The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the LiveReacting listing page.
Run live video streams from an AI agent. Build a channel from a video playlist, start it, schedule it, watch it while it runs, and read the numbers afterwards — from Claude Code, Codex, Cursor, or anything else that speaks MCP.
Streamable HTTP. Stateless. Authenticated with your LiveReacting API key.
Full reference, with every tool, input, response field, limit and error code: developers.livereacting.com/mcp (REFERENCE.md in this repo).
API access is included with every paid plan. Create a key in the
Developer section of the LiveReacting Studio.
Keys start with lr_.
With an API key, an account on the Free plan gets 403 API_ACCESS_REQUIRES_PAID_PLAN on every
call. With OAuth (ChatGPT, Claude), the account can connect, and each tool call returns an error
that explains the plan limit. The tools start working as soon as the plan includes API access,
without connecting again.
Create the key first, put it in your shell profile as LIVEREACTING_API_KEY, then paste the
config below.
Every example that cannot read that variable uses lr_… as a placeholder. Do not paste a
real key into a file you commit. Where a harness supports an environment variable or a
prompt, the block uses it.
In .mcp.json. Claude Code expands ${VAR} in a remote server's headers
(docs), so the key stays out of the file:
claude mcp add --transport http livereacting https://mcp.livereacting.com/mcp --header "Authorization: Bearer lr_…" also works, but it writes the key itself into the stored config.
In ~/.codex/config.toml:
Then export LIVEREACTING_API_KEY=lr_… in your shell profile. Prefer this over
http_headers, which puts the key in the config file.
In ~/.cursor/mcp.json. Cursor resolves ${env:NAME} in headers
(docs):
Or install the LiveReacting plugin from the Cursor Marketplace
and set LIVEREACTING_API_KEY under Plugins → Configure. The plugin is
.cursor-plugin/plugin.json and mcp.json in this repo.
In .vscode/mcp.json. The inputs block makes VS Code prompt for the key instead of
storing it in the repo:
Install the extension. Gemini CLI asks for the key once and keeps it in the system keychain:
Or, without the extension, in ~/.gemini/settings.json. Gemini CLI expands $VAR and ${VAR} in every string of
the file before it validates it, headers included
(docs), so the key stays out of it:
Install this repo as a power: Powers panel → Add Custom Power → Import power from
GitHub, then paste https://github.com/LiveReacting/livereacting-mcp. The power reads
LIVEREACTING_API_KEY from your environment (mcp.json).
In your Zed settings.json, under context_servers, add an HTTP server pointing at the
URL with an Authorization header.
Add a Remote Extension (Streamable HTTP) in Goose's settings, URL
https://mcp.livereacting.com/mcp, with a custom header
Authorization: Bearer lr_….
| Harness | How the key is passed |
|---|---|
| Claude Code | ${VAR} in .mcp.json headers |
| Codex CLI | bearer_token_env_var |
| Cursor | ${env:VAR} in mcp.json headers |
| VS Code Copilot | headers with a prompted input |
| GitHub Copilot coding agent | static header secret only. This product does not support OAuth remote servers, so the key path is the only path here, now and later |
| Gemini CLI | $VAR in settings.json headers |
| Kiro | ${VAR} in the power's mcp.json headers |
| Zed | headers in settings.json |
| Goose | remote extension header |
| ChatGPT | connectors take OAuth or nothing, never a pasted key |
Protocol revisions: 2024-10-07 through 2025-11-25. If your client reports a version
error, that is the first thing to check. 2026-07-28 is not supported yet — no published
TypeScript SDK speaks it.
30 tools in five toolsets. Every name is prefixed livereacting_, so one search in a
client that defers tool loading finds the whole server.
| Tool | What it does |
|---|---|
list_destinations | Your connected YouTube, Facebook, Twitch, LinkedIn, X and custom RTMP destinations |
validate_destinations | Checks a destination's connection still works, before you rely on it |
| Tool | What it does |
|---|---|
list_projects | Your projects |
get_project | One project's configuration |
create_project | A new project |
update_project | Change a project's settings |
get_scenes | Scenes in a project, or one scene |
manage_scene | Create, rename or delete a scene |
activate_scene | Switch the active scene, including during a broadcast |
create_scene_with_playlist | A scene plus its video playlist in one call |
add_layer | Add one video playlist, audio playlist or single video layer to an existing scene |
get_layer | One layer, with its playlist truncated for context |
set_playlist | Replace a playlist's contents |
| Tool | What it does |
|---|---|
start_stream | Go live |
stop_stream | End the broadcast |
get_stream_status | Live state, per destination, including expired connections |
schedule_stream | Go live at set times |
cancel_schedules | Cancel scheduled broadcasts |
sync_stream | Push editor changes to a running stream |
list_broadcasts | Past broadcasts for a project |
| Tool | What it does |
|---|---|
list_media | Files and folders in your library |
get_media_file | One file's details |
upload_file | Start uploading a local file: returns part URLs and curl commands |
complete_upload | Finish an upload once every part is sent |
import_media | Import from Google Drive, Dropbox or YouTube |
import_from_url | Import from a direct link to a media file on any public host |
get_import_status | Poll an import, including every video of a playlist import |
encode_media | Re-encode a file for streaming |
get_encode_status | Poll an encode |
| Tool | What it does |
|---|---|
get_stream_analytics | YouTube analytics, or per-video playlist analytics (that type on request) |
Both are query parameters on the URL, so no client support is needed.
read_only=1 drops every tool that can change anything. Useful for a monitoring agent you
do not want touching a live broadcast.
An agent that promises any of these is lying to your customer.
upload_file returns signed part URLs
and ready curl commands; the agent sends the bytes itself, then calls complete_upload.150 requests per minute per account, shared across your API keys.
REST and MCP are counted separately, so the two surfaces get 150 each rather than sharing one budget. Do not plan around that: it is an implementation detail of how the two are metered, not a promise.
A 429 comes back as a tool error that says the number and how long to wait, because the model
is the one reading it. The polling tools — get_stream_status, get_import_status,
get_encode_status — each document a minimum interval. Poll tighter than that and you spend
your own budget on nothing.
set_playlist replaces the whole playlist. It is not an append. To add one video, read the
layer first, then send the complete list back. The tool tells you when a write shrinks the
playlist, but it will still do what you asked.
Starting and stopping a stream is not idempotent. If a call times out, ask
get_stream_status what happened. Do not retry — a retry can start a second attempt.
Full MCP reference: developers.livereacting.com/mcp Full REST API reference: developers.livereacting.com Support: hello@livereacting.com Privacy policy: livereacting.com/privacy-policy