Web
Search and read the web
Two search providers and a page reader that returns Markdown.
Brave Search · DuckDuckGo Search · Read Webpage (URL → Markdown)
Private AI agents over company data
DocsGPT agents are private AI agents over company data. Each one combines a knowledge source, a system prompt and a set of tools, and comes in one of four types: Classic, Agentic, Research or Workflow. Agents are published with a share link, an API key or a schedule, and run inside your own deployment.
classic · agentic · research · workflow · mcp · cron · share link · api key
Last updated:
docsgpt — research agent
- clarify scope · active contracts, EU counterparties
- search contracts · automated decision-making, profiling
- search regulation · deployer obligations, arts. 26 and 50
- synthesise with citations
Four agent types
Every agent is one of four types. The type decides when retrieval happens and how many steps the model may take; knowledge, prompt and tools are set the same way for all four.
Whatever the type, any answer or tool call can be held for human approval before it goes out. See Guardrails and human-in-the-loop.
type · knowledge · system prompt · tools · publish
| What it does | Retrieval | Fits | |
|---|---|---|---|
| Classic | RAG Q&A with pre-fetched chunks, plus tool use | Chunks retrieved before the model runs | Assistants over a known knowledge base |
| Agentic | LLM decides when to retrieve (internal_search tool); multi-step tool calling | On demand, as a tool | Tasks that mix retrieval with actions |
| Research | Clarify → plan → research → synthesise with citations; budget controls (steps, timeouts, tokens) | Iterative, planned | Long questions across many sources |
| Workflow | Executes predefined node graphs with branching and shared state — Workflows → | Per node | Deterministic document processes |
Classic agents answer with pre-fetched chunks plus tool use. Agentic agents let the model decide when to retrieve (internal_search) and call tools over several steps. Research agents clarify, plan, research and synthesise with citations under step, timeout and token budgets. Workflow agents execute predefined node graphs with branching and shared state.
Tools
Fourteen built-in tools cover web search, databases, REST APIs, MCP servers, remote machines, messaging, memory and image generation. Anything else is a custom tool written with the Python SDK. Tool calls can be gated by human approval, and Remote Device commands additionally pass approval modes, a denylist and an audit log.
Data and systems
Databases, REST APIs, MCP servers
Query PostgreSQL; call any REST endpoint with LLM-filled parameters, no code; connect remote MCP servers with none, Bearer, API-key, Basic or OAuth auth (v0.15).
PostgreSQL Database · API Tool · MCP Tool
Devices and messaging
Machines and notifications
Operate a paired machine through the docsgpt-cli daemon, with approval modes, a denylist and an audit log; send messages and push notifications.
Remote Device · Telegram · Ntfy push
Working memory
Memory, notes and tasks
Persist facts across turns, keep scratch notes and track a task list inside a run.
Memory · Notepad · Todo List
Generation
Images
Generate images from a prompt inside a conversation or run.
image generation
Custom
Your own tools
Write a tool in Python with the SDK and register it like any built-in.
python sdk
Scheduled agents
Agents run on a cron schedule or once at a set time, backed by RedBeat. An agent can also schedule itself with the scheduler tool, using relative ("30m") or ISO-8601 times.
The Schedules UI validates each schedule, auto-pauses schedules that run away and keeps a history per run.
Scheduling
- Schedule types
- cron · one-off (RedBeat)
- Scheduler tool
- relative ("30m") or ISO-8601 times
- Schedules UI
- validation · auto-pause for runaway schedules · per-run history
- Typical uses
- nightly reports · hourly compliance monitoring · re-indexing
| Cron | Next run | Last runs | |
|---|---|---|---|
| weekly-compliance-digest | 0 7 * * 1 | 2026-09-14 07:00 | ✓ ✓ ✓ ✓ · 12 s avg |
| nightly-contract-scan | 0 2 * * * | 2026-09-08 02:00 | ✓ ✓ ✗ ✓ · 1 timeout |
| monthly-vendor-review | 0 9 1 * * | 2026-10-01 09:00 | ✓ ✓ ✓ · 41 s avg |
Publish, share, govern
An agent moves from draft to published. Once published it can be shared by public link, called with a per-agent API key (with rotation, v0.19) or a webhook, and scheduled.
Agents export and import as YAML with secrets stripped; organisations ship templates in premade_agents.yaml; agent folders keep a large estate organised.
draft → publish · share link · api key (rotation) · webhook · yaml · premade_agents.yaml · folders
Agent lifecycle
- 01Draftknowledge · prompt · tools · type
- 02Publishdraft → published
- 03Sharelink · api key · webhook · schedule
- 04Exportyaml, secrets stripped
Text version
An agent is drafted with its knowledge, prompt, tools and type, then published. A published agent is shared by public link, per-agent API key or webhook, or put on a schedule. It can be exported as YAML with secrets stripped and imported into another instance.
How it runs privately
Agents, their tools and their schedules execute in the DocsGPT backend, inside your deployment. Tool credentials are encrypted at rest, and shared tools run server-side on the owner's credentials without exposing secrets to other users.
Four deployment postures, one code base. Compare deployment options →
Inside your network
- agent runtime · scheduler (redbeat)
- tool execution · mcp client
- encrypted tool credentials
- run history
managed · on-premises · air-gapped
Cloud
Arc53 operates the backend; agents and tools run in Arc53's cloud. Web tools and MCP servers are reachable by default.
Air-gapped
Same runtime with a local LLM and no external network. Tools that call external services — web search, remote MCP servers, Telegram — are unreachable by design; tools that target systems inside the network (PostgreSQL, API Tool, Remote Device), memory and schedules work.
Example
Government infrastructure, local inference
UK Department for Work and Pensions: an internal knowledge assistant on government infrastructure, with local inference.
Source: DWP
Next: Workflow agents in detail · approval gates and guardrails · API keys, webhooks and channels · the CLI and Remote Device daemon · public sector · deployment options · pricing
Frequently asked questions
Can AI agents run on a schedule?
- Yes. Any published agent can run on a cron schedule or as a one-off at a set time (RedBeat), and agents can create schedules themselves with the scheduler tool using relative or ISO-8601 times. The Schedules UI validates schedules, pauses runaway ones and keeps a history per run.
Does DocsGPT support MCP?
- Yes. The MCP Tool connects remote MCP servers as tools, with five authentication modes: none, Bearer token, API key, Basic and OAuth (v0.15).
Can agents run fully self-hosted?
- Yes. Agents, their tools and their schedules execute in the DocsGPT backend, which runs in your cloud account, your data centre or air-gapped; tool credentials are encrypted at rest. See deployment options.
What is the difference between a Classic and an Agentic agent?
- A Classic agent retrieves chunks from its knowledge before the model runs and answers with them (plus any tools). An Agentic agent lets the model decide when to retrieve, through an internal_search tool, and to call tools across several steps.
Can I export an agent to another instance?
- Yes. Agents export and import as YAML with secrets stripped, and an organisation can ship templates in premade_agents.yaml so every instance starts with the same agents.
See DocsGPT on your documents
A 30-minute demo on your use case, or a free on-premises review with an engineer.