MIA Screen Share, Bot Desktop and Your Own Agent
MIA can now do more than talk. In a Google Meet call, MIA can share its screen, open a web page and work with it the way a person at a keyboard would. It can present its own desktop with a spreadsheet, a document, a code editor or a terminal, and it can post in the meeting chat. You can also make your own agent (Claude Code, Codex, Hermes, GitHub Copilot CLI and others) the brain behind MIA. Your agent runs on your computer with your own login, and MIA handles joining the call, listening and speaking.
This guide covers:
- Screen tools: MIA presents a web page and reads, clicks, types, scrolls and looks at it.
- Bot desktop: MIA presents its own desktop and works in apps.
- Meeting chat and quiet behaviour: MIA reads the chat, acts without narrating, and speaks only when asked.
- Your agent: bring Claude Code, Codex and other agent CLIs into the meeting.
- Complete examples
- Limits · Troubleshooting · FAQ
Applies to: Google Meet for screen tools and the bot desktop. Google Meet, Zoom and Microsoft Teams for your own agent and meeting chat input. Support: docs.meetstream.ai · API: api.meetstream.ai
New to MIA? Start with the MIA Reference and the Create Agent Guide. This guide assumes you already know how to create an agent and send it to a meeting with
agent_config_id.
At a glance
Each of these is off unless you turn it on, so existing agents behave exactly as before.
How it fits together
MIA hears the meeting, turns speech into text, and passes each turn to the brain. The brain answers, and can call tools: screen tools, chat, your MCP servers and your custom functions. MIA carries out each tool call in the meeting and speaks or posts the reply.
1. Screen tools
With screen tools on, MIA can present a web page into a Google Meet call and use it like a person:
- “Mia, share the pricing page on meetstream.ai.”
- “Scroll down. What does the Enterprise plan include?”
- “Click Contact sales and type my email.”
- “Go to the next slide.”
- “Send that table in the chat.”
Turn it on
Add "screen" to agent.tools. The agent’s model must be able to see images (see Supported models).
In the dashboard: Agents → your agent → Tools → Screen tools.
Tip:
agent.toolsis the same list that holds the built-in tools. To keep them, list them together, for example["screen", "current_time"].
The tools MIA gets
The desktop tools screen_app_open and screen_write_file are covered in Bot desktop.
What a session looks like
How MIA picks pages
MIA opens only pages that someone named or pasted for it. It never opens a page it picked up from other people’s conversation, and it never chooses one on its own. This stops anyone in the meeting from making MIA present something it shouldn’t.
Posting to the chat
send_chat_message can post three things, together or apart:
- Text: MIA’s own words, such as a summary, an answer or a link.
- Page text (
include_page_text): the exact text of the page being shared, pasted as is. - Screenshot (
attach_screenshot): a picture of the shared page.
If the meeting’s chat does not accept images (for example, a guest bot in some Google Meet calls has a text-only chat), the text is still sent. MIA posts a link to the screenshot that stays valid for one hour instead. If the chat is restricted to the host’s organization, MIA is told that images are not possible.
Supported models
Screen tools need a model that can see images. The API rejects "screen" on other models.
When MIA uses screen_look or meeting_share_look, the picture is described with your own provider key (the same key the agent uses).
Example: create an agent with screen tools
Then send it to a Google Meet call:
2. Bot desktop
The bot desktop gives MIA its own computer screen to present. MIA can open apps, write and run code, fill in a spreadsheet, and switch smoothly between a web page and the desktop in the same Meet presentation.
- “Mia, open a spreadsheet and put in last quarter’s numbers: January 120, February 135, March 150. Add a total.”
- “Write a Python script that prints the first ten Fibonacci numbers and run it.”
- “Open a document and draft the three action items we just agreed.”
Turn it on
Set both agent.desktop: true and agent.tools: ["screen"]:
Bots for desktop agents are started on machines with the desktop installed, so no other setting is needed when you create the bot. Agents without desktop: true never get the desktop tools.
Desktop apps and tools
screen_app_open opens an app on the desktop (and starts presenting the desktop if needed):
screen_write_file writes a whole file at once into the ~/work folder and opens it in the code editor. MIA uses this to write code instead of typing it key by key, then runs it in the terminal.
screen_share can also start straight on the desktop: desktop_app set to calc, writer or editor, or desktop for an empty desktop.
