rbi-source-mcp 0.8.4

An MCP server for RBI Master Directions: version-aware, citation-first, knows what's withdrawn.

RBI Source MCP

The entire RBI corpus, version-aware, citation-first, in your AI workflow.

An MCP (Model Context Protocol) server that brings authoritative Reserve Bank of India regulatory information into any MCP-capable client (Claude Code, Claude.ai, Claude Desktop, ChatGPT, Cursor, Claude Cowork, Cline, Continue, Goose, Zed, ...) with hybrid retrieval over Master Directions, Standalone Circulars, Master Circulars, Press Releases, and FAQs, citation-first responses, and explicit withdrawal/supersession detection.

Two ways to use it:

• Try the hosted demo: https://rbi-source.harshil.ai/mcp/ (free, no auth, no install) · client setup ↓
• Self-host in 3 lines: pip install rbi-source-mcp + rbi-source-fetch-corpus + Claude Desktop config · details ↓

⚠️ Unofficial, community-maintained open-source tool. RBI Source MCP is not affiliated with, endorsed by, or sponsored by the Reserve Bank of India. RBI® is a trademark of the Reserve Bank of India. For takedown or legal inquiries, open an issue.

Why

Fintech builders writing a clause, PRD section, or partner agreement currently ping the compliance team in Slack and wait hours-to-days to know if their language conflicts with a current RBI rule. Or they Google rbi.org.in, land on an ASP.NET page that may or may not be a withdrawn circular, and ship against rules superseded three years ago.

RBI Source MCP collapses that loop. Paste a clause into Claude/Cursor with this MCP connected; the LLM gets back ranked relevant RBI provisions with paragraph anchors, official URLs, RBI references, and current/withdrawn status. Retrieval-only by design — the MCP doesn't issue compliance verdicts; the LLM synthesizes from the cited material; the user decides; a qualified human compliance reviewer signs off before anyone acts.

Coverage

Content family	Docs	Chunks	Source page
Master Directions	~340	~50k	BS_ViewMasDirections.aspx
Standalone Circulars	~290	~1.5k	BS_ViewListofstandalonecirculars.aspx
FAQs	~100	~1.8k	FAQView.aspx
Press Releases	~55	~360	BS_PressReleaseDisplay.aspx
Master Circulars	~20	~2.8k	BS_ViewMasterCirculardetails.aspx
TOTAL	~810 docs	~57k chunks

Plus ~10k withdrawn-circular records indexed for rbi_check_current lookup (metadata only, no full-text indexing). Live counts refresh daily — see GET /health on the running server for the current numbers.

Hybrid retrieval = FTS5 (BM25 sparse) + sqlite-vec (bge-base-en-v1.5 768-dim dense, embedded via Cloudflare Workers AI on the hosted endpoint) fused via Reciprocal Rank Fusion (k=60). Every response carries a mandatory legal disclaimer. Eval gate: ~25 hand-labeled compliance cases, must pass at ≥80%; currently 100%.

Quick start

Two paths. Pick one.

A. Use the hosted demo (zero install)

The maintainer runs a free, no-auth instance:

https://rbi-source.harshil.ai/mcp/

Public, unauthenticated, retrieval-only. Hosted in Mumbai. Every response carries the mandatory legal disclaimer. Good for evaluation; rate-limited to 60 req/min/IP.

curl -sS https://rbi-source.harshil.ai/health
# → {"version":"0.8.x","status":"ok","documents":803}

B. Self-host (3 lines, runs locally)

pip install rbi-source-mcp                       # or: uv tool install / pipx install
rbi-source-fetch-corpus                          # downloads ~80MB sigstore-signed corpus
# Add to claude_desktop_config.json: {"mcpServers": {"rbi-source": {"command": "rbi-source-mcp"}}}

Defaults to the local bge-base-en-v1.5 embedder (no API keys). Full self-host guide ↓

Connect from your client

Client	How
Claude Code (CLI)	claude mcp add rbi-source --transport http https://rbi-source.harshil.ai/mcp/
Claude.ai / Claude Desktop	Settings → Connectors → Add custom connector → paste the URL. Requires a paid plan (Pro / Team / Enterprise).
ChatGPT	Settings → Connectors → Create custom connector → paste the URL. Requires a paid plan (Plus / Pro / Team / Enterprise).
Cursor	Settings → MCP → Add new MCP server → type streamable-http, URL: https://rbi-source.harshil.ai/mcp/
Claude Cowork	Settings → Connectors → Add custom connector → paste the URL.
Cline (VSCode)	~/.cline/mcp_settings.json, see snippet
Continue.dev	~/.continue/config.json, see snippet
Goose (Block)	~/.config/goose/config.yaml, see snippet
Zed	~/.config/zed/settings.json, see snippet
Anything else that speaks MCP streamable-HTTP	Point it at https://rbi-source.harshil.ai/mcp/

After connecting, four tools (rbi_check_compliance, rbi_search, rbi_get_document, rbi_check_current) become available. Try:

"Use the RBI Source MCP. What are the net-worth requirements for a payment aggregator?"

