What We Shipped: Geta.Team v2.0.2 to v2.0.11

Five releases in three weeks. Here's everything that shipped in Geta.Team from mid-February to today, starting with the most recent.

v2.0.11 — March 4, 2026

Claude Auth Infinite Loop Fix (Critical)

On instances with active employees, connecting Claude via OAuth was stuck on "Waiting for authentication..." forever. Fresh instances worked fine. The root cause: the auth detection code was running on ALL sessions instead of just the dedicated auth session. A single guard (sessionId === 'claude-auth') fixed it.

Related: after completing OAuth, Claude wasn't appearing as a provider option because the auth/complete endpoint set claude_is_auth but not claude_configured. Both flags are now set together.

Smarter Admin Defaults

• Admin role = full access. No more toggling "Can Manage Connectors" separately for admin users. If you're an admin, you get everything
• SMTP-aware invitations. If your SMTP isn't configured, you get a "Copy Link" button instead of a "Send Invitation" button that silently fails. When SMTP is configured, you get both options
• Per-user MFA toggle. Users can now skip MFA email verification from their account settings
• Provider selection is mandatory. The CLI choice popup can no longer be dismissed by clicking outside, which was leaving sessions in an undefined state. If only one provider is available, it auto-selects without showing the popup
• Invitation URL fix. Emails no longer contain https://null/... URLs when no APP_URL is configured. Added req.get('host') as a fallback
• Email visibility. Employee email and the Email tab are now hidden entirely when the email system isn't configured

v2.0.7 — February 26, 2026

Phone Calls Got a Complete Rebuild

The biggest architectural change in the batch. We migrated the entire telephony system from SIP/RTP to Twilio ConversationRelay.

What that means in practice: we eliminated the SIP server, the RTP bridge, and the manual audio encoding/decoding pipeline. Twilio now handles speech-to-text and text-to-speech natively. The backend receives clean text transcriptions via WebSocket, processes them through the LLM, and returns text responses.

The result is simpler infrastructure, fewer failure points, and support for multiple TTS providers: Google Wavenet, Amazon Polly (Neural), and ElevenLabs premade voices.

We also added echo suppression (a _ttsPlaying flag mutes voice activity detection during playback with a 300ms grace period) and fixed transfer calls that were hanging up prematurely. The search_contacts tool now keeps phone numbers and emails private during calls — your employee confirms the contact exists but never reads sensitive info aloud.

Voice Cloning

You can now clone your own voice and assign it to your AI employee's phone calls.

The UI lives in the Chat sidebar under "Voices." Record up to 15 seconds of audio reading a prompt, give it a name, select a language, and submit. The cloned voice is generated via Inworld's inworld-tts-1.5-max model and available immediately for phone calls.

Each user manages their own voices with preview and delete controls. The cloned voices are per-user — nobody else on the instance can see or use yours.

Permission Fixes

• Admin users were getting 403 errors on voice tools and call usage endpoints — the access check wasn't verifying is_admin
• Owner (user ID 1) was locked out of Configuration because can_manage_connectors was false in the database. Now permanently safeguarded in code
• Conference calls temporarily disabled due to a ConversationRelay limitation (a call leg can't be in a Conference and ConversationRelay simultaneously)

v2.0.4 — February 19, 2026

84,000+ Skills, Browsable From Your Dashboard

The Skills Marketplace is live. Your AI employee can now browse, search, and install from over 84,000 community-built skills powered by the skills.sh ecosystem — directly from the dashboard.

The flow: open the Skills panel, search or browse by popularity, read the skill's documentation, click Install. Your AI employee reads the instructions and creates the skill autonomously. No code, no CLI, no config files.

The marketplace is paginated, searchable with debounce, and each skill shows install counts so you can see what's battle-tested. Skill detail pages render the full README with proper markdown formatting.

White Label: Make It Yours

Enterprise Edition users can now fully rebrand the platform:

• Custom logo — upload your own, replaces the default everywhere (sidebar, login page, favicon, emails)
• Custom brand name — appears in the browser tab, sidebar header, login page, and all system emails (MFA, password reset, invitations)
• Custom primary colors — separate pickers for light and dark mode with hex input for precision
• Custom CSS editor — full CodeMirror editor with syntax highlighting. Override any CSS variable in the system. A collapsible reference section shows every overridable variable

The logo is stored as base64 in the database (no filesystem dependency), and we cleaned up 894 hardcoded colors across the codebase to make theming actually work. That's a 76% reduction in hardcoded values — from 1,170 down to 276 (the remaining are brand-specific SVG colors for WhatsApp, Telegram, Slack, etc.).

TTS Voice Preview

The TTS settings panel now plays real Gemini voice samples. No more guessing what a voice sounds like — hit play and hear it before you assign it.

UX: Panel Cross-Closing

All sidebar panels are now mutually exclusive. Opening any panel closes all others. No more stacking sheets on top of each other.

v2.0.3 — February 18, 2026

UI Sketcher Expansion

For teams using the UI Sketcher skill to generate interface mockups:

• Dark theme support — all presets and components render in dark mode
• 16 new components — icons (80+ SVGs), progress bars, spinners, avatars, badges, ratings, skeleton loaders, search bars, file uploaders, steppers, accordions, timelines, and more
• 8 chart components — bar, line, pie, donut, area, sparkline, gauge, and stat-with-chart
• 8 new presets — profile page, pricing page, analytics dashboard, error page, kanban board, calendar, file manager, and email inbox
• Auto-install — UI Sketcher now installs automatically for new employees

Multi-Language Memory

The memory system now understands "recent" in 15 languages. Ask your AI employee for "derniers souvenirs" in French, "letzte Erinnerungen" in German, or "最近の記憶" in Japanese, and it correctly sorts results by date.

Supported languages: English, French, Spanish, German, Italian, Portuguese, Dutch, Russian, Chinese, Japanese, Korean, Arabic, Hebrew, Hindi, and Turkish.

URL Detection Fixes

Raw URLs and file paths in chat messages are now properly detected and clickable — including inside markdown-formatted messages, which previously missed them entirely.

v2.0.2 — February 16, 2026

18 Bug Fixes

The biggest bug fix round of the batch. Highlights:

The one that was breaking everything: The markdown rendering engine had a single character difference that was destroying all formatting in long reports — join(' ') instead of join('\n'). Every newline in every long-form response was being replaced with a space. If you noticed your AI employee's detailed reports looking like one giant paragraph, that was it.

Chat reliability:

• Messages no longer disappear on page reload (cache filter was excluding Claude sessions)
• Reports with ASCII diagrams render as a single collapsible block instead of being split into multiple sections
• Context summaries render as a single card with proper formatting
• Indented code blocks now detected correctly
• Code blocks display properly in PDF exports

Mobile UX:

• Removed the dropdown menu on mobile chat input — all icons now directly visible
• Fixed horizontal scrollbar with dynamic viewport widths (75vw on mobile, auto on desktop)

Other fixes:

• File links in chat now open in the File Explorer instead of downloading
• Folder navigation loads in parallel for faster file browsing
• OAuth URL cleanup (was containing terminal garbage characters)
• Admin invite emails fail gracefully with a manual invite link fallback
• Input textarea disabled for 4 seconds after WebSocket connection to prevent typing before the connection stabilizes

What's Next

We're shipping weekly. The roadmap includes deeper integrations, more voice options, and continued reliability improvements across the board.

If you're already on Geta.Team, update with a standard Docker rebuild. All database migrations run automatically.

Want to test the most advanced AI employees? Try it here: Geta.Team