MeetStream for Hermes

Let your Hermes agent join, talk in, and summarize live meetings.

Your Hermes Agent sends a MeetStream bot into Zoom, Google Meet or Microsoft Teams, chats in the call, speaks through a MIA voice agent, and pulls summaries and transcripts afterwards. It does this with 19 tools and 6 skills over the hosted MeetStream MCP server.

imageDocs imageMCP imageSkills imageLicense

Install · Try it · Demo · Architecture · Security · Troubleshooting

You run nothing extra. No server, no bridge, no audio pipeline. The plugin is a declarative Agent Plugins v1 package that Hermes loads natively. It contains one MCP connection to https://mcp.meetstream.ai/mcp, six skills, and your own API key read from Hermes' environment.

Install

New to this? INSTALL.md is the full step-by-step guide. It covers the terminal and the desktop app, setting up a voice, your first real meeting, and troubleshooting, with a check after every step.

Prerequisites: Hermes Agent (terminal or desktop app) with a model configured (hermes model), and a MeetStream API key from app.meetstream.ai/api-keys.

# 1. Install and enable the plugin
hermes plugins install meetstream-ai/hermes-plugin --enable
#    desktop alternative: open hermes://plugin/install?repo=meetstream-ai/hermes-plugin&enable=1

# 2. Give Hermes your MeetStream key (saved to ~/.hermes/.env, never to config.yaml)
hermes config set MEETSTREAM_API_KEY ms_your_key_here

# 3. Start a NEW session so the MCP tools and skills load
hermes                     # terminal; in the desktop app: quit, reopen, click New chat

Then ask: "List my MeetStream bots." A clean answer means the key, the plugin and the MCP connection are all working.

Important

Start a new chat after installing or setting the key. Hermes loads MCP tools only when a session starts. An already-open chat answers "I don't have a MeetStream tool" even when everything is configured correctly.

Note

• Hermes installs this type of plugin disabled. If you installed without --enable, run hermes plugins enable meetstream.
• In the desktop app the plugin is listed under Capabilities → Plugins → Installed.
• The app has no screen for the MeetStream key; Settings → Providers is for model keys only. Set it with hermes config set as shown above. The app and the terminal share the same settings.

What you can ask

Plain language. Hermes picks the tools, and the skills tell it how to use them well.

"Join my meeting at https://meet.google.com/abc-defg-hij and introduce yourself."

"Is the bot in yet? Post 'Recording started - say stop if you'd rather not' in the chat."

"Who's in the call right now, and what have people said in chat?"

"Summarize that meeting. Give me decisions and action items with owners."

"Send a bot to my 3pm on my calendar."

"Make the bot leave."

Speaking in the meeting

The MeetStream MCP server has no "say this sentence" tool. Speech comes from a MIA voice agent, which MeetStream hosts, and you attach one with a single agent_config_id on create_bot. When the bot is admitted, the agent speaks its configured first_message, such as an introduction. After that it talks with participants using its own model and prompt.

Create a voice agent once with the official CLI and the ready-made config in this repo:

export MEETSTREAM_API_KEY=ms_your_key_here        # the CLI reads the key from your shell
npx -y @meetstream/cli mia create -f examples/mia-intro-agent.json
# -> prints JSON containing "agent_config_id": "..."

Important

Attach the agent with only agent_config_id. Ignore the CLI's follow-up hint about --socket-ws / --live-audio-ws. Those flags are for self-hosted bridges, and adding them to a hosted MIA bot is the most common cause of an agent that joins but never speaks. The speak-in-meeting skill enforces this.

examples/mia-intro-agent.json uses the provider set already verified live against MeetStream: pipeline mode, OpenAI gpt-4o with tts-1/alloy, Deepgram nova-3, 48 kHz mono. Edit first_message to change what the bot says on arrival. Model and voice providers must be enabled on your MeetStream account.

Real-meeting demo

The full chain is Hermes agent → MeetStream MCP → MeetStream bot → real meeting → bot speaks.

git clone https://github.com/meetstream-ai/hermes-plugin.git && cd hermes-plugin
hermes plugins install meetstream-ai/hermes-plugin --enable
hermes config set MEETSTREAM_API_KEY ms_your_key_here
export MEETSTREAM_API_KEY=ms_your_key_here          # only needed to auto-create the voice agent

# Start a Google Meet (or Zoom/Teams) call yourself, then:
scripts/demo.sh 'https://meet.google.com/abc-defg-hij'

scripts/demo.sh does the following:

1. Checks that Hermes is installed, the plugin is enabled and MEETSTREAM_API_KEY is configured.
2. Creates the MIA voice agent from examples/mia-intro-agent.json via @meetstream/cli. Skip this by passing an id as the second argument or setting MEETSTREAM_MIA_AGENT_ID.
3. Runs one Hermes turn (hermes -z) with the join-meeting, speak-in-meeting and meeting-chat skills preloaded. Hermes calls create_bot with the voice agent attached, polls get_bot_status until the bot is admitted, then calls send_chat_message.

Verified end to end on 2026-10-06 against a live Google Meet call: the bot was dispatched, admitted, reached InMeeting, and posted its chat greeting.

Your part: when a participant named Hermes appears in the waiting room, admit it (you have about 5 minutes). You should then hear the introduction and see a hello in chat. Clean up with hermes -z "Remove my MeetStream bot from the meeting at <url>".