Config snippets for Cline / Continue / Goose / Zed

Cline (VSCode) — in MCP Settings:

{ "mcpServers": { "rbi-source": { "type": "streamableHttp", "url": "https://rbi-source.harshil.ai/mcp/" } } }

Continue.dev — in ~/.continue/config.json:

{ "experimental": { "modelContextProtocolServers": [ { "transport": { "type": "streamable-http", "url": "https://rbi-source.harshil.ai/mcp/" } } ] } }

Goose — in ~/.config/goose/config.yaml:

extensions:
  rbi-source:
    type: streamable_http
    uri: https://rbi-source.harshil.ai/mcp/
    enabled: true

Zed — in ~/.config/zed/settings.json:

{ "context_servers": { "rbi-source": { "command": null, "settings": { "url": "https://rbi-source.harshil.ai/mcp/", "type": "http" } } } }

Verify the connection

curl -sS -X POST https://rbi-source.harshil.ai/mcp/ \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list","params":{}}' \
  | head -c 500

You should see an SSE event listing the four tools.

Endpoints

• GET / — homepage (HTML for browsers, JSON corpus banner for curl/MCP clients via Accept-header detection)
• GET /health — liveness probe; 200 with documents count when healthy, 503 if corpus is empty/missing
• POST /mcp/ — MCP streamable-HTTP transport (note the trailing slash; clients that don't follow 307s should hit /mcp/ directly)

Verified end-to-end on Claude Code (CLI) and Claude.ai (Settings → Connectors). The other client snippets follow each tool's documented MCP-config format but haven't been individually smoke-tested — if you hit issues with a specific client, please open an issue with the error output.

Tools

rbi_check_compliance(text, topic_hint?, limit?) — headline

Paste free text (a clause, PRD section, draft policy paragraph, code comment); get back ranked relevant provisions with citations: paragraph anchor, official URL, RBI reference, last-updated date, current/withdrawn status.

Retrieval-only by design. The MCP returns the cited material; the LLM consuming it produces any summary or conclusion; the user verifies with a qualified human compliance reviewer. The tool does not issue compliance verdicts.

topic_hint (optional) biases retrieval to one MD: pa / kyc / ppi / cards / e_mandate / afa (full alias list in the tool's input schema). Unknown values are ignored — search spans the full corpus. Out-of-scope inputs (recipes, sports articles, anything unrelated) return low_confidence: true so consuming LLMs can decline to synthesize.

rbi_search(query, filters?, limit?) — direct retrieval

Same hybrid engine as rbi_check_compliance, exposed for cleaner keyword queries: "what are the net-worth requirements for PAs". Returns ranked chunks with citations across the whole corpus or filtered by topic.

rbi_get_document(document_id, include_text?, as_of?) — fetch a document

Returns metadata + table of contents (chunk anchors, sections, page numbers) for any document. Pass include_text=true to get the full assembled body. as_of is reserved for v1.1 (when document version history lands).

rbi_check_current(url_or_ref) — safety / utility tool

Paste an RBI URL; learn whether it's current, withdrawn, or out-of-corpus. Useful for verifying a citation. Three-step lookup:

1. Withdrawn-circulars list → returns withdrawn with replacement reference
2. Active corpus (any indexed family) → returns current with full citation
3. Neither → returns not_withdrawn with honest caveat

Supported URL patterns:

• BS_ViewMasDirections.aspx?id=<MD_ID>
• NotificationUser.aspx?Id=<NOTIF_ID> (checked against withdrawn list + active corpus)

Other inputs return a structured unsupported_at_v0.1 response. Never silently fails.

Mandatory disclaimer

Every tool response includes a _disclaimer field at the top of the JSON object plus an _llm_instruction telling the consuming LLM to surface the disclaimer. Five required points: (1) not legal advice, (2) retrieval-only — no compliance verdicts, (3) provisions may have been amended/withdrawn since the last corpus refresh, (4) verify with a qualified human compliance reviewer before acting, (5) unofficial — not affiliated with the Reserve Bank of India.

The legal posture is preserved on error responses too: if a tool dispatch raises (DB unavailable, embedder OOM, etc.), the wrapped error envelope still carries _disclaimer + _llm_instruction at the top. Guarded by tests/test_error_envelope.py.

