Docs / Getting Started

Get Snipara running on one project, then keep that project in scope.

Start by choosing one of the two primary surfaces: Hosted MCP for agents, API / SDK for apps and integrators. Keep local helper tooling optional.

1. Create a project

Use the dashboard to name the workspace you want Snipara to remember.

2. Connect your client

Use Hosted MCP for LLM agents, or API / SDK for software integrations.

3. Query the right scope

Use docs and parsed binaries for narrative context, and the code graph for structure-heavy questions.

Primary path

For LLM agents: use Hosted MCP

Use this for Claude Code, Codex, Cursor, Windsurf, ChatGPT Desktop, and any MCP-compatible assistant.

Primary path

For product integration: use API / SDK

Use this for backend-to-backend integrations, bulk upload, sync, provisioning, health, and analytics.

Recommended start

Use Hosted MCP when you want an agent connected quickly.

Add the hosted endpoint to the MCP client you already use. No local package is required unless you need stdio compatibility, browser-login testing, or local helper workflows.

https://api.snipara.com/mcp/YOUR_PROJECT_SLUG

Use create-snipara when you want scaffolding. Use snipara-companion only when local terminal workflows help.

Keep it high level

Snipara does not replace your LLM. It prepares the context you send to it.
Hosted MCP is the canonical agent surface.
API / SDK is the canonical app and integrator surface.
`rlm_context_query` stays the default for docs and narrative knowledge.
Memory stores durable decisions and preferences; it is not the document source of truth.
Code structure lives in the code graph, not in a generic doc search result.
Supported PDF, DOCX, PPTX, SVG, and VSDX files are parsed into searchable text during GitHub sync.
Shared context is project-aware, so the same team can behave differently per project.

Code Graph v1

In v1, rlm_context_query stays doc-first. It remains the right default for docs, notes, markdown, and narrative context.

For structural code questions, use rlm_code_callers, rlm_code_imports, rlm_code_neighbors, or rlm_code_shortest_path.

Manual path

HTTP direct is the cleanest hosted setup.

If your client accepts remote MCP, use the project slug and an API key. Keep the slug as the human-readable identifier across all transports.

{
  "mcpServers": {
    "snipara": {
      "type": "url",
      "url": "https://api.snipara.com/mcp/YOUR_PROJECT_SLUG",
      "headers": {
        "X-API-Key": "rlm_YOUR_API_KEY"
      }
    }
  }
}

See integrations Integrate with API

Integration guide

Compare hosted HTTP, SSE, and local stdio without changing the story every time.

Also useful

Shared context

Learn how reviewed memory and project-scoped context travel between runs.