ByteAsk Embedded MCP

Page-cited answers from embedded & firmware reference docs — for coding agents that can't afford to guess a register value.

Official MCP Registry Namespace: ai.byteask/embedded-docs · Remote MCP Endpoint: https://mcp.byteask.ai/mcp

Quickstart · Tools · Connect a client · Configuration · Hosted server · Contributing

---

ByteAsk Embedded MCP is the open-source server behind ByteAsk Embedded Docs: a source-grounded, page-cited evidence-retrieval MCP server for coding agents (Claude Code, Codex, Cursor) that write firmware / driver / protocol code and need exact facts — SunSpec points, register offsets, Modbus function codes, trip thresholds, SCPI commands, API symbols.

It returns verbatim snippets with page citations — never an authored answer — and when nothing is relevant enough it says no match rather than fabricate. Every document is treated equally: no authority layer, no filters.

Note

What's in this repo: the MCP server — tools, transports (stdio + Streamable HTTP), bearer auth, DNS-rebinding protection, result rendering — plus a small, pluggable retrieval interface.

What's not in this repo: the retrieval engine and the document corpus. How documents are parsed, chunked, embedded, and ranked, and the licensed source material itself, sit behind the SearchBackend seam and power the hosted endpoint at https://mcp.byteask.ai/mcp. This repo ships an in-memory SampleBackend (a few illustrative, public-knowledge records) so the server runs out of the box.

Why

• Cited, or nothing. Every hit is verbatim source text with a section + page citation. On a miss it returns an honest "no confident match" — it never invents a register value.
• Built for coding agents. The tool descriptions and triggers are tuned so agents call search_docs reflexively the moment they see a hex literal, a Modbus code, an IEEE clause, a SCPI verb, or an MCU part number — before answering from memory.
• Two transports, one server. stdio for local agents, Streamable HTTP for hosted.
• Bring your own retrieval. The search engine is a two-method interface — swap in anything behind BYTEASK_BACKEND without touching the server.
• Zero-setup demo. The bundled SampleBackend runs immediately. No API keys.

Quickstart

Requires Python ≥ 3.10 and uv.

uv sync
uv run byteask-embedded-mcp        # run as an MCP server (stdio)

That's it — the bundled SampleBackend serves a couple of illustrative records, so search_docs works immediately. Run the offline tests with uv run pytest.

Connect a client

Hosted (no install)

The hosted server speaks Streamable HTTP at https://mcp.byteask.ai/mcp and is backed by the full licensed corpus.

Claude Code:

claude mcp add --transport http byteask-embedded-docs https://mcp.byteask.ai/mcp

Codex, Cursor, and other clients (mcp-remote)

{
  "mcpServers": {
    "byteask-embedded-docs": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://mcp.byteask.ai/mcp"]
    }
  }
}

Local (this repo)

The project-scoped .mcp.json registers the stdio server for clients that read it. Manually, for Claude Code:

claude mcp add byteask-embedded-docs -- uv run byteask-embedded-mcp

Tools

Input is natural language (or an exact identifier). Output is compact markdown.

Tool	What it does
search_docs(query, limit=8)	Search the corpus; return ranked, page-cited evidence. Each hit has a document title, a section + page citation, the verbatim snippet, and a result_id. A "no confident match" response means not found — do not fabricate.
get_context(result_id)	Expand a hit to its full source section.
request_document(request)	Ask for a missing document to be added (logged server-side).

Example output:

## Results for "what Modbus function code writes multiple registers"

### Sample — Modbus Application Protocol (illustrative) — §6.12, p.30
> Function code 16 (0x10), Write Multiple Registers, writes a block of contiguous
> holding registers (1 to 123 registers) in a remote device. ...
_ref: sample:modbus-fc16_

Plug in your own retrieval

The server depends only on a two-method interface (backend.py):

class SearchBackend(Protocol):
    def search(self, query, limit=8, effort=None) -> dict: ...
    def get_context(self, result_id, effort=None) -> dict: ...

Implement it, expose a factory make_backend(config) -> SearchBackend, and point the server at it:

BYTEASK_BACKEND="my_pkg.my_module:make_backend"

The exact return-value contracts are documented at the top of backend.py.

Configuration

All settings are environment variables (loaded from .env; see .env.example).

Variable	Default	Notes
BYTEASK_BACKEND	—	module:callable returning a SearchBackend; empty → SampleBackend
BYTEASK_LOGS	logs	where query / request JSONL logs are written
MCP_TRANSPORT	stdio	stdio (local agents) or http
MCP_HTTP_HOST / MCP_HTTP_PORT	127.0.0.1 / 8000	HTTP bind address
MCP_HTTP_AUTH_TOKEN	—	bearer token for HTTP (empty = unauthenticated, dev only)
MCP_ALLOWED_HOSTS	—	comma-separated hosts allowed in the Host header (* disables)
LOG_LEVEL	INFO	stderr log verbosity

Running over HTTP

MCP_TRANSPORT=http MCP_HTTP_AUTH_TOKEN=$(openssl rand -hex 32) \
  uv run byteask-embedded-mcp --host 0.0.0.0 --port 8000

Clients then send Authorization: Bearer <token>. The bundled bearer check is a shared-secret stub — replace it with real auth (OAuth 2.1 resource server, mTLS, or a trusted reverse proxy) before exposing publicly. DNS-rebinding protection stays on independently via MCP_ALLOWED_HOSTS.

Hosted server

You don't need to run anything to use ByteAsk Embedded Docs. The hosted server gives Claude Code, Codex, Cursor, and any MCP client exact, page-cited facts from embedded and firmware reference docs — register maps, protocol function codes, SCPI commands, standard thresholds, datasheet specs. The guarantee: verbatim source, or "no match" — never an invented value.

Name	byteask-embedded-docs
Endpoint	https://mcp.byteask.ai/mcp (Streamable HTTP)
Docs & per-client setup	https://docs.byteask.ai/embedded

This repository is the open-source server that powers that endpoint.

Project layout

src/byteask_embedded_mcp/
  server.py     # FastMCP app + 3 tools (search_docs, get_context, request_document)
  backend.py    # SearchBackend protocol + in-memory SampleBackend (swap for real retrieval)
  render.py     # structured result -> compact markdown
  http_auth.py  # Streamable HTTP entrypoint + stub bearer-token guard
  config.py     # server config (transport, logging, backend selection)
  schemas.py    # Hit / Section result types
  obs.py        # per-call JSONL logging
tests/          # offline unit tests (renderer, backend, server tools)
assets/         # README demo GIF + its deterministic generator

Security

• stdout stays clean in stdio mode (it is the JSON-RPC channel); all logs go to stderr / logs/*.jsonl.
• The HTTP bearer check is a stub — unauthenticated if no token is set, a shared secret at best. Harden it before exposing widely.
• DNS-rebinding protection is on by default for the HTTP transport.

Contributing

PRs and issues are welcome.

uv sync            # install (incl. dev tools)
uv run pytest      # run the offline test suite

A few conventions to keep the server clean:

• The backend seam is the extension point. Retrieval internals (parsing, chunking, embeddings, ranking) are intentionally out of scope here — build them behind SearchBackend in your own package, not in this repo.
• Keep the dependency surface small and the stdio path free of the HTTP stack.
• Add a test for new behavior; the suite is fully offline (no network, no keys).

License

MIT © ByteAsk