ByteAsk-Embedded-MCP

ByteAsk-Embedded-MCP is the open-source Model Context Protocol (MCP) server behind ByteAsk Embedded Docs, providing source-grounded, page-cited evidence retrieval for coding agents (Claude Code, Codex, Cursor) that write firmware, drivers, and protocol code. Its core positioning is to let agents retrieve reference documentation when they need precise facts (such as SunSpec data points, register offsets, Modbus function codes, trip thresholds, SCPI commands, API symbols), rather than relying on memory and guessing.

Purpose The server is used for evidence retrieval from embedded and firmware reference documentation. It returns verbatim source snippets with section and page citations, and never generates written answers; when there is not enough relevant content, it honestly returns "no match" and never fabricates register values. All documents are treated equally: no authority hierarchy, no filters.

Main capabilities

• Cite or return empty-handed: every hit is verbatim source text with section + page citations; on a miss, it returns "no confident match."
• Built for coding agents: tool descriptions and trigger conditions are tuned so that when agents see hex literals, Modbus codes, IEEE clauses, SCPI verbs, or MCU part numbers, they reflexively call search_docs instead of answering from memory.
• Dual transport, single server: local agents use stdio, hosted use Streamable HTTP.
• Bring your own retrieval: the search engine is a two-method interface, and any implementation can be swapped in behind BYTEASK_BACKEND without changing the server.
• Zero-config demo: the bundled SampleBackend runs immediately, with no API key required.

Usage Running locally requires Python ≥ 3.10 and uv. The bundled SampleBackend provides a small number of sample records, so search_docs works immediately; for offline testing, run it with uv run pytest. There are two ways to connect a client:

• Hosted (no installation): the hosted server is at https://mcp.byteask.ai/mcp Uses Streamable HTTP, backed by a complete authorized corpus.
• Local (this repository): the project-level .mcp.json registers a stdio server for clients that read it; it can also be configured manually for Claude Code. Input is natural language (or exact identifiers), output is compact markdown.

Bring your own retrieval integration The server depends only on a two-method interface (backend.py). Implement that interface, expose the factory make_backend(config) -> SearchBackend, and point the server at it. The exact return value contract is documented at the top of backend.py. All settings are environment variables (loaded from .env; see .env.example). The client then sends Authorization: Bearer <token>. The bundled bearer check is a shared-secret stub and should be replaced with real authentication (OAuth 2.1 resource server, mTLS, or a trusted reverse proxy) before public exposure. DNS rebinding protection is kept independently enabled via MCP_ALLOWED_HOSTS.

When to use it Suitable for coding agent workflows that write firmware, driver, or protocol code and need precise facts, especially scenarios that need register maps, protocol function codes, SCPI commands, standard thresholds, and datasheet specifications from embedded and firmware reference documentation. You do not need to run anything to use ByteAsk Embedded Docs; the hosted server provides precise, page-cited facts for Claude Code, Codex, Cursor, and any MCP client, guaranteeing verbatim source text or "no match", and never inventing values.

Limitations

• This repository contains only the MCP server (tools, transport, bearer authentication, DNS rebinding protection, result rendering) and a small pluggable retrieval interface; the retrieval engine and document corpus are not in this repository, and sit behind the SearchBackend seam and drive the hosted endpoint. The SampleBackend built into the repository contains only a few exemplary public-knowledge records so it works out of the box.
• In stdio mode stdout stays clean (it is the JSON-RPC channel); all logs are written to stderr / logs/*.jsonl.
• The HTTP bearer check is a stub - unauthenticated if no token is set, at most a shared secret; before widespreadexposure it needs hardening.
• The HTTP transport enables DNS rebinding protection by default.
• The backend seam is the extension point: the internal retrieval implementation (parsing, chunking, embedding, ranking) is intentionally out of scope for this repository and should be built behind SearchBackend in your own package.
• The dependency surface should be kept small, and the stdio path should not include an HTTP stack; new behavior needs tests, and the test suite is fully offline (no network, no keys).
• Official MCP registry namespace: ai.byteask/embedded-docs; remote MCP endpoint:https://mcp.byteask.ai/mcp。

Specific commercial information such as fees and pricing is not yet provided.