Market and Knowledge Base

Install an integration from the Market

1. Open the Market and synchronize its catalog. Catalog synchronization also runs when the view opens.
2. Choose an integration and check its runtime requirements.
3. Install the package, fill in required settings and secrets, and attach it to a project.
4. Check server health before using it in a client.

Catalog definitions can come from a remote JSON manifest or an uploaded local manifest. The default Market catalog is https://mcpbox.sh/catalog.json; it is separate from the CLI catalog.

A package is reusable across projects. Each project attachment creates a managed MCP server with that project's settings. Check existing attachments before uninstalling a shared package.

Manifests can describe required binaries, environment variables, secret fields and post-install health checks. Missing system dependencies can block installation. The go_install strategy needs Go and an installable Go module with a concrete executable. Docker-backed entries require Docker and use a limited docker run --rm -i workflow; the Market is not a Compose orchestrator.

Create a local knowledge collection

1. Open Knowledge Base and create a collection.
2. Choose the source directory on the machine running MCPBox.
3. Index the collection and review the result.
4. Search a known term to check that the expected files appear.
5. Connect the collection in the project's RAG view.

Collections are global: one collection can be connected to several projects, and a project can use several collections. Disconnecting a project link and deleting a collection are different operations; check all consumers before deleting shared knowledge.

Supported local files

The local index currently accepts .go, .js, .php, .md, .txt, .csv, .xlsx, .docx, .pptx and .pdf. PDFs need a text layer. The built-in extractor does not perform OCR on scans or images.

Common generated directories such as .git, node_modules, vendor, dist and build are skipped. Results contain source paths and, when available, sheet, slide or page hints. Local indexing uses Bleve full-text search, without mandatory embeddings or a vector database. Search for concrete terms found in the documents; this mode does not provide semantic retrieval.

Enable automatic reindexing when creating or editing a bleve_only collection if you want MCPBox to scan it periodically (a ten-minute loop). It is off by default and is not enabled for RagBox-only or dual-service collections. After changing files, you can also trigger indexing manually.

Collection modes and uploaded files

A collection can use bleve_only (local), ragbox_only (external), or dual (both). For RagBox-backed collections, select an existing workspace or configure the vector connection needed by your deployment.

The collection card also supports file upload. The keep mode retains the uploaded source in the managed collection folder; transient removes the temporary source after processing. Transient sources are unavailable for a later full reindex, so retain your original documents elsewhere. Review both local and RagBox processing status: an accepted external ingestion job can still be processing.

Search from an AI client

Connected collections are available through the built-in search_project_knowledge MCP tool. Connect the collection to the project used by the client, refresh tools and ask the client to search the knowledge base. Creating a global collection alone does not attach it to every project.

Use RagBox when you need external retrieval

The optional RagBox integration connects separately deployed ingestion and search services. Configure the service URL, tenant, API token and optional admin token, then synchronize and bind knowledge collections. Detailed setup and modes are in external integrations.

MCPBox's local Bleve search and RagBox retrieval are distinct paths. Available semantic or vector behavior depends on the connected RagBox deployment and its configuration.

Refreshed Knowledge Base interface

Cards are compact: ! opens help or collection details; Search opens one query/results dialog, including loading and errors. Closing it during a request does not reopen it when results arrive.

If upload/reindex fails with missing bearer token, processing is not confirmed. This is a known server-side limitation, including custom admin-port setups; retain the original and inspect Logs. The interface refresh does not fix this backend error.