Architecture

   Runtime (single-process Python, any Linux/macOS host)
   ──────────────────────────────────────────────────────
   ./data/db.sqlite      FTS5 + sqlite-vec, atomic-swap on refresh
   ./data/db-prev.sqlite one-command rollback target

   Default embedding path: bge-base-en-v1.5 in-process via
   sentence-transformers (~440 MB model on disk, ~3-8s cold load,
   ~80-120ms per query on CPU). With a 10K-entry process-local LRU
   cache, repeat queries skip the cost entirely.

   Optional Cloudflare Workers AI path: same model, ~250ms/call over
   HTTPS, no torch on the box. Picked via
   RBI_EMBEDDING_PROVIDER=cloudflare.

   Both paths emit 768-dim vectors → ONE corpus, ONE schema, no
   client-side dim juggling.

   HTTP-mode public surface (rbi-source-mcp-http)
   ──────────────────────────────────────────────
   GET  /             homepage HTML (browsers) / corpus banner JSON (clients)
   GET  /health       deep liveness probe (counts docs, re-checks
                      sqlite-vec, surfaces query_cache hit/miss stats)
   POST /mcp/         MCP streamable-HTTP transport
   *    OAuth 2.1     RFC 8414, 9728, 7591 ceremonial endpoints

   Corpus build (GitHub Actions)
   ─────────────────────────────
   Daily at 02:00 UTC (07:30 IST):    diff vs latest-corpus  (~5 min, ~50 CF Neurons)
   Monthly on the 1st at 03:00 UTC:   full rebuild safety net (~50 min, ~3000 Neurons)

   crawl 5 families → diff via content-hash → re-extract changed PDFs
   → embed (CF Workers AI in CI; bge-base @ 768-dim)
   → build new SQLite (FTS5 + chunks_vec virtual table)
   → smoke gate (rbi-source-eval, ≥80% absolute AND <5pp regression)
   → sigstore-sign + publish corpus.sqlite.xz as a GitHub Release
   → GC: keep last 14 daily + last 12 monthly timestamped tags + the
     `latest-corpus` moving alias; delete the rest

Self-host

Three lines, no Docker, no Cloudflare account, no crawling.

# 1. Install (persistent — stays available across Claude Desktop restarts)
uv tool install rbi-source-mcp                  # or: pipx install rbi-source-mcp

# 2. Get the prebuilt corpus (~80 MB compressed, refreshed daily via GitHub Actions)
rbi-source-fetch-corpus                         # → ~/.local/share/rbi-source-mcp/db.sqlite

# 3. Wire to Claude Desktop (claude_desktop_config.json) or Claude Code:
#    {
#      "mcpServers": {
#        "rbi-source": { "command": "rbi-source-mcp" }
#      }
#    }

That's it. The default embedding path runs bge-base-en-v1.5 in-process via sentence-transformers (~440 MB model auto-downloaded on first query, cached at ~/.cache/huggingface/hub/). No external API needed. CPU-only — works on a laptop.

Use uv tool install (or pipx install) instead of uvx rbi-source-mcp — uvx is a one-shot ephemeral runner; Claude Desktop restarting the server every time would re-resolve deps and re-load the model. uv tool install keeps a persistent venv with the model cached.

Verify the install

rbi-source-doctor                               # runs preflight: Python, sqlite-vec, corpus, HF cache, model

Should print all-green. If anything fails, the output tells you the specific fix.

Register with Claude Code (stdio)

claude mcp add rbi-source -s user -- rbi-source-mcp

Pin a corpus version (optional)

rbi-source-fetch-corpus defaults to the moving latest-corpus release. For reproducible deploys, pin to an immutable timestamped tag:

rbi-source-fetch-corpus --tag corpus-2026-05-04-1234567

Browse available tags at github.com/harshilmathur/RBI-Source-MCP/releases.

Cryptographic verification (optional)

By default the fetch script verifies SHA256 (mandatory). For full sigstore signature verification against the GitHub Actions OIDC identity:

pip install 'rbi-source-mcp[verify]'            # pulls sigstore-python (~25 MB)
rbi-source-fetch-corpus --verify-sigstore

This refuses to install any corpus that wasn't signed by the official corpus-release.yml workflow on main. Defends against PyPI compromise + tampered release assets.

Cloudflare Workers AI (alternative embedder)

If you'd rather embed via the Cloudflare API instead of running torch locally:

export RBI_EMBEDDING_PROVIDER=cloudflare
export RBI_EMBEDDING_MODEL=@cf/baai/bge-base-en-v1.5
export RBI_EMBEDDING_DIM=768
export CF_ACCOUNT_ID=...                        # https://dash.cloudflare.com → Account ID
export CF_API_TOKEN=...                         # https://dash.cloudflare.com/profile/api-tokens → "Workers AI: Read"

Both providers emit 768-dim vectors against the same bge-base-en-v1.5 model (BAAI weights for local; Cloudflare-hosted variant for the API). The prebuilt corpus matches; runtime + corpus mismatches are caught by rbi-source-doctor's corpus_runtime_match check before any retrieval happens.