You can also run the demo by hand in an interactive hermes session:

Join my meeting at https://meet.google.com/abc-defg-hij using MeetStream agent_config_id <id> and introduce yourself. Once you're admitted, say hi in the chat too.

Architecture

 You ──prompt──▶ Hermes agent (model, tool choice, conversation state)
                    │  mcp__meetstream__create_bot / send_chat_message / get_transcript …
                    │  Streamable HTTP, Authorization: Bearer <your key, read at connect time>
                    ▼
            mcp.meetstream.ai/mcp  (hosted, multi-tenant, stores no keys)
                    │  MeetStream REST API
                    ▼
            MeetStream bot + MIA ──joins, records, chats, speaks──▶ Zoom / Meet / Teams

Package layout

INSTALL.md                   Step-by-step install and first-meeting guide
plugin.json                  Agent Plugins v1 manifest (name: meetstream)
mcp.json                     One streamable-http server -> https://mcp.meetstream.ai/mcp
skills/<name>/SKILL.md       6 skills, loaded via skills_list / skill_view
examples/mia-intro-agent.json  MIA voice agent config for the demo
scripts/demo.sh              End-to-end real-meeting demo
scripts/check.py             Repository checks (manifests, skills, tool names, secrets)

Hermes discovers plugin.json and loads it through its Agent Plugins v1 adapter. That adapter is Hermes' only native mechanism for a standalone plugin repo to ship an MCP server, which is why this repo does not use plugin.yaml with Python. The server keeps its mcp.json name, so tools appear as mcp__meetstream__<tool>. Skills are namespaced as agent-plugin-meetstream-ee716565:<skill>.

The 19 tools

These come from the hosted server, not from this repo, so they stay current automatically.

Full reference: MeetStream MCP server.

Skills

The skills are adapted from the MeetStream Cursor plugin and trimmed to what the MCP tools actually support. REST-only features are named as REST-only, never presented as tools. That includes creating MIA agents, connecting calendars and signed-in bot accounts.

Relationship to the MIA × Hermes bridge

labs/mia-hermes-bridge solves the opposite direction, and the two are complementary:

Use this plugin to send, control and debrief bots from your Hermes agent. Use the bridge when Hermes itself should be the participant who talks. Neither reimplements the other, and neither reimplements MeetStream transport.

Your API key

• Yours, per user. Hermes interpolates ${MEETSTREAM_API_KEY} into the Authorization header at connect time, from your environment or ~/.hermes/.env. It is sent with each request to the hosted MCP server.
• Never stored server-side. mcp.meetstream.ai is multi-tenant and stateless. It builds a fresh client per request from the header and keeps no key database (source).
• Never in this repo. mcp.json contains only the ${MEETSTREAM_API_KEY} placeholder. scripts/check.py scans every file for key-shaped strings.
• Fails closed. If the variable is unset, Hermes refuses to connect and names the missing variable. It never sends a literal ${…} header.
• Revocable. Each person should use their own key so it can be revoked individually at app.meetstream.ai/api-keys.
• Profiles. Under a multi-profile Hermes gateway, each profile resolves the key from its own .env.

Prefer the key never leaving your machine? Skip the plugin's hosted connection and run the same server locally over stdio in ~/.hermes/config.yaml. A config.yaml server named meetstream takes precedence over the plugin's:

mcp_servers:
  meetstream:
    command: npx
    args: ["-y", "@meetstream/mcp"]
    env:
      MEETSTREAM_API_KEY: "${MEETSTREAM_API_KEY}"

Config-only setup without the plugin (tools but no skills)

# ~/.hermes/config.yaml
mcp_servers:
  meetstream:
    url: "https://mcp.meetstream.ai/mcp"
    headers:
      Authorization: "Bearer ${MEETSTREAM_API_KEY}"

Verify the installation

hermes plugins show meetstream            # Status: enabled
hermes config env-path                    # the .env that should contain MEETSTREAM_API_KEY
hermes -z "Call the MeetStream list_bots tool and tell me how many bots I have."

Inside a session, ask "Which MeetStream tools and skills do you have?" Hermes should list the mcp__meetstream__* tools and find the agent-plugin-meetstream-… skills through skills_list.

Troubleshooting

First look at what Hermes logged: grep -h meetstream ~/.hermes/logs/*.log | tail -5. The full table is in INSTALL.md.

More: MeetStream errors · Debugging bots · Hermes MCP docs.

Development

git clone https://github.com/meetstream-ai/hermes-plugin.git && cd hermes-plugin

python3 scripts/check.py           # manifests, skills, links, secret scan, no-transport guard, hermes validate/doctor
python3 scripts/check.py --live    # + cross-check every tool the skills mention against the real server (Node 18+, no key needed)

hermes plugins validate .          # Hermes' own catalog gate
hermes plugins doctor --ci .       # Hermes' runtime contract check

To test local edits in Hermes, push them to a branch (or any git URL) and reinstall:

hermes plugins install https://github.com/<you>/hermes-plugin.git --force --enable

Rules for contributors are in AGENTS.md. The short version: no code that talks to meetings, no keys in files, every skill name matches its directory, and never describe a tool the MCP server does not expose.

The rest of the ecosystem

Built by MeetStream.ai · MIT