> ## Documentation Index
> Fetch the complete documentation index at: https://www.voiceflow.com/docs/llms.txt
> Use this file to discover all available pages before exploring further.
# Voiceflow Knowledge MCP
> A public, read-only MCP server that answers questions about how Voiceflow works, grounded in the documentation, with a citation on every answer.
The **Voiceflow Knowledge MCP** is a second, separate server: where the [Voiceflow MCP server](/docs/mcp/overview) works on *your* projects, this one knows how Voiceflow works — the product, the REST API, the `vf` CLI and the MCP server — and answers questions about it from the published documentation, with a citation on every answer. It is public, needs no account and no key, and changes nothing.
Connect it to any MCP client that speaks Streamable HTTP, and your assistant can answer "how do I…", "what is…" and "does X still exist" questions about Voiceflow without leaving the conversation — or use it alongside the Voiceflow MCP server, so the assistant that edits your agent can also look up how a setting works before it changes it.
| | Voiceflow MCP server | Voiceflow Knowledge MCP |
| --- | --- | --- |
| URL | `https://mcp.voiceflow.com/mcp` | `https://www.voiceflow.com/docs/mcp` |
| Knows | your account: projects, agents, transcripts | how Voiceflow works: the docs, the API, the CLI, what changed |
| Sign-in | OAuth, your Voiceflow account | none |
| Can change things | yes (its grant covers reading and changing) | no — every tool is read-only |
## Connect a client
The server's local name below is `voiceflow-knowledge`; pick any name that does not collide with a server you already have.
Run this from your terminal, not inside a Claude Code session:
```bash
claude mcp add --transport http voiceflow-knowledge https://www.voiceflow.com/docs/mcp
```
Add `--scope user` to make it available in every project. There is no sign-in step: `/mcp` shows it connected straight away.
Add the server to `~/.cursor/mcp.json` (every project) or `.cursor/mcp.json` (this project):
```json
{
"mcpServers": {
"voiceflow-knowledge": {
"url": "https://www.voiceflow.com/docs/mcp"
}
}
}
```
`url` is the only key. Nothing prompts you to sign in.
Add the server to `.vscode/mcp.json` (this workspace) or the user configuration (Command Palette, then **MCP: Open User Configuration**):
```json
{
"servers": {
"voiceflow-knowledge": {
"type": "http",
"url": "https://www.voiceflow.com/docs/mcp"
}
}
}
```
The wrapper key is `servers`, not `mcpServers`, and `"type": "http"` is required — VS Code would otherwise treat the entry as a local process to launch.
Add the server to `~/.codex/config.toml`:
```toml
[mcp_servers.voiceflow-knowledge]
url = "https://www.voiceflow.com/docs/mcp"
```
To confirm it works, check the client's own view of the server rather than the answer: in Claude Code, `/mcp` lists `voiceflow-knowledge` as connected with its four tools; Cursor, VS Code and Codex show the server and its tools in their MCP settings. A citation alone proves nothing — an assistant with web access or prior knowledge can cite `voiceflow.com/docs` on its own. Then ask it **how far back a Voiceflow agent remembers earlier turns** and watch the call to `ask_voiceflow` in the transcript.
## What the assistant gets
Four tools, all read-only. The assistant picks them itself; you ask in plain language.
- **`ask_voiceflow`** — one grounded answer: the passage that answers the question, the matching facts on that page (settings, defaults, limits, prices), the concept it names, and a citation. The first call for any "how do I", "what is", "how does X work", API, CLI or error question; most questions need nothing else.
- **`search_voiceflow`** — a ranked list of pages (up to 10), each with its best matching section, a citable URL with `#anchor`, a snippet and why it matched. For browsing candidates, comparing two readings, or when the first answer came back weak.
- **`read_voiceflow_page`** — one page or section as text, by URL, with its heading trail, the headings around it, and the rows of a table rendered from data. For exact wording, a code sample, a table, or a request/response schema.
- **`get_voiceflow_changelog`** — dated changelog entries (added, improved, changed, deprecated, fixed) and tombstones: things that no longer exist, with what replaced them, each linking the page that describes the feature today. For "what's new since", "when did X ship", "is X deprecated", "does X still exist", "what replaced X".
The answers describe Voiceflow **as it works right now**. Questions about history — what changed, when, and what replaced what — belong to `get_voiceflow_changelog`; when the assistant asks another tool about history, the server says so and names the changelog tool.
Four resources ride alongside the tools: `voiceflow://glossary` (every Voiceflow term, its definition, the phrasings people use for it, and where it is documented), `voiceflow://outline` (where every page lives), `voiceflow://skill` (how to build, operate and measure agents over the APIs, the MCP server and the CLI) and `voiceflow://manifest` (what this deployment indexed).
## How it answers well
- **The question, in the user's words.** `question` is the only thing a call needs. The assistant may add `terms` in Voiceflow's own vocabulary (a glossary term, a heading, a setting name) and an `intent` (`how_to`, `what_is`, `api`, `cli`, `error`, `exact`) for a sharper first call — both are optional.
- **Weak terms are dropped, never refused.** A term that is a whole sentence, a generic word, or a word the documentation never uses is dropped with a note saying so; the question is still answered.
- **The answer suggests the next call.** When a call carried no terms, or the answer's confidence is low, it comes with `suggestedTerms` — the nearest terms in Voiceflow's vocabulary — so a second, sharper call needs nothing the first answer did not hand over.
- **Only a malformed call is refused**, and the refusal says exactly how to fix it, with an example.
- **Every answer carries a citation** — a `voiceflow.com/docs` URL, with `#anchor` for a section — and the assistant is told to cite it verbatim.
## Limits
- **Read-only, public, no sign-in.** The server exposes no tool that changes anything, on Voiceflow or anywhere else.
- **Rate-limited per client address**: 20 answers a minute (tool calls and resource reads), and 120 requests of any kind a minute; past either, the server answers `429` with a `retry-after` header. The connection handshake, the tool list and pings are not answers — a client can connect freely; only a loop pays.
- **English.** The documentation is in English and the server searches it in English; an assistant translating a user's question before asking gets the best answers.
The server was renamed on 2026-09-17 — the tools were `answer_voiceflow_docs`, `search_voiceflow_docs`, `read_voiceflow_docs` and `changes_voiceflow_docs`, and the server called itself *Voiceflow Documentation*. The earlier names still answer (with a note naming the new one) until version 2.2.0 of the server; update a client that uses them.