用ElevenLabs CLI管理语音代理

Command Line Interface to interact with ElevenLabs voice agents

ElevenLabs CLI — Agents as Code Command-line interface for the ElevenLabs platform. The CLI does two things: Full API access — every ElevenLabs API endpoint is available as a subcommand (elevenlabs <resource> <method>). Agents as Code — manage Conversational AI agents from local configuration files, with templates, branches, and push/pull sync. Table of contents Installation Authentication Quick start Speaking text Waiting for Flows results Agents as Code Data residency UI components Usage Documentation Advanced Common flags Environment variables Telling us what you are doing Telemetry Output formats Shell completion Development Installation Homebrew (macOS / Linux) brew install elevenlabs/tap/elevenlabs Scoop (Windows) scoop bucket add elevenlabs https://github.com/elevenlabs/scoop-bucket scoop install elevenlabs npm (macOS / Linux / Windows) npm install -g @elevenlabs/cli Installs the same binary through a thin launcher, picking the right build for your platform. Use npx @elevenlabs/cli <command> to run it once without installing it globally. Shell (macOS / Linux) curl --proto '=https' --tlsv1.2 -LsSf https://github.com/elevenlabs/cli/releases/latest/download/elevenlabs-cli-installer.sh | sh PowerShell (Windows) powershell -ExecutionPolicy ByPass -c "irm https://github.com/elevenlabs/cli/releases/latest/download/elevenlabs-cli-installer.ps1 | iex" Build from source Install the Rust toolchain, then: git clone https://github.com/elevenlabs/cli.git cd cli cargo build --release The binary lands at ./target/release/elevenlabs. To run it as elevenlabs from anywhere, install it with cargo, which copies it into ~/.cargo/bin: cargo install --path . Authentication Every request is authenticated with an ElevenLabs API key, sent as the xi-api-key header. Set it in the environment: export ELEVENLABS_API_KEY=xi-... A .env file in the working directory is loaded automatically, so putting ELEVENLABS_API_KEY=xi-... there works too — handy for keeping a key scoped to one project. For a one-off call, pass it per command instead: elevenlabs user get --xi-api-key xi-... Quick start List available commands: elevenlabs --help Call an API endpoint: elevenlabs <resource> <method> Run elevenlabs <resource> --help to see available methods for a resource. Speaking text Turn text into speech and play it, without picking a file path or finding a player: elevenlabs say "this came from the terminal" echo "build finished" | elevenlabs say # or: elevenlabs say - The voice, model, audio format and player default to whatever you have stored, and each can be overridden per run: elevenlabs say config # show the current defaults elevenlabs say config voice 21m00Tcm4TlvDq8ikWAM elevenlabs say config model eleven_multilingual_v2 elevenlabs say config player mpv elevenlabs say config voice --unset # back to the built-in default elevenlabs say "one off" --voice JBFqnCBsd6RMkjVDRZzb --model eleven_flash_v2_5 # faster, no audio tags elevenlabs say "save it" --output out.mp3 # write a file, skip playback Defaults are stored in ~/.elevenlabs/config.json alongside data residency. Out of the box say uses eleven_v4 — the most expressive model, and (with v3) the only family that reads audio tags like elevenlabs say "[whispers] it worked" as delivery rather than as words. The audio format follows the player unless --output-format or an --output filename says otherwise. Playback shells out to whichever player is on the box, preferring ones that read stdin so audio starts before the download finishes: ffplay, mpv, afplay (macOS), paplay/aplay (Linux), Media.SoundPlayer (Windows). Install ffmpeg or mpv if none are present, or point --player at your own. config is a subcommand, so speaking that exact word needs elevenlabs say -- config. Waiting for Flows results Flows image, video and text-to-speech generations, and template runs, finish in the background: create returns straight away with status: pending. create-and-wait takes the same flags as create, then polls until the result is completed or failed and prints it (the same JSON as get): elevenlabs flows image create-and-wait \ --json '{"model_id": "bytedance-seedream-5-lite", "prompt": "a corgi on a beach"}' elevenlabs flows templates list # find a template and its input ports elevenlabs flows templates runs create-and-wait \ --template-id <template_id> --inputs '{"prompt": "a corgi"}' --wait-timeout 600 elevenlabs flows video create-and-wait --json - --query content_url < request.json It waits between checks for as long as the server's Retry-After header asks, never less than --min-poll-interval (default 3s). Progress is written to stderr when stderr is a terminal or with --debug; stdout only carries the final result, so --format, --query and pipes behave as they do for get. failed is printed like any other result, and the command exits 1. --wait-timeout <seconds> gives up after that long (default: wait indefinitely). It exits 1 with an error whose reason is waitTimeout, carrying the id and the get command to resume with. The work keeps running on the server. --dry-run prints the create request and sends nothing, as on create. For production, prefer setting webhook on create and handling the webhook event over holding a process open. Agents as Code Manage Conversational AI agents from local configuration files. elevenlabs agents init scaffolds a project; agent configs live as JSON on disk and sync to ElevenLabs. Pulled configs are stored as raw wire JSON and pushed back verbatim, so they round-trip losslessly. Project structure your_project/ ├── agents.json # Agent registry: ids + branch mappings → config paths ├── tools.json # Tool registry ├── tests.json # Test registry ├── agent_configs/ # Agent configuration files ├── tool_configs/ # Tool configuration files └── test_configs/ # Test configuration files Commands # Scaffold a new project (pass a path, or --override to reset an existing one) elevenlabs agents init [path] [--override] # Create an agent from a template (or an existing f