Connapse is an open-source, self-hostable AI agent knowledge backend, with the goal of providing AI agents with persistent, searchable memory, avoiding the problem of starting from scratch in every new session with no access to historical research and documents.
Purpose
Connapse enables AI agents to query and build their own knowledge base: documents can be uploaded, or it can be pointed at an existing Amazon S3 bucket or local file system, and then agents query through MCP tools, REST API, or CLI to obtain structured answers with source citations. It is suitable for AI agent developers who need a knowledge backend that can both be queried and built, .NET/Azure teams who want a RAG platform that fits their existing tech stack and cloud identity, and enterprise teams that need project-isolated knowledge bases, RBAC, and audit trails.
Main Capabilities
- Container-isolated knowledge: Each project has its own independent vector index, storage connector, and search configuration, with no cross-contamination between projects, teams, or customers.
- Hybrid search: Combines vector similarity with keyword full-text search, supporting configurable fusion strategies (convex combination, DBSF, AutoCut).
- Multi-provider AI: Embeddings and LLMs can be switched among Ollama, OpenAI, Azure OpenAI, and Anthropic, with runtime, per-container switching and no restart required.
- Index existing storage: Connect to MinIO, a local file system (real-time file monitoring), or Amazon S3 (IAM authentication), and files can remain in place.
- Four access methods: Web UI, REST API, CLI (native binary), and an MCP server for Claude.
- Enterprise-grade authentication: Multi-level RBAC (Cookie + OAuth 2.1 + PAT + JWT), with AWS identity linking support; after identity linking is configured, per-user search result filtering is performed according to AWS permissions (optional, unrestricted when not configured).
- One-click deployment: Docker Compose integrates PostgreSQL + pgvector, MinIO, and optional Ollama, with built-in structured audit logs and rate limiting.
- Multi-format ingestion: PDF, Office documents, Markdown, and plain text are automatically parsed, chunked, and embedded.
- Real-time processing: Post-Platform ingestion, with real-time progress updates via SignalR.
- Runtime configuration: Change chunking strategies, embedding models, and search settings per container without restarting.
- Cloud identity linking: AWS IAM Identity Center (SAML) identity linking; once configured, per-user filtering based on AWS permissions (optional, unrestricted when not configured).
- Invite-only access: Administrators control user registration, with four roles (Admin / Editor / Viewer / Agent).
- Agent management: Dedicated agent entities with API key lifecycle, scoped permissions, and audit trails.
- Audit logging: Structured auditing of uploads, deletions, container operations, and authentication events.
- CLI distribution: Native self-contained binaries (Windows/Linux/macOS) as well as a .NET global tool via NuGet.
- Cross-model search: Switch embedding models mid-project, with old vectors automatically falling back from semantic to hybrid search.
Usage
- Deployment: Requires Docker and Docker Compose, .NET 10 SDK (for development), and optionally Ollama (for local embeddings). After running Docker Compose, the first startup pulls images (about 2-5 minutes), initializes PostgreSQL and the pgvector extension and runs EF Core migrations, creates the MinIO bucket, seeds the administrator account, and starts the web application.
- Integration with MCP clients such as Claude: Create an agent API key in the Web UI under Settings → Agent API Keys, then add the corresponding snippet to the client configuration (such as claude_desktop_config.json, VS Code, or Cursor MCP configuration). The MCP server supports Streamable HTTP transport and is compatible with clients such as Claude Desktop, Claude Code, VS Code, and Cursor.
- CLI: The connapse CLI is released separately at Destrayon/connapse-cli and supports authentication commands (auth login, auth whoami, auth pat), etc.
- Example MCP tool usage: Create containers, upload documents, search container contents, list files, retrieve full text, delete and update files, delete folders, view document and chunk statistics, etc.
Use cases
- AI agent developers need agents that can both query and build knowledge bases.
- .NET / Azure teams want a RAG platform that fits their existing tech stack and cloud identity.
- Enterprise teams need project-isolated knowledge bases, RBAC, and audit trails.
- Those who don't want to repeatedly upload files and want to point directly at existing Amazon S3 or file system storage.
Limitations
- The project is under active development, currently at v0.3.2, approaching production readiness, and has not yet released a stable 1.0 version.
- The Azure Blob Storage connector and Azure AD cloud identity linking were removed after the v0.3.0 release (#476), and the implementation of the AWS IAM Identity Center device authentication flow was also removed (#435); per-user AWS search filtering was rebuilt based on ISearchScopeResolver + AWS S3 Access Grants, is optional, and fails closed once AWS identity linking is configured; deployments without it configured have no per-user filtering.
- Production environments need to set a strong Identity__Jwt__Secret.
- Uploads have a file type allowlist, and the maximum file size depends on server configuration; after a document is uploaded, it takes time to be embedded before it can be searched.
- Common issues: Docker not running or port conflicts cause connection refused at localhost:5001; API keys are shown only once at creation; MCP client configuration changes require a restart.
Security and licensing
- The security policy is detailed in SECURITY.md, including security limitations and the roadmap.
- MIT license, permitting commercial and private use, with the only requirement being to include the copyright notice and license in substantial portions of the software.
Roadmap
- v0.4.0: Communication connectors (Slack, Discord).
- v0.5.0: Knowledge platform connectors (Notion, Confluence, GitHub).
- v1.0.0: Production-ready stable version.
