roshan-baaz-mcp

A self-hostable Model Context Protocol server for Roshan AI's Baaz (باز) Persian semantic-search service.

Built from the public API documentation at docs.roshan-ai.ir. Unofficial community integration.

What is this?

Baaz (باز) is Roshan AI's Persian-native semantic search and retrieval service. You give it documents; it chunks them, computes embeddings, and stores them across an Elasticsearch + Weaviate backend. You can then run semantic queries, find similar documents, check document status, fetch full documents, and manage indices — all tuned for Persian (فارسی) content.

roshan-baaz-mcp wraps that HTTP API as MCP tools so Claude — or any MCP-compatible client — can call Baaz as first-class tools.

Baaz is self-hostable, and a single deployment is rarely enough: teams run many independent Baaz instances (per region, per tenant, per environment). This server treats named instances as a core concept — one MCP process can route every tool call to the right Baaz deployment via an optional instance argument.

Highlights

• 100% endpoint coverage — every documented Baaz endpoint is a tool.

• Multi-instance / self-hosting first — route by instance; tokens are never exposed by list_instances.

• Guardrails — http(s) URL validation, ~10MB body cap, pagination clamps, and token redaction in every error path.

• Offline docs tool — roshan_baaz_docs describes the service and tools without any network call.

Related MCP server: rag-mcp-server

Tool reference

Every tool accepts an optional instance: str | None selecting which configured Baaz deployment to call (defaults to default_instance).

See the Baaz API guide for full request/response field semantics.

Install

Requires Python 3.10+ (developed and tested on 3.11).

git clone https://github.com/roshan-research/roshan-baaz-mcp.git
cd roshan-baaz-mcp
python -m venv .venv && source .venv/bin/activate
pip install -e .            # add ".[dev]" for the test/lint toolchain

Run it (stdio transport by default):

python -m roshan_baaz_mcp            # or: roshan-baaz-mcp
python -m roshan_baaz_mcp --help     # transports, host/port, log level

Configuration (multi-instance)

Configuration is read from environment variables via pydantic-settings.

Shorthand (single instance)

The fastest way to point at one Baaz deployment. This synthesizes an instance named default:

export ROSHAN_BAAZ_BASE_URL=https://baaz.roshan-ai.ir
export ROSHAN_BAAZ_TOKEN=your-secret-token

Nested (one or many named instances)

Prefix ROSHAN_BAAZ__, nested delimiter __. Each instance under INSTANCES__<NAME>__:

# Two self-hosted regions behind one MCP server
export ROSHAN_BAAZ__INSTANCES__TEHRAN__BASE_URL=https://baaz.tehran.example.ir
export ROSHAN_BAAZ__INSTANCES__TEHRAN__TOKEN=tehran-token
export ROSHAN_BAAZ__INSTANCES__MASHHAD__BASE_URL=https://baaz.mashhad.example.ir
export ROSHAN_BAAZ__INSTANCES__MASHHAD__TOKEN=mashhad-token
export ROSHAN_BAAZ__DEFAULT_INSTANCE=tehran

Tools then target a deployment with instance="mashhad"; omit it to use the default. Call list_instances to discover configured names (tokens are never returned).

Use with an MCP client

Add the server to your MCP client (e.g. Claude Desktop) config:

{
  "mcpServers": {
    "roshan-baaz": {
      "command": "python",
      "args": ["-m", "roshan_baaz_mcp"],
      "env": {
        "ROSHAN_BAAZ_BASE_URL": "https://baaz.roshan-ai.ir",
        "ROSHAN_BAAZ_TOKEN": "your-secret-token"
      }
    }
  }
}

Architecture

The MCP client talks to roshan-baaz-mcp over a transport (stdio by default). The server resolves the requested instance, applies guardrails, and calls the Baaz HTTP API.

Request flow

A typical retrieval session: index documents, run a semantic query, read the ranked results, then optionally expand with similar documents.

Self-hosting & scaling

Because Baaz is self-hosted, one MCP process can front many Baaz deployments, routing each tool call by its instance argument (region / tenant / environment).

Scaling notes:

• Stateless — the server holds no per-request state; run as many replicas as you need behind your scheduler. Configuration comes entirely from the environment.

• One process, many instances — add instances by setting more ROSHAN_BAAZ__INSTANCES__<NAME>__* variables; no code changes.

• Rate limits — Baaz limits to ~100 requests/minute per deployment; spread heavy load across instances and back off on 429.

• Body cap — requests are capped at ~10MB (enforced client-side before the call); batch large indexing jobs accordingly.

• Horizontal autoscaling — a Kubernetes HPA manifest and Helm chart ship in deploy/.

Deployment

Ready-to-use manifests live in deploy/:

• Docker — Dockerfile and a two-instance docker-compose.yml.

• Kubernetes — raw manifests in deploy/kubernetes/ (namespace, configmap, secret, deployment, service, ingress, hpa, kustomization).

• Helm — chart in deploy/helm/roshan-baaz-mcp/.

• Terraform — example in deploy/terraform/.

See deploy/README.md for details.

Testing

pip install -e ".[dev]"
python -m pytest -q          # all HTTP is mocked with respx; no network
ruff check src tests

Optional live tests run against a real Baaz deployment when you opt in:

export ROSHAN_BAAZ_LIVE=1
export ROSHAN_BAAZ_BASE_URL=https://baaz.roshan-ai.ir
export ROSHAN_BAAZ_TOKEN=your-secret-token
python -m pytest tests/live -q

There is also a no-network smoke test that asserts every tool has a non-empty description and an instance parameter:

python scripts/smoke_test.py

Examples

examples/inspect_server.py prints the registered tools and their schemas without any network call. See examples/README.md.

License

MIT. Baaz (باز) and Roshan AI are products/trademarks of Roshan AI; this is an unofficial, community-maintained integration built from public docs.