Skip to main content
Glama

of-mcp — Open Finance Brasil docs as an MCP server

A local MCP server that turns the public Confluence space at openfinancebrasil.atlassian.net/wiki/spaces/OF into four tools any MCP-compatible client (Claude Desktop, Claude Code, Cursor, etc.) can call:

Tool

What it returns

search_docs

Top BM25 chunks (compact snippets + page id + URL)

get_page

Full markdown of one page; optional section= slice

list_sections

Page tree (root pages, or children of a page id)

answer_question

Verbose RAG context (full chunks + citations) for the host LLM to answer

Why this design (token economy)

  • No embeddings. SQLite FTS5 (BM25) is local, instantaneous, and costs zero tokens. For keyword-heavy technical docs this is usually as good as semantic search. Optional [semantic] extra is left as a hook if you need re-ranking.

  • Heading-based chunks (~300 tokens). search_docs returns 6 small snippets by default — typically a few hundred tokens total — instead of whole pages.

  • No internal LLM call. answer_question returns the context, the calling assistant does the synthesis in its own context window. Avoids paying twice.

Setup

git clone <this-folder> of-mcp
cd of-mcp
python -m venv .venv
source .venv/bin/activate          # Windows: .venv\Scripts\activate
pip install -e .
cp .env.example .env                # optional — defaults are fine

No Atlassian account or token needed. The OF Confluence space is public, and the crawler hits the REST API anonymously. The ATLASSIAN_EMAIL / ATLASSIAN_API_TOKEN variables in .env.example exist only as a fallback in case Atlassian ever restricts the space — leave them empty.

Build the index

of-mcp-crawl              # full crawl, incremental on subsequent runs
of-mcp-crawl --force      # re-index every page
of-mcp-reindex            # re-chunk from cached pages, no network

The index lives in ./data/of.db (SQLite + FTS5).

Wire it up to Claude Desktop / Claude Code

Add this to your MCP client config (e.g. ~/.config/claude/claude_desktop_config.json):

{
  "mcpServers": {
    "of-mcp": {
      "command": "/absolute/path/to/of-mcp/.venv/bin/of-mcp",
      "env": {
        "OF_MCP_DB_PATH": "/absolute/path/to/of-mcp/data/of.db"
      }
    }
  }
}

For Claude Code:

claude mcp add of-mcp /absolute/path/to/of-mcp/.venv/bin/of-mcp

Restart the client. You should see the four tools available.

Usage examples

In your MCP-enabled chat:

use search_docs with query "iniciação de pagamento erros 422"

open the page returned and read the "Erros" section using get_page

use answer_question for "como funciona consentimento recorrente?"

Project layout

of-mcp/
├── pyproject.toml
├── .env.example
├── README.md
├── data/                    # SQLite index lives here (gitignored)
└── src/of_mcp/
    ├── server.py            # FastMCP server + tool definitions
    ├── crawler.py           # Confluence REST client + crawl orchestrator
    ├── chunker.py           # HTML → Markdown → heading chunks
    ├── search.py            # BM25 wrapper + output formatters
    ├── db.py                # SQLite schema + FTS5 triggers
    ├── config.py            # env loader
    └── scripts/
        ├── crawl.py         # `of-mcp-crawl` entrypoint
        └── reindex.py       # `of-mcp-reindex` entrypoint

Re-indexing on a schedule

The crawler is incremental (skips pages whose Confluence version is unchanged), so a daily cron is cheap:

30 4 * * *  cd /path/to/of-mcp && .venv/bin/of-mcp-crawl >> crawl.log 2>&1

License

MIT — do whatever you want, but the OF documentation itself belongs to the Open Finance Brasil structure governance.

Available Tools

4 tools
answer_questionA

Retrieve the most relevant chunks formatted as RAG context.

The MCP itself does NOT call an LLM — it returns the raw context plus citations so the calling assistant can answer using its own context window (saving tokens vs. running a second model here).

Inputs: question (required), max_chunks. Returns: result

get_pageA

Return the full markdown of a page.

Args: page_ref: Confluence page id, exact title, or substring of the title. section: Optional heading text. If provided, only the matching section (and its sub-sections) is returned.

Inputs: page_ref (required), section. Returns: result

list_sectionsA

List pages in the documentation tree.

Args: parent_id: If omitted, returns root pages of the OF space. Otherwise returns immediate children of the given page id.

Inputs: parent_id. Returns: result

search_docsA

Search the Open Finance Brasil docs (BM25). Returns compact snippets.

Args: query: Natural-language query in Portuguese or English. limit: Max number of hits (default 6, hard-capped at 20).

Inputs: query (required), limit. Returns: result

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A4.4/5.0

    Scored across 4 tools

    Disambiguation5/5

    Each tool serves a distinct purpose: answer_question returns RAG context, get_page retrieves full page markdown, list_sections navigates the doc tree, and search_docs performs keyword search. No overlap in functionality.

    Naming Consistency5/5

    All tools follow a consistent verb_noun snake_case pattern (answer_question, get_page, list_sections, search_docs), making it easy to infer their actions.

    Tool Count5/5

    With 4 tools, the server is well-scoped for documentation retrieval and search. Each tool earns its place, covering essential operations without unnecessary bloat.

    Completeness5/5

    The tool set covers the full lifecycle of documentation access: search (search_docs), context retrieval (answer_question), page content (get_page), and navigation (list_sections). No obvious gaps.

    Related MCP Connectors