All the other screen tools (screen_type, screen_key, screen_mouse, screen_drag, screen_click, screen_look, …) work on the desktop too. Mouse positions are in pixels of the 1280×720 screen.
What a session looks like
Safety of the desktop
- The desktop apps have no network access.
- The desktop is started only when MIA first presents it, and it is removed when the bot leaves.
- Everything runs inside the bot, which is deleted after the meeting. Nothing on your own computer is touched.
3. Meeting chat and quiet behaviour
Three optional switches change how MIA uses the chat and its voice. They work well with screen tools and with your own agent.
All three must be true or false when present.
chat_input: MIA reads the chat
Each chat message reaches MIA as “[chat from NAME]: text”.
- MIA replies or acts only when the message is for it: the message names MIA (the bot name or a wake word), or someone just told MIA out loud to expect something in the chat (“Mia, I’ll paste the link in the chat”).
- Every other message, including a bare pasted link, is added to the conversation as context and gets no reply. A pasted link is not a request, so nobody can make MIA present a page just by pasting it.
- When a request comes from the chat, MIA answers in the chat, not out loud.
quiet_actions: act without narrating
MIA stops saying “Okay, opening it now…” or “Done!”. After a successful action it stays silent unless it was asked to read, describe or answer something. It still speaks when an action fails.
Actions covered: share, stop, open, click, type, scroll, key, mouse, drag, open app, write file, and send chat message.
speak_only_when_asked: speak only when addressed
For a quiet assistant that works in the background:
- MIA hears every turn and acts on clear requests right away (“share the pricing page”).
- It speaks only when addressed by name: its name or a wake word in the turn, or a follow-up within 30 seconds of that. A question without its name (“can you hear me?”) gets no answer.
- “Mia, stop talking” (or “shut up”, “be quiet”) ends that window at once.
- It does not confirm actions out loud.
Requirements and cost:
- Pipeline mode only. The API rejects it in realtime mode.
- The wake-word gate is not used in this mode, so the model runs on every turn instead of only after the wake word. Expect more model usage.
- With this option on, screen and chat actions are refused unless MIA was addressed by name.
Example
4. Your agent: bring Claude Code, Codex and more
Make your own agent CLI the brain of MIA. Your agent keeps its model, its prompts, its tools and access to your files. MIA keeps the meeting side: joining the call, speech-to-text and voice.
- Your agent runs on your computer, signed in with your own subscription.
- MeetStream never sees your agent’s login and charges no model tokens for its replies.
- Everyone in the meeting can hear your agent answer, and it can use the meeting through MIA’s tools: post in the chat, share its screen and more.
Good uses:
- “Claude, what does the
create_bothandler do when the meeting link is invalid?” with your agent reading your own repository during a design review. - A standup where your agent answers questions about open branches and recent changes.
- A demo where your agent shares a page and walks through it while you talk.
Supported agents
A warm session keeps one agent process running for the whole meeting, so answers start in about a second instead of waiting for the agent to start on every turn.
How a turn works
- When the bot starts joining, the bridge starts one session of your agent for that meeting and warms it up with a short test turn, so no one waits later.
- When someone talks to MIA, the words go to your agent as text. The answer streams back, and MIA starts speaking the first sentence while your agent is still writing the rest.
- When someone interrupts, the bridge stops your agent at once and keeps the conversation, so the next question still has the context.
- When your agent calls a meeting tool (for example
send_chat_messageorscreen_share), MIA runs it in the meeting and sends the result back, and your agent continues the same turn. Wake words, quiet actions, speak only when asked, and interruptions all still apply.
What you need
- macOS or Linux, with Python 3.10 or newer.
- One supported agent, installed and signed in.
- A MeetStream account with a speech-to-text provider and a voice provider set up for MIA.
Step 1: Install and run the Agent Kit
The bridge is the open-source MeetStream Agent Kit (github.com/meetstream-ai/agent-kit). It uses only the Python standard library.
Or with pipx: pipx install meetstream-agent-kit. Or download the single file meetstream-agent-kit.pyz from the latest release and check it against SHA256SUMS:
The guided setup asks a few questions:
- Which agent should answer. It lists the agents it finds on your computer and checks that each is signed in.
- The folder your agent works in, for example your project’s repository.
- The most access this computer allows (see Access levels).
- How MeetStream reaches your computer:
- Cloudflare quick tunnel: no account needed. The address changes each time it starts.
- ngrok: needs a free ngrok account. A free static domain keeps one address.