Build the corpus yourself (advanced)

If you want a build fresher than the latest daily release (or want to fork the indexer):

git clone https://github.com/harshilmathur/RBI-Source-MCP.git
cd RBI-Source-MCP
uv sync
uv run rbi-source-crawl                         # crawl list pages
uv run rbi-source-index-all                     # index Master Directions
uv run rbi-source-index-all-circulars           # Standalone Circulars
uv run rbi-source-index-press-release --bulk    # Press Releases
uv run rbi-source-index-faq --bulk              # FAQs
uv run rbi-source-index-master-circular --bulk  # Master Circulars
uv run rbi-source-eval                          # gate at ≥80%

Crawl + index takes ~30-60 min. The same pipeline runs daily in .github/workflows/corpus-release.yml (diff mode, ~5 min) and publishes the result as a GitHub Release.

Run as an HTTP server (advanced)

For an internal team behind a reverse proxy:

rbi-source-mcp-http --host 0.0.0.0 --port 8080

Daily corpus refresh

The workflow at .github/workflows/corpus-release.yml runs daily at 02:00 UTC (diff mode: ~5 min, ~50 CF Neurons) plus a monthly full rebuild on the 1st at 03:00 UTC as a silent-re-upload safety net. Each run: crawl + index + smoke gate (≥80% absolute, <5pp regression) + sigstore-sign + publish to Releases. Self-hosters re-run rbi-source-fetch-corpus whenever they want fresh data; the latest-corpus alias always points at the most recent build.

The Releases page is auto-pruned: latest 14 dailies + latest 12 monthlies kept as immutable rollback targets, the rest deleted (last successful run takes ~30s for the GC pass).

Roadmap

Shipped surface is in CHANGELOG.md. What's next:

v0.9 (next release): Press Releases full archive (5y year-by-year crawl) · RBI Speeches crawler · pin GitHub Actions to commit SHAs · MCP Registry listing

v1.0+: Notifications archive (thousands of docs; year-iterating crawler) · document_versions historical snapshots → enable compare_versions and as_of queries · amendment chain extraction → unblocks find_updates and trace_relationships tools · OCR pipeline for scanned MDs · low-confidence threshold tuned with a fusion-score floor

Contributing

See CONTRIBUTING.md. Short version:

• Eval cases (src/rbi_source_mcp/eval/cases.py) — clause + expected-provision pairs from real fintech compliance work — are the highest-leverage contribution
• Corpus quality reports — file an issue with the "corpus quality" template, include the verbatim tool input + response
• Parser fixes for any of the five content-family list pages
• Topic-hint mapping additions in mcp/check_compliance.py
• New content families (open an issue first to align on scope)

Security

See SECURITY.md. Don't open public issues for vulnerabilities — use a private security advisory.

Telemetry

The hosted instance at rbi-source.harshil.ai captures anonymous server-side usage events to PostHog so I can see which tools get called, latency, and error rate. Self-host installs never phone home — telemetry only activates when POSTHOG_API_KEY is set in the environment, which it isn't in any path the OSS code ships with.

What the hosted instance captures per tool call:

• tool name (rbi_search / rbi_check_compliance / rbi_get_document / rbi_check_current)
• latency (ms), status (ok / error / input_too_large / db_unavailable)
• shape-only metadata: limit, has_filters, include_text, topic_hint enum value, length-bucket of the input (e.g. "500-2000")
• exception class name on failure, server version, region (when MCP_REGION or FLY_REGION is set)

What it does not capture: query strings, clause text, document IDs, response bodies, URLs passed to rbi_check_current, IP, or any client identifier. Events are anonymous ($process_person_profile: false) and disable_geoip=True is set on the SDK.

The website uses Cloudflare Web Analytics (cookieless, no consent banner needed).

To self-host with your own collector: pip install 'rbi-source-mcp[hosted]' (pulls the PostHog SDK) and set POSTHOG_API_KEY + optional POSTHOG_HOST. Leave them unset and you can pip install rbi-source-mcp without the extra — the default for every install.

License

MIT. See LICENSE.

Disclaimer

This tool surfaces RBI source material; it does not provide legal advice. Use citations to inform compliance review, not replace it. RBI publications are subject to amendment without notice; this tool's last_updated_at timestamps are best-effort. Always verify with a qualified compliance reviewer before acting on any provision returned.