for answering Auggie CLI and Cosmos questions from the docs.
• Daemon Spaces: Added CLI support for plain daemon Spaces.
• Daemon Indexing: Added `--allow-indexing` support to `auggie daemon` so spawned child agents can use codebase indexing.

Improvements

• Session Metadata: Session JSON now records per-exchange model IDs, summarization billing, and sub-agent histories.
• Indexing Visibility: `auggie daemon` now reports indexing status at startup and explains how to enable it when possible.
• MCP Reliability: MCP OAuth refreshes are more resilient, and stale shared MCP opt-ins now surface CLI warnings.
• CLI Welcome: The welcome screen now highlights updated model availability.

Bug Fixes

• Fixed `auggie upgrade` on Windows.
• Fixed `auggie mcp add --header` to use `-H` and avoid conflicting with help.
• Fixed daemon auth-session file handling to write atomically and recover from malformed files.
• Fixed markdown code fences from carrying over between agent turns.
• Fixed out-of-order messages in long paginated conversation turns.
• Fixed subagent delegate tool rendering.

Improvements

- Lighter custom environment images : Custom environments no longer require Node.js. Only bash and git are needed, allowing smaller and leaner Docker base images. Commonly requested

New Features

• General-Purpose Subagent: Added a built-in general-purpose subagent option for broader multi-step research and implementation tasks.

Improvements

• Model Defaults: Added a configurable default model setting, and `/model` changes now apply only to the current session.
• Markdown Links: The TUI now shows destination URLs for markdown links.
• Daemon Update Hints: The app now warns when a connected daemon is out of date.
• Cloud Error Visibility: Cloud sessions now surface CLI startup errors more directly in Cosmos.
• VFS Diagnostics: VFS sync warnings and MCP merge collision errors now include clearer troubleshooting details.
• Lazy Skill Includes: Lazy-loaded skill bodies can now include nested skills.

Bug Fixes

• Fixed cloud agents to retry transient knowledgebase sync failures and stop cleanly on unrecoverable errors.

New Features

- Service account attribution for automated workflows -- Tenant admins can now configure workflows to run triggered sessions as a specific service account user, ensuring proper attribution and permissions for automated tasks.

- Workflow service account support -- Tenant admins can now configure automated workflows to run as a specific service account user for proper attribution and permissions.

- Scoped webhook subscriptions -- Agents can now subscribe to events from a specific custom webhook instead of receiving all webhook deliveries tenant-wide, enabling more targeted event-driven workflows.

- Cloud agent hosted artifacts -- Cloud agents can now create and share hosted artifacts via the CLI and API, available across all environments

Improvements

- Webhooks page redesigned as card layout -- Webhooks are now displayed as cards with owner avatars, sorted newest-first, with a cleaner create dialog and inline help text.

- Faster expert list loading -- Expert list pages load significantly faster for tenants with many experts, eliminating timeout errors when switching tabs. Commonly requested

- Org-scope shared files writable by all team members -- All authenticated tenant users can now write to org-scope shared files, removing the previous admin-only restriction.

- New Slack tool capabilities for agents -- Agents can now verify Slack token identity and retrieve permanent message links, enabling richer Slack integration workflows.

- Webhook management redesign -- Webhooks are now displayed as cards with owner avatars and newest-first ordering, replacing the previous table layout.

- Org-scope shared files now writable by all users -- Non-admin users can now write to org-scope shared files, removing the previous admin-only restriction.

- Expanded Slack tool capabilities -- Agents can now verify Slack token identity and retrieve permanent message links, enabling richer Slack integration workflows.

- File attachments in web conversations -- Non-image file attachments are now available to all production users in Cosmos web sessions

- Faster cloud sessions list -- The "all sessions" page now loads faster, with optimized queries and parallel data fetching.

- Startup spinner replaces stop button -- While a cloud session is starting, the stop button is replaced with a spinner and tooltip, preventing confusion from clicking a non-functional control.

- More reliable session keepalive -- Session liveness tracking is now active across all environments, reducing the chance of sessions timing out prematurely.

- Faster sessions list -- The "all sessions" page now loads significantly faster

- Clearer startup state -- The stop button shows a spinner with a tooltip while a cloud session VM is starting, instead of displaying a non-functional stop icon

- More reliable session keepalive -- Sessions are now less likely to time out prematurely during active use

Bug Fixes

- Archiving cloud sessions now properly stops background processes -- Archiving a session running in a shared daemon pool now stops the underlying process, freeing up pool slots and preventing orphaned sessions from consuming resources.

- Expert list loading performance -- Expert list loading is significantly faster for tenants with many experts, eliminating timeout errors and slow tab switching.

- Session archive resource cleanup -- Archiving a cloud agent session now properly stops the underlying process, preventing orphaned sessions from consuming resources.

- Event processing reliability -- A single misconfigured event subscription filter could previously block event delivery for all other subscriptions in the same group; this no longer happens

- First message display in new sessions -- The initial message in a new Cosmos session now reconciles correctly with the server, eliminating a visual glitch where the first message appeared without a stable identifier

- Event delivery reliability -- Fixed an issue where a single misconfigured event subscription could delay or block event processing for all other subscriptions in the same account. Commonly requested

- Initial message display in new sessions -- The first message in a new session now appears correctly without a brief visual glitch during reconciliation

- Cloud sessions recover from deleted snapshots -- Sessions that referenced a deleted snapshot were previously stuck and unable to restart; they now automatically recover by starting a fresh environment, with a notification that previous filesystem state was lost.

- Expert daemon pool routing fixed -- Experts configured with daemon pools now correctly route triggered sessions to the selected pool, instead of failing with environment errors.

- Terminal and Files tabs hidden for session viewers -- Tenant admins and other session viewers no longer see broken Terminal and Files tabs that briefly appear then error out.

- Automatic recovery from deleted snapshots -- Cloud sessions that referenced a deleted VM snapshot were previously stuck and unable to restart; they now automatically recover by starting a fresh VM

- Expert daemon pool routing -- Experts configured with daemon pools now correctly route triggered sessions to the selected pool, instead of failing with environment errors

- Terminal and Files tabs for session viewers -- Tenant admins and other session viewers no longer see broken Terminal and Files tabs that briefly flash then error out
Bug Fixes

- Shared session daemon status -- Shared cloud agent sessions backed by daemon pools now correctly display connected status instead of appearing disconnected

Improvements

• Network Reliability: CLI network requests now use a consistent HTTP client for more reliable shutdown and connection cleanup.

New Features

• General-Purpose Sub-Agent: Added a built-in general-purpose sub-agent for deep research, code search, and multi-step workflows.