- An address you already have, for example a named Cloudflare tunnel.
It runs one short test turn with your agent and asks Start now? On yes, it starts the bridge and the tunnel and prints the two values to paste into MeetStream:
Next time, run meetstream-agent-kit and press Enter to start with the same answers. Leave the window open while you want your agent available in meetings.
Prefer to run each piece yourself?
Use
--tunnel ngrok, or leave out--tunneland run your own, for examplecloudflared tunnel --url http://127.0.0.1:8787.
Step 2: Save the pairing token in MeetStream
In the dashboard, open Integrations → Models → Local agent, paste the pairing token and save. The token is shown only once on your computer and is stored encrypted by MeetStream, like any other provider key. It is never stored in the agent configuration.
Step 3: Point your MIA agent at your bridge
In the dashboard: open the agent, go to Model, choose Your agent as the provider, and paste the Bridge URL. Set the access level, the display label and the activation code, then click Test connection. Save the agent first: the test uses the saved settings.
With the API: use pipeline mode with model.provider: "local_agent":
model.modelis optional. It is passed to your agent’s own model setting (for examplehaikufor Claude Code). Leave it out, or usedefault, to keep your agent’s default.model.system_promptis optional. Your agent has its own instructions, and MIA adds its meeting rules.transcriberandvoiceare required: MIA still does the listening and speaking.
Step 4: Send the bot to a meeting
There is nothing else to do per meeting. Keep the bridge and the tunnel running.
Activation code
The activation code is on by default (require_code: true). Your agent answers nobody until you prove you are at your computer:
- When the bot joins, MIA posts in the meeting chat: “I’m off until activated. To use Claude Code, paste the 6-digit code it shows on your computer here in the chat.” (The name comes from your
label.) - The bridge shows a 6-digit code in its terminal and as a desktop notification. MeetStream never sees the code before it is pasted.
- Whoever pastes the code in the chat becomes the owner for that call. The owner is tied to their participant id, not their display name. Only the owner can start a turn. What others say still reaches your agent, as context with the owner’s next turn.
- If the owner leaves, MIA locks again and the bridge shows a new code.
Limits:
- Each code works once and is valid for 5 minutes.
- After 3 wrong codes a new code is shown. After 12 wrong codes MIA stays off for the rest of the call.
- Only chat messages that contain a 6-digit number are checked as codes.
To turn it off, set "require_code": false and list owner_names (or owner_ids) instead. Only those people can then start a turn.
Note: Display names can be copied by guests. Use the activation code for anything beyond reading.
Access levels
What your agent may do on your computer during a meeting. The levels mean the same for every agent; the bridge maps each one to that agent’s own controls.
Your computer has the final say. serve --max-access sets the highest level your computer allows. The setting in MeetStream can only lower it for an agent, never raise it. A leaked API key or a dashboard change can never give anyone more access than you allowed on your computer.
- Full needs an id-based owner: the activation code (default), or
owner_only: truewithowner_ids. Display names never count. Without that, the bridge drops the meeting to Ask me first, and the API rejectsaccess: "full". - Antigravity and custom commands cannot ask before acting, so Ask me first runs them read-only. A custom
--commandalways runs with your own permissions.
model.local_agent reference
Rules for the rest of the config:
modemust be"pipeline".transcriberandvoiceare required.model.modelandmodel.system_promptare optional.agent.tools: ["screen"]is allowed: screenshots fromscreen_lookandmeeting_share_lookgo to your agent as images.
Test the connection
The dashboard’s Test connection button calls:
When your agent’s login expires
- The login is checked when you pair, when the bridge starts, and each time a meeting starts.
- If it is missing or expired, the bridge shows a desktop notification with the one-line fix (for example
claude auth login). MIA sends anagent_errorwith codelocal_agent_login_requiredto your webhooks and the dashboard, and posts one chat line with the same fix once the meeting is live. Then it stays silent. - MIA checks again every 30 seconds and carries on by itself once you have signed in again.
- There is no fallback model. MeetStream never answers in your agent’s place.
Security
- The bridge listens on 127.0.0.1 only. Your tunnel is the only way in.
- Every request from MIA needs your pairing token. The token is shown once, and only its SHA-256 hash is kept on your computer (in
~/.config/meetstream-agent-kit, readable only by you).meetstream-agent-kit revoke <id>makes it stop working at once. - Meeting tools reach your agent on a separate secret for each meeting, never on a command line.
