AnythingLLM
A local-first AI workspace for chatting with documents, building retrieval knowledge bases, switching model providers and connecting Agent or MCP tools from a desktop or Docker deployment.
What AnythingLLM provides
AnythingLLM brings document chat, workspaces, model selection, embeddings, vector storage, Agent, MCP and workflow features into one interface. The smallest useful test is a workspace with one non-sensitive document, a local or limited model and a check of the retrieved citations before real material is imported.
Data paths and permissions
A local-first interface does not mean every request is local. Cloud models, embeddings, vector databases, speech, model downloads, external tools and MCP servers can create separate network paths. Magic features may request microphone, screen or accessibility access, while Agents and tools may reach files or commands.
Maintenance note
This page reviews AnythingLLM 1.15.0, desktop and Docker modes, RAG retrieval, local and cloud providers, Magic permissions, Agent/MCP boundaries, telemetry and the MIT license. Content review date: 2026-08-23.
Save to your cloud drive
Open the cloud drive to get the file directly, or save it for convenient access on another device.
Quark Cloud Drive
RecommendedSave AnythingLLM to this cloud drive
Baidu Netdisk
Save AnythingLLM to this cloud drive
AnythingLLM 1.15.0 local knowledge base and Agent setup guide
Create an isolated test workspace, verify parsing, embeddings and citations, and only then connect real documents, cloud providers, Agents or system-level Magic features.
Before you start
- Check the Windows, macOS or Linux architecture and reserve disk space for the app, models and vector indexes.
- Prepare non-sensitive test documents and decide which LLM, embedding and vector database components are local or cloud-hosted.
- Keep Agent, MCP, microphone, screen and accessibility permissions off until a specific task needs them.
Installation steps
- 01
Install the matching 1.15.0 build
Select a signed desktop package or a controlled Docker deployment, verify the version and review the initial privacy and telemetry settings.
- 02
Build a minimal workspace
Choose a test model, local embedding and vector store, import one non-sensitive document and check indexing, citations and deletion.
- 03
Trace the data path
Inspect the chat model, embedding, vector, speech and tool settings separately, and use low-privilege, revocable credentials for external providers.
Quick start
- 01
Test questions against citations
Ask questions whose answers can be located in the source, inspect the cited chunks and treat the source document as authoritative when the model disagrees.
- 02
Limit Agent and MCP tools
Enable one trusted tool at a time, restrict file directories, network and commands, and observe harmless tasks before widening access.
- 03
Enable Magic only when needed
Test dictation, selected-text and completion features without sensitive windows, review screen and accessibility permissions and turn them off afterward.
Usage tips
- Retrieval quality depends on parsing, chunking, embeddings, search and generation together; replacing the model alone rarely fixes a broken index.
- Desktop mode is mainly single-user; team permissions and server isolation belong in a carefully configured Docker deployment.
- Anonymous telemetry can be disabled, but cloud providers, model downloads, tools and MCP servers still have their own network behavior.
Troubleshooting and uninstall
Why does a document answer have no useful citation?
Inspect parsing, chunk size, embedding model, workspace binding and source encoding with a small document, then re-index before changing the generation model.
Why does a local deployment still make network requests?
Check telemetry, updates, model and embedding downloads, cloud providers, speech, Agents, MCP and external tools separately, then revoke exposed credentials immediately.
- Export settings and revoke credentialsSave the workspaces and source-document list that must remain, revoke cloud API keys and MCP tokens, and do not treat a vector index as an original-document backup.
- Remove the app and local dataStop AnythingLLM, uninstall the desktop app or tear down the controlled deployment, then clear chats, indexes, models, caches and logs only after checking retention needs.
Frequently asked questions
Does local-first mean documents never leave the computer?
No. Local models and embeddings can stay on the device, but cloud providers, speech, model downloads, vector services, Agents and MCP tools may send data elsewhere.
Is an AnythingLLM vector index a backup of the source document?
No. Keep the original documents and a tested backup separately; an index may be deleted, re-created or tied to a particular embedding model.
Are Agents and MCP tools just interface plugins?
They are executable integrations that may access files, commands or networks. Install trusted components, constrain their scope and review logs before using them with real data.