DocsGPT is an open-source (MIT) AI agent platform for enterprise documents, built by Arc53. It connects an organisation's files and systems to private AI assistants, agents and workflows with cited answers, and runs in the cloud, on your own servers or fully air-gapped.
DocsGPT 1.0.0 · MIT · github.com/arc53/DocsGPT ↗ · What's new ↗
Last updated:
Documents go in; cited answers and governed actions come out.
Knowledge
30+ file formats with OCR for scans, websites, wikis your agents can edit and knowledge graphs; every answer cites its passage. Knowledge →
pdf · docx · xlsx · pptx · md · html · png · mp3
Connectors
Google Drive, SharePoint, Confluence, GitHub, Notion, Linear and any MCP server: sync into Knowledge, give agents tools, or both. Connectors →
sync · agent tools · allow · ask first · off
Search
Cited answers with a citation reader, hybrid BM25 + vector and GraphRAG, testable before anyone relies on it. Search & Retrieval →
classic · hybrid · graphrag · citations: title · source · page
Agents
Classic, Agentic, Research and Workflow agents with built-in and connected tools, schedules and team sharing. Agents →
tools · schedules · teams · api key
Workflows & artifacts
No-code builder with If / Else branches, code and shared state; DOCX, PPTX, XLSX and PDF outputs. Workflows → · Artifacts →
start · ai agent · set state · if / else · code · end
Channels
OpenAI-compatible API, an MCP server for Claude and Cursor, chat and search widgets, Slack, Telegram and Discord bots. API, MCP & Widgets →
/v1/chat/completions · /mcp · widget · slack · telegram
1. Start (chat · api · webhook · schedule) → 2. Classify (small model) → 3. Extract (writes state) → 4. Decide (on shared state) → 5. Notify (template). Branch from Decide: else · confidence < 0.8 → Human approval → Notify.
Guardrails check content at four intervention points under a company-wide floor; human approval can hold any answer or tool call; admins govern connectors, roles and quotas; identity comes from your IdP over OIDC and SCIM 2.0; every sign-in, change and guardrail decision lands in one exportable Activity feed; every run leaves a trace.
| Where it acts | Page | |
|---|---|---|
| Guardrails | User input, retrieved sources, tool results and the final answer: 4 intervention points, 7 checks, and a company-wide floor no agent can weaken | Guardrails → |
| Human approval | Any chat answer or tool call held for one-click review; changes through connected services ask first | Human-in-the-loop → |
| Connector governance | Admins choose which connectors members may use, whose account shared tools run with, and whether agents may write | Connectors → |
| SSO and SCIM | Sign-in with your identity provider; accounts created and removed by your directory | Admin → |
| Roles and sharing | Instance and team roles; Viewer or Editor on every agent, knowledge source, prompt and tool | Admin → |
| Activity feed | One append-only timeline of sign-ins, changes and guardrail decisions, exported as CSV or NDJSON | Activity → |
| Usage quotas | Token and spend limits per user, with team and instance defaults | Quotas → |
| Traces | Every run as a step-by-step trace, exported to Langfuse, Datadog, Grafana or any OpenTelemetry backend | Traces → |
| Access tokens | Scoped, expiring tokens for scripts, the CLI and CI pipelines | API → |
Cloud and Managed are operated by Arc53 and use an external network. On-premises is operated by the customer with egress optional (a hybrid on-prem platform with a cloud LLM is on-premises, not air-gapped). Air-gapped is operated by the customer with no external network at all.
| Who operates it | External network | Details | |
|---|---|---|---|
| Cloud | Arc53, multi-tenant | Yes | Cloud deployment → |
| Managed | Arc53, dedicated instance | Yes | Managed deployment → |
| On-premises | Customer, with Arc53 support | Optional | On-premises deployment → |
| Air-gapped | Customer, with Arc53 support | None | Air-gapped deployment → |
The same platform, from one seat on every posture. Only the operator and the network boundary change. Compare deployment options →
MIT licence, a public repository and no model lock-in: cloud providers and local engines, picked per chat or per agent.
DocsGPT is released under the MIT licence at github.com/arc53/DocsGPT. It runs on cloud providers or on local engines such as Ollama, vLLM and llama.cpp, and stores vectors in any of six vector stores, so the stack can be swapped without rewriting the application.
ollama · vllm · llama.cpp · faiss · pgvector · qdrant · milvus · elasticsearch · mongodb atlas
GitHub stars
18,317
on GitHub
Contributors
239+
Licence
MIT
Vector stores
6
Used by people from
24 h → 16 min
A large Central European commercial bank processes 10,000+ inbound documents a day; time-to-action fell from 24 hours to 16 minutes, with over 99.5 % classification accuracy, running fully air-gapped on the bank's existing NVIDIA T4 GPUs.
Source: EU bank, name under NDA
Uploads, links, crawler, wiki; 30+ formats, OCR, chunking
/product/knowledge
Cited answers, hybrid search, GraphRAG
/product/search
Classic, Agentic, Research and Workflow agents
/product/agents
Visual builder with conditions, code and state
/product/workflows
DOCX, PPTX, XLSX, PDF outputs; sandboxed Python
/product/artifacts
Drive, SharePoint, Confluence, GitHub, MCP; sync and agent tools
/product/connectors
Checks, redaction, approval gates, audit
/product/guardrails
YAML test suites, LLM-as-judge, CI output
/product/benchmarking
Cloud providers, local engines, Custom Models, fallback
/product/models
OpenAI-compatible API, MCP server, chat and search widgets, bots
/product/integrations
SSO, SCIM, roles, Activity feed, quotas, traces
/product/admin
Chat, local tools, agents as code, bench
/product/cli
A 30-minute demo on your use case, or a free on-premises review with an engineer.