- Everything said or typed in the meeting reaches your agent as untrusted text, labelled by speaker, with standing rules: never reveal secrets, never run destructive commands, never act for anyone but the owner.
- Tokens, secrets and request headers are never logged.
- Your agent’s login never leaves your computer.
Agent Kit commands
serve, doctor and bench take --agent, --command, --binary, --workdir, --owner, --model, --effort and --timeout. serve also takes --port, --max-access (readonly, confirm or full), --no-prime (skip the warm-up turn) and --tunnel (cloudflare, ngrok, url or none) with --tunnel-url for an ngrok static domain or your own address. Every command has --help.
Tip for fast replies: a small model at low effort answers fastest, for example Claude Code with
--model haiku --effort low. Runmeetstream-agent-kit bench --agent claude --model haiku --effort lowto time your own setup.
5. Complete examples
Example A: Product demo presenter (screen tools, quiet)
A realtime agent that presents pages and walks through them, without narrating each click.
In the meeting:
Example B: Live coding on the bot desktop
A pipeline agent that writes and runs code on its own desktop.
In the meeting:
Example C: Quiet meeting helper that lives in the chat
Listens all meeting, acts on clear requests, speaks only when addressed.
Example D: Claude Code in your design review
Your own Claude Code, reading your repository, with the activation code and read-only access.
On your computer:
Save the printed pairing token under Integrations → Models → Local agent, then create the agent:
In the meeting:
6. Limits and good to know
Screen tools and desktop
- Screen tools and the bot desktop work in Google Meet today. On Zoom and Microsoft Teams, they are not offered to the agent.
- The shared screen is 1280×720. Mouse positions and screenshots use that frame.
- Starting a share or opening a page takes a few seconds. A click takes one to two seconds.
- The browser MIA presents runs separately from the meeting browser and can reach only public websites.
screen_typetakes up to 2,000 characters per call.send_chat_messagetakes up to 2,000 characters.screen_write_filetakes up to 200,000 characters.- If MIA’s presenting browser or desktop crashes, it restarts by itself, up to 3 times in 10 minutes.
- Meet decides what appears to the audience. If Meet ends the bot’s presentation, MIA stops sharing.
Your agent
- Pipeline mode only.
- Your computer, the bridge and the tunnel must stay on and online for the whole meeting. If they stop, MIA posts one chat line, sends an
agent_error, and checks again every 30 seconds. - The activation code works in Google Meet. In Zoom and Microsoft Teams, set
require_code: falseand useowner_names. - A Cloudflare quick tunnel gets a new address each time it starts. Update the Bridge URL in MeetStream when it changes, or use a named tunnel or an ngrok static domain for a fixed address.
- How fast your agent answers depends on your agent, its model and your computer. With a warm session, small models start answering in about one second.
Updating agents with the API
PUT /api/v1/miareplaces the wholeagentblock. When you update an agent, sendtools,desktop,chat_input,quiet_actionsandspeak_only_when_askedagain, or they are cleared.
7. Troubleshooting
meetstream-agent-kit doctor checks your computer and prints a fix for every problem it finds:
8. FAQ
Does MeetStream see my agent’s login, code or files? No. Your agent runs on your computer with your login, and the login never leaves your computer. Your agent sends MIA only the text of its answers and its tool calls.
Who pays for my agent’s model? You do, through your own agent subscription. MeetStream charges no model tokens for your agent’s replies. MIA’s speech-to-text and voice work as for any other agent.
Can someone in the meeting make my agent change files on my computer?
Not by default. Your agent starts read-only. Ask me first needs you to approve every action in the bridge terminal. Full needs your computer to allow it with --max-access full and the activation code, so only you can talk to your agent. Meeting speech and chat are treated as untrusted text.
Can MIA open any website someone mentions?
MIA opens only pages someone names or pastes for it, never pages it picks itself or overhears from other people’s conversation. With chat_input, a bare pasted link is context, not a request.
Does the bot desktop touch my computer? No. The desktop runs inside the bot, has no network access, and is removed when the bot leaves.
Can I use MCP servers and custom functions together with screen tools or my own agent? Yes. Screen tools are added next to your MCP servers and custom functions. Your own agent also gets them, through the bridge, along with the meeting tools.
Does my agent work in Zoom and Microsoft Teams? Yes, your agent can answer in Google Meet, Zoom and Microsoft Teams. Screen tools and the activation code are Google Meet only for now.
