Skip to content
Start free

Open-source AI agent platform for enterprise documents

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.

Book a demoStart free

v0.21.0 · MIT · github.com/arc53/DocsGPT ↗

Last updated:

docsgpt — product tour
docsgpt — product tour

What DocsGPT does

Documents go in; cited answers and governed actions come out.

  1. 01

    Knowledge

    30+ file formats, Google Drive, SharePoint/OneDrive, Confluence, Amazon S3, OCR for scans. Knowledge & Connectors →

    pdf · docx · xlsx · pptx · md · html · png · mp3

  2. 02

    Search & agents

    Cited answers, hybrid BM25 + vector, GraphRAG; Classic, Agentic, Research and Workflow agents. Search & Retrieval → · Agents →

    classic · hybrid · graphrag · citations: title · source · page

  3. 03

    Workflows & artifacts

    No-code builder with CEL routing; DOCX, PPTX, XLSX and PDF outputs from sandboxed code. Workflows → · Artifacts →

    ai agent · set state · condition · code · end

  4. 04

    Channels

    OpenAI-compatible API, chat and search widgets, Slack, Telegram and Discord bots. API, Widgets & Channels →

    POST /v1/chat/completions · widget · slack · telegram · discord

app.docsgpt.cloud — Agents › Edit
MetaContract reviewerAnswers questions about supplier contracts from the contracts library only, with citations.
SourceContracts library1,204 chunks · re-embedded 2026-09-06
Promptcontract-review-v3Add
ToolsCode ExecutorMemoryGoogle Search Tool
Agent typeClassic
publish · share link · api keySave
app.docsgpt.cloud — Sources › Add source
Select the way to add your source: Upload File, Crawler, Link, GitHub, Reddit, Google Drive, Amazon S3, New wiki.
app.docsgpt.cloud — Sources › Add source
Select the way to add your source: Upload File, Crawler, Link, GitHub, Reddit, Google Drive, Amazon S3, New wiki.
Workflow builder
  1. triggerDocument uploadtrigger
  2. ai agentClassifysmall model
  3. ai agentExtractwrites state
  4. conditionDecidecel over state
  5. endNotifytemplate
humanHuman approvalapprove · reject · logged
Text version

1. Document upload (trigger) → 2. Classify (small model) → 3. Extract (writes state) → 4. Decide (cel over state) → 5. Notify (template). Branch from Decide: else · confidence < 0.8 → Human approval → Notify.

Governed by design

Guardrails check content at four intervention points with seven built-in checks; human approval can hold any answer or tool call; identity comes from your IdP over OIDC and SCIM 2.0; every auth event lands in an append-only audit log; OpenTelemetry exports to your monitoring.

 Where it actsPage
GuardrailsUser input, retrieved sources, tool results, final answer — 4 intervention points, 7 checks; flag, redact or blockGuardrails →
Human approvalAny chat answer or tool call held for one-click reviewHuman-in-the-loop →
OIDC SSOSign-in with your identity providerAdmin →
SCIM 2.0IdP-driven user provisioning and deactivationAdmin →
Audit logAppend-only record of logins, denials, role changes, forced logoutsAudit log →
OpenTelemetryTraces and metrics to the collector you already runAnalytics →
app.docsgpt.cloud — Agent › Guardrails
Guardrails settings for an agent, 3 checks active: Enable guardrails on; Enforcement mode 'Monitor only' with the note that monitor mode records what would have happened without changing any answer; seven checks (Banned terms, Grounding in sources, Prompt injection, Personal information, Custom policy, Credentials and secrets, Link policy), each with its intervention points (User input, Answer, Retrieved sources, Tool result) and a Flag only action; Blocked-response message; Continue if a check fails; Check timeout 2000 ms.
app.docsgpt.cloud — Agent › Guardrails
Guardrails settings for an agent, 3 checks active: Enable guardrails on; Enforcement mode 'Monitor only' with the note that monitor mode records what would have happened without changing any answer; seven checks (Banned terms, Grounding in sources, Prompt injection, Personal information, Custom policy, Credentials and secrets, Link policy), each with its intervention points (User input, Answer, Retrieved sources, Tool result) and a Flag only action; Blocked-response message; Continue if a check fails; Check timeout 2000 ms.
app.docsgpt.cloud — Analytics
Analytics over 30 days: 217 messages, 858,949 tokens, 120 tool calls, 82% run success, feedback +29 / −9; charts of messages per day, token usage (generated and prompt), scheduled runs (completed, failed, skipped) and tool usage by tool.
app.docsgpt.cloud — Analytics
Analytics over 30 days: 217 messages, 858,949 tokens, 120 tool calls, 82% run success, feedback +29 / −9; charts of messages per day, token usage (generated and prompt), scheduled runs (completed, failed, skipped) and tool usage by tool.
app.docsgpt.cloud — Logs
Logs view: filters for level and event type, a search box, and a timeline of chat answers, scheduled runs, a workflow run and a webhook run with ISO timestamps.
app.docsgpt.cloud — Logs
Logs view: filters for level and event type, a search box, and a timeline of chat answers, scheduled runs, a workflow run and a webhook run with ISO timestamps.

