Use anyscribe from your AI agent
You don't have to memorize flags. Hand a link to your agent and say "transcribe this" — it drives anyscribe for you.
scribeandascliare shorter aliases foranyscribe— every command on this page works typed as any of the three.
Two ways in
The Claude Code skill teaches Claude how to use the CLI. Claude runs anyscribe commands in your terminal the way you would, so anything the command line can do, the skill can reach: transcribing, batches, provider setup, log reading, updates, local model management. It's the fuller path, and it only works in Claude Code.
The MCP server exposes a fixed set of ten tools over MCP — the Model Context Protocol, the standard way an AI app talks to an outside tool. It's narrower by design, but it works in any MCP host: Claude Desktop, Cursor, Claude Code, and anything else that speaks the protocol.
Not sure which? If you live in Claude Code, install the skill. If you live in Claude Desktop or Cursor, use MCP. Installing both is fine — they don't conflict.
Claude Code (skill)
pip install anyscribe
anyscribe install-skill
That copies the skill into ~/.claude/skills/anyscribe/. Claude Code picks it up on the next session — no restart dance beyond opening a new one.
You may already have it. If Claude Code is installed on this machine, anyscribe installs the skill for you the first time you run any anyscribe command, and quietly refreshes it whenever the copy on disk falls behind the installed version. No prompt, nothing to opt into. Run anyscribe install-skill when you want to re-install or confirm it explicitly — say, after deleting the folder or debugging why Claude isn't picking it up.
Then either type /anyscribe to invoke it explicitly, or just ask in plain English — the skill activates on its own when your request is about transcription:
- "Transcribe this YouTube link and drop it in my vault:
https://youtube.com/watch?v=..." - "Transcribe these five URLs with the cheapest provider, and tell me which ones failed."
- "Switch me to Deepgram and re-run the last one with speaker labels."
New to skills? A skill is a folder of instructions Claude reads when your request matches. Nothing runs in the background — the files just sit there until Claude needs them.
MCP (Claude Desktop, Cursor, other hosts)
pip install "anyscribe[mcp]"
The server command, anyscribe-mcp, ships with the base package — it's already on your PATH. What the [mcp] extra adds is the MCP library the server needs in order to run, kept out of the base install so a plain pip install anyscribe stays light. Without the extra, the command exists but stops on a missing import; with it, anyscribe-mcp is the server your agent talks to.
Seeing
No module named 'mcp.server.fastmcp'? You're on an anyscribe release older than 0.16.3 with the newest MCP library. The MCP project shipped a version 2 that renamed the piece anyscribe was importing, and older anyscribe releases didn't ask for a specific version — so a fresh install picked up the new one and the two no longer fit. Upgrading fixes it:
bash pip install -U "anyscribe[mcp]"0.16.3 and later pin the range explicitly, so this can't recur.
Claude Code:
claude mcp add anyscribe -- anyscribe-mcp
Claude Desktop — add this to your claude_desktop_config.json:
{ "mcpServers": { "anyscribe": { "command": "anyscribe-mcp" } } }
Other hosts (Cursor and friends) want the same two facts in whatever shape their config file takes: the command is anyscribe-mcp, and it needs no arguments.
Configure anyscribe first. The MCP server uses the same
~/.anyscribe/config.yamland API keys as the CLI. Runanyscribe onboardonce before wiring up MCP, or the tools will start by telling your agent there's no provider configured.
What your agent can reach (skill vs MCP)
Both paths hit the same engine — the difference is how much of it is exposed.
| Capability | Skill (CLI) | MCP |
|---|---|---|
| Transcribe one URL or file | ✓ | ✓ transcribe |
| Transcribe many at once | ✓ | ✓ batch_transcribe |
| Download media without transcribing | ✓ | ✓ download |
| Browse what's in your vault | ✓ | ✓ list_transcripts |
| Delete a transcript | ✓ | ✓ delete_transcript |
| Read your settings | ✓ | ✓ get_config |
| Change a setting | ✓ | ✓ set_config |
| See available providers | ✓ | ✓ list_providers |
| Check an API key works | ✓ | ✓ test_provider |
| Health check | ✓ | ✓ doctor |
Read run logs (logs) |
✓ | — |
Update anyscribe (update) |
✓ | — |
Run the setup wizard (onboard) |
✓ | — |
Set up offline transcription (local setup) |
✓ | — |
Manage Whisper models (model list / pull / rm) |
✓ | — |
| Menu-bar tray | ✗ | ✗ |
The tray is the one row where the agent runs out of road partway. Claude Code can start it (anyscribe tray) or set it to launch at login (anyscribe install-service), because those are ordinary commands. What no agent can do is use it: the tray's whole point is a menu-bar icon you click, and there's nothing to click on from a transcript. If you want anyscribe always running in the background, see Getting Started → Keep it running.
Why the short list on MCP? Every MCP tool is a permanent promise with a fixed shape. Ten cover the work; the rest are one-off maintenance jobs better done by a human at a terminal — or by Claude Code, which has the whole CLI anyway.
Agent-friendly flags
Anything your agent needs to parse, it should ask for as JSON:
anyscribe "<url>" --json --quiet
Most commands that report results accept --json, and commands that would stop to ask "are you sure?" accept --yes (-y) to answer in advance — the pair that keeps an agent from hanging on a prompt it can't see. Full details, including the exact JSON shape, live in Commands → For scripts and agents.