Nov 21, 2026

Slates - Your every agent in one workspace, AI employees on demand, and a team that works with you.

personalised agents ai employees always up agents multi agent system mcp server

Problem: AI can now do real work, but every agent lives in its own app and remembers only itself. You have to babysit chats, agents can't reach you when they need a decision, and you can't see, audit or undo what they did. Small business owners can't hire a team, and today's "AI employees" just wait in a chat box.

What Slates is: Slates is the headquarters for your AI agents.

1. One workspace for every agent. Claude, ChatGPT, Gemini, Codex and your own agents connect through Slates' MCP server and REST API, with one-click OAuth connectors and per-agent tokens. Agents build 14 kinds of living boards for you, from dashboards and kanban to trackers, habits, goals and charts. Boards are grouped into Spaces alongside notes, an inbox and a generated Today screen showing what needs you and who changed what.
2. Agents that work as a team:
   - The Desk: you leave instructions, agents hand work to each other and ask you questions.
   - Shared memory: what one agent learns about you, they all know.
   - Playbooks: your own way of doing recurring jobs, which every agent follows.
   - Manage your agents: set scopes and revoke any of them instantly.
   - Full history: every change is versioned, with undo.
3. It works while you're away: automations (morning brief, recaps, feeds), push reminders, the Ask Slates assistant, and the Slates runner, which actually runs your agents with spend caps, a real browser, GitHub tools and a live "Agents at work" board.
4. AI employees with voices:
   - Agent Studio hires you an AI employee in seconds. Say "manage my dairy business" and you get a Dairy Manager with its own boards, routines and Murf voice.
   - Call Slates and talk: it acts on your workspace live.
   - Standup puts your whole AI team in one Agora voice call. Each agent reports in its own voice, and your real Claude Code, Codex and Antigravity answer live. They hear each other, jump in and yield when interrupted.
   - When an agent needs a decision, Slates calls you (in-app call, real phone, or push). You say "go ahead" and it's done.

How we used Agora: Conversational AI Engine is the voice of Slates:
- managed speech recognition and LLM, with Murf voices
- llm.mcp_servers, so voice agents use Slates' tools mid-call
- llm.tools: a custom ask_agent tool that brings real coding agents into the call
- one speak-only agent per voice, making a multi-voice room
- speak / interrupt / history for floor control and live captions
- per-agent tokens, and agents limited to their own boards
- Web SDK with echo cancellation, noise suppression and voice barge-in

How we used Murf: Falcon voices give every agent its own identity (Indian, American and British English, chosen by role and region). They're streamed through Agora in every call, and Murf's streaming API generates each new employee's spoken introduction.

Tech stack: Next.js 16, Supabase (row-level security, realtime), PWA with web push, Slates MCP server and REST API (OAuth 2.1), Agora Conversational AI and Web SDK, Murf Falcon, Node runner (Claude Code / Codex / Antigravity CLIs, Playwright), Twilio, Gemini, Cloudflare Tunnel.

Guardrails:
- Per-agent scoped tokens, plus approval before archiving or deleting.
- Versioned history with undo.
- Every outbound call says it's an AI.
- Spoken decisions count only once confirmed; negations are never a yes.
- Call tokens expire, are limited to the agent's own boards, and carry no destructive tools.
- Signed webhooks and call contexts.

Challenges:
- Multi-party voice turn-taking. Nothing open-source solved it, so we built floor control on Agora's interrupt and speak APIs.
- Streaming speech fooled "finished talking" detection.
- Speaker echo triggered false interruptions.
- Staying honest: agents that can't truly join a call are narrated by Slates, never faked.

What's next: Agora SIP phone calls, real-time captions over RTM, a cloud-hosted runner, A2A so other companies' agents can join calls, and an agent template marketplace where anyone can publish an AI employee others hire in one tap.

This build was uploaded as a hackathon project

Hackathon

Build with Agora - BLR

View All Projects
Give a star to encourage!Discussion
Start a new conversation!
Login to join the discussion