Snipara Mcp

What is Snipara MCP?

Snipara MCP (snipara-mcp) is a lightweight stdio MCP connector that exposes Snipara’s hosted Project Intelligence APIs to MCP-compatible AI coding agents (Claude Code, Cursor, Codex, OpenAI Agents, etc.). It provides durable project context, reviewed memory, code-graph insights, and shared guidance so agents don’t start cold every session.

How to use it

• Quick run (no local install): uvx snipara-mcp.
• Pip install: pip install snipara-mcp (or pip install "snipara-mcp[rlm]" for runtime helper).
• Login & init: snipara login then snipara init to configure a project and MCP settings.
• Configure MCP clients: either point the client to the hosted HTTP endpoint (preferred if supported) or configure the client to run the local stdio command (e.g. uvx snipara-mcp) with environment variables:
  • SNIPARA_API_KEY, SNIPARA_PROJECT_ID (or SNIPARA_PROJECT_SLUG)
  • Optional: SNIPARA_API_URL, SNIPARA_TOOL_PROFILE (core or full)
• Example: set env vars and run the stdio server for clients that expect a local process.

Key features

• Small stdio MCP connector exposing the same compact default MCP contract as hosted Snipara.
• 13 core agent-callable tools (retrieval, memory, inbox review, context, code graph, etc.).
• Retrieval and recall APIs: snipara_context_query, snipara_get_chunk, snipara_recall, etc.
• Durable memory and structured end-of-task commits with a why block (decision, rationale, alternatives, constraints, outcome).
• Unified conversational Inbox review and agent-assisted decision proposals with human-approval apply semantics.
• Configurable code-impact traversal (depth, direction, edge kinds) and code-graph utilities (snipara_code_*).
• Agent-readable tool contracts, detailed behavior annotations, and safe parameter guidance.
• Credential-free discovery for initialize / tools/list (actual tool calls still require authentication).
• Security floor: requires MCP Python SDK >= 1.28.1 (session isolation, task ownership fixes, WebSocket Host/Origin validation).
• Thin, easy-to-audit client; heavy state and processing remain on the hosted Snipara service.

Use cases

1. Provide persistent project context to coding agents so they can make informed edits across sessions.
2. Let agents recall past decisions, reusable memory, and team guidance when automating code changes.
3. Plan and evaluate code impact and risky changes using code-graph queries and impact traversal.
4. Review and apply evidence-backed ProjectDecision drafts with human admin approval.
5. Support stdio-only MCP clients or local workflows where hosted HTTP streaming isn’t available.
6. Integrate with IDEs, agent orchestrators, and CI tools to surface project intelligence to LLM-driven workflows.

FAQ

• Q: When should I use the hosted HTTP endpoint vs snipara-mcp? A: Use hosted HTTP MCP when your client supports streamable HTTP (preferred). Use snipara-mcp when your client expects a local stdio process or for local compatibility.

• Q: How do I authenticate? A: Provide SNIPARA_API_KEY and SNIPARA_PROJECT_ID (or login via snipara login to store tokens in ~/.snipara/tokens.json). Tools/list may be visible before auth but actual calls require credentials.

• Q: What tools are available? A: Core tools include context retrieval, search, chunk read, recall/memory APIs, inbox review, document upload/sync, settings/index operations, and code-graph tools. Full specialist tools are available with SNIPARA_TOOL_PROFILE=full.

• Q: Is this secure for production? A: The connector enforces a security floor by requiring a recent MCP SDK with session isolation and host/origin checks. Authentication is mandatory for tool calls and the hosted service enforces access and project ownership checks.

• Q: How do I install and develop locally? A: Install via pip or run uvx snipara-mcp. For development: pip install -e ".[dev]", run tests with pytest, and lint with ruff.

• Q: License? A: MIT.