Purpose
DocsGPT is an open-source platform developed by Arc53 for building private AI agents, assistants, and enterprise search on top of an organization's own documents. All answers cite sources, and it is licensed under MIT. It can be deployed in the cloud, on your own servers, or in a fully air-gapped (zero external network) environment, and supports hybrid mode (local platform + cloud LLM).
Main Capabilities
Knowledge and indexing: Supports more than 30 file formats, provides connectors for Google Drive, SharePoint, Confluence, GitHub, S3, and others, and supports OCR for scanned PDFs, wikis, and knowledge graphs.
Agents: Provides four agent types: Classic, Agentic, Research, and Workflow, supports scheduled tasks, connected tools, and arbitrary MCP servers; it can also act as an MCP server, allowing Claude or Cursor to retrieve its knowledge.
Workflows: A visual drag-and-drop builder, including nodes such as Start, AI Agent, Set State, If/Else, Code, Note, and End, with a step-by-step trace retained for each run.
Governance: 7 guardrail checks that can flag, redact, or block; human approval for answers and tool calls; roles, usage quotas, exportable audit logs, and run traces; OIDC SSO and SCIM 2.0 support through Admin.
How to Use
Quick start: macOS/Linux or Windows, requires Docker, and provides a one-command quick start method. Self-host using the community edition. The cloud version is available at app.docsgpt.cloud.
Use Cases
Aimed at regulated environments, suitable for organizations that need data residency, air-gapped deployment, or local inference. Officially disclosed use cases include: the UK Department for Work and Pensions (DWP) running an internal knowledge assistant on government infrastructure using local inference; AMD using it to build an internal technical documentation assistant; LexEU used for multilingual EU legal research covering more than 2.5 million pages in 6 EU languages, with retrieval times of 100–200 milliseconds; a large Central European commercial bank processing more than 10,000 incoming documents per day in a fully air-gapped environment (using existing NVIDIA T4 GPUs), reducing action time from 24 hours to 16 minutes, with classification accuracy exceeding 99.5%.
Limitations
Based only on existingmaterials, more features, pricing details, or platform compatibility cannot be confirmed; the official last updated time is 2026-10-04.
