Categories
Self-Hosted AI Agents Knowledge Base MCP Server RAG Agent Memory Vector Search Semantic SearchStars
Forks
Watchers
Developer links
Connapse
Give your AI agents a library to search and a memory that survives the session: Connapse serves both over MCP and a REST API. Upload files or whole folders, or index them where they live: SFTP servers (including folders on your own computer), Amazon S3, Azure Blob Storage, GitHub, Confluence, Jira, or a custom source server. Queries run vector and keyword search together and fuse the rankings, an optional cross-encoder reranker re-scores the top results, and each passage comes back labeled with its source file. Agents save dated facts with their sources into personal or team memory spaces, recall them later, correct them without losing history, and forget them on request, while a Memory page shows each space's timeline. Results from GitHub, Confluence, Jira and custom sources are checked against each reader's own access, and a passage is left out whenever that check cannot be made. Optional link search follows a GitHub issue to its closing pull request or a Jira issue to its epic, and code tools find definitions, callers, and what a change could affect. More than twenty named tools suit Claude Code, Cursor, and VS Code; each key gets only the tools its scopes allow, and deletes can wait for approval. Embeddings run locally by default, with OpenAI or Azure OpenAI one setting away. Running on a dedicated VPS on RepoCloud with guaranteed CPU, RAM, and SSD, full root SSH access, and a browser serial console. MIT licensed.
Benefits
- Agents remember between sessions
- Facts are saved with their dates and the documents they came from, recalled in a later session, and corrected without erasing history, so decisions stop vanishing when a chat ends.
- One knowledge base for every agent
- Index uploads, SFTP folders, S3, Azure Blob, GitHub, Confluence, and Jira once, then Claude Code, Cursor, VS Code, or your own agent searches the same library over MCP or REST.
- Answers trimmed to each reader
- Results from GitHub, Confluence, Jira, and custom sources are checked against the asker's own access in the source system, and a passage is dropped whenever that check cannot be made.
- Runs locally with one command
- A single Compose command starts Connapse, generates its own passwords, and downloads a local embedding model, so nothing leaves your machine until you choose OpenAI or Azure OpenAI in Settings.
Features
- Hybrid search with reranking
- Vector and keyword search run together and fuse their rankings; an optional cross-encoder reranker raises nDCG@10 from 0.565 to 0.660 on a 50,000-document enterprise benchmark.
- Agent memory spaces
- Each fact keeps a subject, attribute, value, validity dates, and cited sources in a personal or team space; corrections keep history, and forgetting removes every version.
- MCP server and REST API
- More than 20 MCP tools cover search, documents, memory, code, containers, and approvals; 3 generic tools reach the REST API; each key sees only the tools its scopes allow.
- Source connectors
- SFTP, Amazon S3, Azure Blob Storage, GitHub, Confluence, Jira, and custom source servers stay searchable and in sync, with scope, sync interval, and pause or sync-now controls per source.
- Linked results and code tools
- Link search brings back records the top results point to, such as a closing pull request; code tools find definitions, references, and the impact of a change.
- Approval for agent deletes
- Held deletes become a batch the key's owner approves on the Approvals page; Connapse then deletes the approved items itself, and the agent reads each outcome with deletion_status.