Runs where your data is

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 itExternal networkDetails
CloudArc53, multi-tenantYesCloud deployment →
ManagedArc53, dedicated instanceYesManaged deployment →
On-premisesCustomer, with Arc53 supportOptionalOn-premises deployment →
Air-gappedCustomer, with Arc53 supportNoneAir-gapped deployment →

The same platform, from one seat on every posture. Only the operator and the network boundary change.

Open source, model-agnostic

MIT licence, a public repository and no model lock-in: cloud providers and local engines behind one Model Registry.

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 seven vector stores, so the stack can be swapped without rewriting the application.

ollama · vllm · llama.cpp · faiss · pgvector · qdrant · milvus · elasticsearch · lancedb · mongodb atlas

Self-host in 5 minutesModels & BYOM →

GitHub stars

18,273

on GitHub

Contributors

226+

Licence

MIT

Vector stores

7

Used by people from

  • AMD
  • Microsoft
  • Siemens
  • Deloitte
  • AstraZeneca
  • PwC
  • HM Government
  • Accenture
  • Payoneer
  • Intel
  • Airbnb
  • Amazon
  • Stanford
  • MIT

Example

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

Documentation

Product areas

Agents

Classic, Agentic, Research and Workflow agents

/product/agents

Workflows

Visual builder with conditions, code and state

/product/workflows

Models & BYOM

Cloud providers and local engines, hot swap

/product/models

CLI

ask, chat, agentic mode, bench

/product/cli

Deployment options

Cloud · Managed · On-premises · Air-gapped — who operates it, where data sits, what leaves the network.

/deploy

Frequently asked questions

What is DocsGPT?

DocsGPT is an open-source platform for private AI agents, assistants and enterprise search over an organisation's own documents, built by Arc53 and deployable in the cloud, on your own servers or fully air-gapped. Every answer cites the title, source and page it came from.

Is DocsGPT free and open source?

Yes. The platform is MIT-licensed and can be self-hosted at no cost. Cloud seats, Managed instances and Enterprise (on-premises and air-gapped) deployments are paid; see pricing.

Can DocsGPT run on-premise or air-gapped?

Yes. With a local LLM, local embeddings and a self-hosted vector store, the whole stack runs inside your network with no external egress. See air-gapped deployment.

Does DocsGPT cite its sources?

Yes. Every answer carries the title, source and page number of the passages it used, and the source panel opens on the cited page.

Which LLMs does DocsGPT support?

Cloud providers (OpenAI, Anthropic, Google, Azure and others) and local engines (Ollama, vLLM, llama.cpp and others), swappable from the Model Registry. See Models & BYOM.

See DocsGPT on your documents

A 30-minute demo on your use case, or a free on-premises review with an engineer.

Book a demoStart freeSee pricing