DocsGPT

DocsGPT is an open-source platform for building AI agents and assistants that answer from your documents, with source citations. Agents can search your knowledge, call tools and APIs, run workflows and schedules, and be reached from the web app, an API, widgets or chat integrations. You can run it on your own hardware with cloud or local models, or use DocsGPT Cloud .

A 32-second tour of DocsGPT on a demo product, the Kettleby Espresso One: a PDF handbook is dropped into Add knowledge and a public GitHub repository is added; an agent answers a question with its sources, a research agent works through its plan step by step, and a workflow routes a question through its nodes; an agent fills a todo list with a tool; the chat widget answers on a web page; a Python script streams an answer through the OpenAI-compatible API; the MCP Inspector calls the search_docs tool on DocsGPT’s MCP server; and docsgpt up starts a self-hosted install.

Get started

Quickstart Concepts Using the web app Add knowledge Choose a deployment Architecture

Build

Agent basics Workflow nodes Customizing prompts Connectors Tools basics Cloud providers

Integrate

API overview OpenAI-compatible API MCP server Chat widget Search widget Community integrations

Deploy and operate

Docker setup Kubernetes App configuration Security checklist Upgrading Troubleshooting

Features

Agents and workflows: Classic and research agents (plus agentic ones through the API or an imported agent YAML file), a visual workflow builder, schedules for recurring and one-time runs, and guardrails that flag, redact or block.
Knowledge: Upload PDF, Word, PowerPoint and Excel files (legacy formats included), OpenDocument, RTF, EPUB, Markdown, reStructuredText, HTML, CSV, JSON and text; images with OCR; and audio, which is transcribed. Add web pages, sitemaps and GitHub repositories, keep wikis the agent can edit, and tune chunking and retrieval per source, including hybrid search and GraphRAG.
Connectors: Sync Google Drive, SharePoint and OneDrive, Confluence, GitHub, Amazon S3, Linear and Reddit into knowledge. One connection can both sync content and give agents that service’s tools.
Grounded answers: Answers cite the sources they came from, and a strict prompt keeps the model to your documents.
Tools: Built-in tools, any REST API through the API tool, MCP servers, artifacts and sandboxed code execution, and shell commands on a paired remote device, with per-action approval.
Models: OpenAI, Anthropic, Google, Groq, OpenRouter, Novita and any OpenAI-compatible server such as Ollama or vLLM (cloud, local), plus per-user custom models and fallback models. Vectors go to FAISS, pgvector, Qdrant, Elasticsearch, Milvus or MongoDB Atlas.
APIs and integrations: The agent API with streaming, an OpenAI-compatible /v1 API, webhooks, an MCP server, a REST API with personal access tokens, HTML and React chat and search widgets, a Chatwoot bridge, and community bots for Discord, Slack and Telegram.
Teams and administration: SSO with OIDC and SCIM provisioning, roles and teams, usage quotas, analytics and logs, and OpenTelemetry observability.
Deployment: A one-command installer, Docker Compose, pip, Kubernetes and air-gapped installs.

DocsGPT is MIT-licensed and developed on GitHub . Contributions are welcome; see the contributing guide .