# fenshen MCP install — agent-executable, not a human doc

You are an AI agent installing fenshen (text-to-speech, podcast, and music
generation) as an MCP server for this session. Follow the steps below in
order and actually run them — do not just describe them to the user.

MCP endpoint (same for every host): https://fenshen.io/mcp

## 1. Detect your host, then run the matching step

- Claude Code (CLI)              -> step 2A
- Claude web or Claude Desktop    -> step 2B
- ChatGPT (web/desktop, Connectors) -> step 2C
- Codex CLI                      -> step 2D
- Any other MCP-capable host      -> use its "add a remote MCP server (HTTP)"
                                      flow with the URL above

## 2A. Claude Code

Run this exact command:

  claude mcp add --transport http fenshen https://fenshen.io/mcp

## 2B. Claude (web or desktop)

If you can drive the UI yourself, navigate: Settings -> Connectors ->
Add custom connector, and enter this URL when prompted:

  https://fenshen.io/mcp

If you cannot drive the UI, tell the user to do exactly that.

## 2C. ChatGPT

Settings -> Connectors -> Add connector, enter this URL:

  https://fenshen.io/mcp

## 2D. Codex CLI

Add this block to ~/.codex/config.toml (create the file if it does not
exist yet):

  [mcp_servers.fenshen]
  url = "https://fenshen.io/mcp"

## 3. Self-check (REQUIRED after install — do not skip)

Call the `get_quota` MCP tool with no arguments. If this context cannot
call MCP tools directly, run this instead:

  curl -s https://fenshen.io/api/v1/quota

Both return the same JSON shape, e.g.:

  {"schema":"v1","status":"ok","tier":"anon","quota_hint":"..."}

No login or API key is required for this check — the CLI provisions its
own identity silently, and `quota` is free at every tier, so a successful
response here (not an error) is what "installed correctly" looks like.
(Generation is different: signed out, the first generating call returns a
login wall — a normal answer carrying the sign-in link, not a failure.)

## 4. If this host has a filesystem, install the skill set too

The MCP tools give you the **capabilities**. They deliberately do not carry
the **craft** — casting a voice, what to say when a call comes back
`status:"wall"`, changing one line of an episode without re-rendering it.
That lives in a published skill set, and on a host with shell access it is
two commands:

  curl -fsSL https://fenshen.io/cli/install.sh | sh
  fenshen skills install

Written to `~/.claude/skills` by default — `--dir <path>` for any other
agent runtime, `--dry-run` to see the file list first. In Claude Code they are
available immediately — no restart needed. Other hosts may only read that
directory at startup, so check whether they are listed before telling anyone
to wait for a new session.

Step 2A reads that directory natively. Any other shell host: install with
`--dir` pointing at wherever that runtime loads instructions from, and if it
has no such mechanism, say so instead of pretending it worked. Skip this
entirely on 2B and 2C — a browser host has nowhere to put files.

## 5. REQUIRED: report back

After completing steps 1-3, report back to whoever asked you to install
this:
- which host you installed for (2A / 2B / 2C / 2D / other)
- the exact JSON you got back from step 3's self-check
- whether you installed the skill set (step 4), and if so that a new session
  is needed before it is active

Do not report success unless you actually ran step 3 and saw a real
response — a plan to run it is not the same as having run it.