blograg 0.0.1

A local MCP server for retrieving paragraphs from one Jekyll-style blog.

blograg

blograg is a local MCP-oriented retrieval tool for one Jekyll-style blog. It uses labelrag as the retrieval core and treats heading-delimited markdown sections as the paragraph unit.

It is designed for local, single-blog usage:

• build a paragraph index from one Jekyll-style repository
• serve that index over MCP Streamable HTTP
• inspect service state from the CLI and a lightweight browser page
• register the HTTP endpoint with local MCP clients such as Codex or OpenClaw

Detailed command reference lives in docs/commands.md.

Scope

Current 0.0.0 scope:

• one local blog directory
• Jekyll-style front matter parsing
• heading-delimited paragraph segmentation
• full rebuild only
• one MCP tool: retrieve_paragraphs

Out of scope:

• incremental indexing
• multiple blog roots
• runtime rebuilds from the MCP server
• alternate storage backends
• broad MCP tool surfaces beyond paragraph retrieval

Installation

python3.11 -m venv .venv
. .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -e '.[dev]'

Quick Start

Initialize local defaults and optional provider secrets:

blograg config wizard

Build an index:

blograg build --blog-dir /path/to/blog --index-dir /path/to/index

Start the managed HTTP service:

blograg start --index-dir /path/to/index

Inspect service state:

blograg status
blograg logs --follow
blograg doctor

Open the browser status page:

http://127.0.0.1:8765/

Register the MCP endpoint with a client:

blograg register --client codex
blograg register --show

Core Commands

Most day-to-day usage is centered on:

• blograg config wizard
• blograg build
• blograg serve
• blograg start
• blograg status
• blograg logs
• blograg doctor
• blograg register

For command-by-command examples and option summaries, see docs/commands.md.

Persistent Config

blograg stores user-level config and secrets in:

• config.toml
• secrets.toml

Default locations:

• macOS/Linux: ~/.config/blograg/
• Windows: %AppData%/blograg/

Useful commands:

blograg config path
blograg config show
blograg config show --all
blograg config set default_index_dir /path/to/index
blograg config set retrieval.retrieval_strategy label_gate_semantic_rank
blograg config set-secret mistral --api-key your-key-here

config show masks secret values and only reports whether each provider key is configured.

MCP Service Model

blograg serve loads an existing index and starts the MCP server. It does not rebuild automatically. If the index is missing or incomplete, run build first.

The default transport is Streamable HTTP. The default HTTP binding is:

• host: 127.0.0.1
• port: 8765

If you need LAN access, bind explicitly:

blograg serve --host 0.0.0.0 --port 8765

Current HTTP endpoints:

• /mcp
• /
• /healthz

The browser page at / is a lightweight status page, not a separate web app.

MCP Client Registration

Register the local endpoint with one client at a time:

blograg register --client codex
blograg register --client openclaw

Inspect current registration state:

blograg register --show
blograg register --show --server-name blograg-local

You can also register an explicit URL:

blograg register \
  --client codex \
  --server-name blograg-local \
  --url http://127.0.0.1:8765/mcp

LLM Usage

blograg build supports the upstream extraction modes:

• heuristic
• spacy
• llm

Example LLM build:

MISTRAL_API_KEY=your-key-here \
blograg build \
  --blog-dir /path/to/blog \
  --index-dir /path/to/index \
  --concept-extractor llm \
  --llm-provider mistral \
  --llm-model mistral-small

If an index was built with --concept-extractor llm, query analysis at serve time still needs access to the corresponding provider API key. You can provide it through:

• blograg config set-secret ...
• environment variables such as MISTRAL_API_KEY

Retrieval Output

The server currently exposes one tool:

retrieve_paragraphs(query: str, top_k: int = 5)

Each result includes:

• paragraph_id
• text
• post_title
• slug
• section_heading
• trace.retrieval_strategy
• trace.score
• trace.score_kind

Index Layout

blograg build writes an outer blograg directory inside the chosen index root:

/path/to/index/
  blograg/
    manifest.json
    paragraphs.json
    labelrag/
      ...

The outer layer stores blograg-specific metadata and paragraph source metadata. The inner labelrag directory is a normal persisted upstream snapshot.

Runtime Notes

• The default build mode is heuristic, so the default path does not require a spaCy model download.
• The default embedding provider still comes from upstream labelrag, so the first real build or query may download the configured embedding model.
• Advanced retrieval runtime settings live under persisted retrieval.* config keys and can also be overridden through serve and start.

Development Checks

pytest
ruff check .
ruff format --check .
pyright
python -m build
twine check dist/*