Docs MCP

Your docs, agent-ready

Index Markdown docs into a production-ready MCP.
Agents search, retrieve, and correctly cite your docs.Offline-firstPerformant
Markdown-based
Get accessView Github
~/acme-project

Trusted by

Solution

Three tools, zero friction

A compact, read-only MCP toolset optimized for agent workflows. Search, retrieve, and improve — all over MCP.

search_docs

Semantic, keyword, or path-prefix search across all indexed docs. Compact hits with scores, snippets, and anchors — optimized for token efficiency.

get_doc

Fetch a full document or a specific section by anchor. Bounded payloads with token estimates prevent context window overflow.

guides/retries
742 tokens
#backoff-strategy
## Backoff strategy
The SDK uses exponential backoff with
jitter to retry failed requests.
Default configuration:
- Max retries: 3
- Initial delay: 500ms
- Max delay: 60s
- Backoff multiplier: 1.5
truncated: false

give_feedback

Agents and users report missing or outdated docs inline. Feedback is logged as structured JSON for docs improvement loops — no separate database required.

give_feedback
medium
fromagent:cursor
doc_idguides/retries
categorymissing_content

"Missing details for per-request override in Python SDK"

logged
stdout

How it works

From Markdown to
agent-ready in minutes

1

1. Configure your repo

Point Speakeasy at your Markdown docs and SDK repos with a simple gen.yaml. Define which folders to index and which SDK references to include.

acme/docs
docs
getting-started.md
authentication.md
webhooks.md
guides
pagination.md
error-handling.md
api-reference
users.md
orders.md
payments.md
sdk-repos
acme-ts-sdk
README.md
reference

2

2. Index your content

Speakeasy scans your docs, builds a full-text search index, and generates vector embeddings — optimized for fast, token-efficient agent queries.

Terminal
$ speakeasy run
 
→ Scanning ./docs — 47 Markdown files
→ Scanning acme-ts-sdk — 12 reference docs
→ Building full-text search index
→ Generating vector embeddings
 
✓ MCP server compiled → @acme/docs-mcp

3

3. Use your favorite client

Add the compiled MCP server to Cursor, VS Code, or any MCP-compatible client. Agents can immediately search, retrieve, and cite your docs.

~/acme-project
Performance

Built for speed, optimized for agents

Hybrid search combining full-text, phrase proximity, and vector similarity. Fast enough for multi-turn agent loops.

$1

To embed a 28.8 MB corpus

5ms

Median search latency (FTS)

$0.001

Per 1,000 queries with embeddings

"Speakeasy was critical in launching our MCP server, and they’ve continued to be a great partner in iterating on the server since then"

Benjamin Woskow

Sr. Director of Eng @ LaunchDarkly

"The MCP we built using Speakeasy just works. It was honestly much simpler than we expected"

Constantine Nathanson

Staff Full Stack Engineer @ Cloudinary

"It took me 30 minutes to set up the first toolset which included testing it out locally to see how it worked. I was extremely impressed with the experience"

James Perkins

Co-founder & CEO @ Unkey

"Speakeasy was critical in launching our MCP server, and they’ve continued to be a great partner in iterating on the server since then"

Benjamin Woskow

Sr. Director of Eng @ LaunchDarkly

"The MCP we built using Speakeasy just works. It was honestly much simpler than we expected"

Constantine Nathanson

Staff Full Stack Engineer @ Cloudinary

"It took me 30 minutes to set up the first toolset which included testing it out locally to see how it worked. I was extremely impressed with the experience"

James Perkins

Co-founder & CEO @ Unkey

Frequently askedquestions

What is Docs MCP?
Docs MCP is an MCP server that turns your Markdown documentation into a high-signal, agent-ready source. Agents like Claude, Cursor, and ChatGPT can search, retrieve, and cite first-party docs through standard MCP tools instead of guessing from training data or scraping HTML.
How is Docs MCP different from a generic search engine or RAG store?
Generic vector stores treat docs as opaque chunks. Docs MCP preserves document structure, returns sections that fit an agent's context window, and exposes tools that match how agents actually work: search, get a specific doc, and submit feedback. The result is fewer hallucinations and lower token cost per question.
Which clients work with Docs MCP?
Any MCP-compatible client. That includes Claude Desktop, Claude Code, Cursor, VS Code with MCP support, ChatGPT with MCP, and your own internal agents. The same Docs MCP server is reachable from every one of them.
What kinds of documentation can be indexed?
Any Markdown content: developer docs, API references, runbooks, internal wikis exported to Markdown, and SDK references. Docs MCP can also index code samples and SDK source so that agents have both the prose and the executable references when they answer.
How is the index kept fresh?
Speakeasy reindexes on every commit through your CI pipeline. Updates land in the MCP server within minutes of merging, so agents are answering from the current docs instead of a snapshot from last week.
Is Docs MCP token efficient?
Yes. Search results return concise, structured snippets with citation links rather than full pages, and the get-doc tool returns only the requested section. Many users see substantially lower token cost compared to feeding entire docs into context.
Where is the Docs MCP server hosted?
Speakeasy hosts Docs MCP as a managed service. Self-hosted deployments are available for organizations that need to keep the index inside their own infrastructure for compliance or data-residency reasons.
Can users send feedback on bad answers?
Yes. Docs MCP includes a give-feedback tool that lets agents (and the humans behind them) flag answers that were wrong, missing, or unclear. The feedback is routed back to the docs team so the source improves over time.

Make your docs work for every AI agent

Talk to usRead the docs