Tech 6 min read

Slack launches an MCP server, and Gemini adds AI music generation

IkesanContents

Slack and Google both pushed harder to integrate AI directly into their own platforms. In both cases the goal is not “use AI as another separate tool,” but “make AI part of the environment people already use.”

Slack: MCP server and real-time search API

On February 17, Slack announced two new features that let generative AI operate Slack more directly. They reached GA about four months after beta started in October 2025.

MCP server

The Slack MCP server endpoint is https://mcp.slack.com/mcp using JSON-RPC 2.0. AI agents can do four categories of operations:

search and discovery

  • search messages and files with filters for date, user, and content type
  • search users by partial name, email, or user ID
  • search channels in both public and private spaces

messaging

  • send messages to any conversation type
  • fetch full channel history and thread conversations

canvas management

  • create and update rich-format documents
  • export Canvas documents as Markdown

user info

  • fetch full user profiles, including custom fields and status

MCP, or Model Context Protocol, is the tool-access standard proposed by Anthropic. When major business tools like Slack provide official MCP servers, agent usefulness goes up another level. I have also written about Blender MCP, Google Developer Knowledge API, and Android MCP servers in the past.

supported clients and auth

The MCP clients confirmed to work are:

  • Claude.ai web, Claude Desktop, and Claude Code
  • Cursor
  • Perplexity

Authentication uses OAuth 2.0 confidential flow. A registered Slack app is required, and scopes are split very granularly. For example, public channel search uses search:read.public, private channel search uses search:read.private, DM search uses search:read.im, and message sending uses chat:write. Private channels and DMs require a user token (xoxp-), while bot tokens can only access public channels.

At the time of writing, SSE transport and Dynamic Client Registration are not supported yet. PKCE support is listed as coming soon.

real-time search API

Slack also announced a search-based API that lets external apps access live workspace data directly. Instead of wiring together webhooks and the Events API, you can now use one search API.

There are two methods: assistant.search.context for full-text message search, and assistant.search.info for workspace search capability metadata. Queries support OR-based alternative terms, filtering by content type or channel type, and retrieving surrounding context messages. Results are capped at 20 per page, and the rate limit is 10 requests per minute per user.

Semantic search is only available in workspaces that have Slack AI Search enabled, which usually means a paid higher-tier plan. External storage or copying of retrieved data is prohibited, and it cannot be used to train AI models.

adoption status

During beta, Slack says more than 50 partners built agents and RTS queries plus MCP tool calls increased 25x. Partners include Guru, Manus, Perplexity, Moveworks, ThoughtSpot, ChatGPT, and Google Agentspace.

OSS alternatives

There are community Slack MCP server implementations on GitHub too. Compared with the official version, the OSS projects sometimes support stealth browser-token connections, user-group management, and SSE transport. The official version, on the other hand, supports Canvas and has enterprise-grade audit logging and access control. For business use the official server is the clear choice, but OSS remains an option for personal use.

Google Gemini: Lyria 3-based music generation

Google integrated DeepMind’s music model Lyria 3 into the Gemini app. It launched on the web on February 18, with mobile rollout following shortly after.

technical characteristics

Lyria 3 combines a language encoder with a diffusion-based audio decoder and supports 48kHz stereo output. That is a major step up from the roughly 24kHz output common in earlier AI music tools. Generated tracks include musical structure such as intros, verses, choruses, transitions, and outros.

Compared with Lyria 2, the main improvements are:

  1. image and video prompts are now supported, not just text
  2. lyric generation is automatic
  3. developer API access is available through Vertex AI

You can generate a 30-second track from text, image, or video without leaving the Gemini chat UI. Style, vocals, tempo, and other musical elements can all be described in natural language. The result can also include lyrics and cover art generated by Nano Banana AI.

SynthID watermarking

All generated tracks include SynthID watermarking. The audio waveform is converted to a spectrogram and an imperceptible watermark is embedded using human hearing characteristics. The watermark survives common transformations such as MP3 compression, speed changes, or re-recording through a microphone. You can upload audio to Gemini or the SynthID Detector portal to check whether Google AI created it.

availability and limits

The feature is for Gemini users age 18 and up and is available even on the free plan, though with generation limits. It supports eight languages: English, German, Spanish, French, Hindi, Japanese, Korean, and Portuguese. It is still in beta and currently limited to 30-second tracks.

For YouTube creators it is being rolled out as Dream Track. The older Dream Track was limited to AI clones of participating artists’ voices; Lyria 3 shifts toward original vocal and instrumental generation and expands the feature to creators worldwide.

competition

itemLyria 3SunoUdio
track length30 secondsup to 2 minutesfull songs
inputtext, image, videotexttext
audio quality48kHzhighhigh
costincluded in Gemini subscriptionfreemium + paid plansfreemium + paid plans

Suno and Udio still lead if you want full songs, but Lyria 3 has the advantage of staying inside Gemini and YouTube, with cost bundled into the existing subscription.

Google says Lyria 3 was trained on music that Google and YouTube have the rights to use under partner contracts, terms of service, and applicable law. Still, Billboard previously reported that DeepMind had used a dataset that included copyrighted major-label recordings, so questions about rights handling remain. Universal Music Group and YouTube already signed a licensing deal in October 2024 that included AI guardrails, but no guarantee is provided that generated output will be dissimilar from existing songs.