Added

• new tool to access screenshares and video snapshots inside the meeting for vision-enabled models (#128, #131)
• client changes to allow multimodal tool results (see below)

Improvements

• reduce deepgram transcription latency (#132)
• add speech event delay to avoid unwanted interruptions (#133)
• improve hybrid VAD (#134)

joinly-client (v0.1.15)

• new and more detailed system prompt and instructions (#116, #135)
• support for google as LLM provider (#118)
• include option for tool loading (#120)
• custom post tool-call callbacks (#123, #125)
• improved tool result parsing including capabilities for multimodal results (#126)
• message history filtering for large tool results (#127)
• update README example (#130)

Improvements

• reduce audio format conversions in pipeline (#113)
• hybrid VAD implementation between webrtc and silero (#114)

Fixed

• unmute automatically on start (#112)
• round start/end times in transcript for consistency (#109)
• keep webrtc package in lite build (#111)

joiny-client (v0.1.8)

• add direct list_tools and session access to joinly client (#104)
• split system prompt and custom instructions (#105)
• use one mcp cient per server, instead of stateless proxy (#106)
• add message history limit to reduce token usage (#107)
• change date formatting to ISO (#108)

Others

• refactor to use type keyword (#103)

Improvements

• upgrade kokoro-onnx to 0.4.9 (#88)
• add model improvemet program opt-out for deepgram (#90, #97)
• default to --client if meeting url is given (#93)
• set default logging to info level for cli (#96)
• remove suffix for platform error messages (#98)

Fixed

• fix missing links in google meet chat (#86)
• get headers to load settings for fastmcp>=2.11 (#89)

joinly-client (v0.1.7)

• fix openai issues with openai>=1.99.2 (#91)
• fix gpt-5 usage by disabling sampling settings (#94)
• wrap tool calls for direct usage with client (#99)
• add debug logging for agent (#100)
• improve agent for gpt-5 (#101)

Others

• update environment names in release workflow (#87)

New: joinly-client (v0.1.5)

• complete rewrite of all client functionality, now using a single joinly-client package (#84)
• support for usage in code and via cli

Added

• LLM token usage tracking for a session (#75)
• STT/TTS API usage tracking, characters or audio minutes (#80)
• setting session configurations via joinly-client

Improvements

• event bus for transcript updates (#65)
• live segment transcription at segment level (#63)
• improve lite image by removing unnecessary dependencies (#79)
• unify logging of live segments (#82)
• --env-file cli option to specify a custom environment file (#74)

Fixed

• set explicit logging level only for own logger (#78)
• correctly fail on unsupported meeting platforms (#77)
• fix fatal server crash on session errors (#64)

Others

• add docker-outside-of-docker to devcontainer (#76)
• improve release workflow (#70, #73)
• cleanup repository root files (#72)

Improvements

• add health check endpoint to MCP server (#66)
• improve default voice selection for ElevenLabs TTS (#68)
• adapt logging levels for less noise in the logs (#69)
• update release workflow pipeline and update cuda image tag (#70)
• add lite image variant without local model weights (#71)

Fixed

• mark zoom waiting room as a successful join to fix potential timeouts (#67)

Improvements

• allow setting session-specific settings from the client (e.g., which STT/TTS), this will be further improved in the next release with client improvements (#57)
• remove redundant leave on exit (#59, #62)
• remove browser agent (#56)

Fixed

• zoom additional passcode handling (#61)
• deepgram misses first word (#60)
• resource subscribe flow (#58)
• enforce maximum message length (#55)
• make opening menu panels more robust and remove deprecated timeout (#54)
• change last segment timing to start in example (#53)
• simplify chat timestamps (#52)
• await meeting provider join before initializing transcript (#51)

Added

• ElevenLabs TTS support via --tts elevenlabs (#47)
• new setting --lang <language_code> to set the language for TTS and STT (depends on support of services) (#46)

Improvements

• streamline speech controller implementation (#41)
• improve error handling and interrupts in speech controller (#41)
• force leave by closing page on failed leave action (#35)
• auto leave on session tear down (#38)
• set docker logging to plain format (#39)

Fixed

• handle exceptions during agent invocation (#45)
• log speech-to-text exceptions (#44)
• ensure aligned segment timestamps in transcript (#40)
• fail on failed deepgram connection (#37)
• propagate ProviderNotSupportedError (#36)
• stop adding a segment for an interrupted speech without any spoken text (#48)
• fix no new segment error due to compact transcript after interruption (#49)

Added

• add get_transcript tool for fetching the meeting transcript with timestamp filters (#21)
• real-time speaker attribution for the transcript, in core app and all platforms (#27)
• new tool get_participants to retrieve the current meeting participants with available meta-data (e.g., host, muted/unmuted) (#28)

Improvements

• better internal meeting time measurement with more accurate start and end times (#18)
• shared meeting clock object for synchronized internal time handling (#22)
• add speech through speak_text tool to the meeting transcript (not included in transcript://live resource, but in get_transcript) (#23, #24)
• length-based TTS pre-chunking for better performance with long texts (#25)
• more compact transcripts by merging nearby segments of the same speaker for better LLM handling (#26)
• browser action improvements, for more robustness and some fixes (#30, #33, #34)
• teams live platform support using the existing teams platform actions (#31)

Fixed

• fix leftover audio in deepgram TTS after interruptions (#19)
• fix rare case where a update notification without new transcript segments crashes the client (#20)
• allow join_meeting after failed join attempts, which previously caused issues (#32)

Added

• add CUDA support for Whisper models and respective Docker build, which can significantly speed up transcription and allows usage of models like distil-large-v3 (default for cuda) (#10)
• add MCP tool get_chat_history for accessing the current meeting chat (#13)

Improvements

• change default Whisper model for CPU to base.en for better quality while stying near real-time (#11)
• change MCP tool response on detected interruptions while speak_text to a text response instead of an error (#12)

Fixed

• fix stuck whisper initialization (#4)
• fix no-speech event set on start, which caused a speak_text before any audio to be stuck (#7)
• fix and add missing mute/unmute actions (#5)
• fix action errors in google meet with multiple participants (#6)