Open Docy is the self-hosted edition of Docy AI. Compose AI agents from nodes that extract, check, branch, search, write and ask a person. Run them on your own infrastructure and keep the record of what each one decided and why.
# clone and configure $ git clone https://github.com/OpenDocy/opendocy.git $ cd opendocy $ cp settings.ini.sample settings.ini $ cp .env.sample .env # replace change-me # install and start $ uv sync $ uv run docy install-ai $ uv run docy build-web $ uv run docy run ✓ http://127.0.0.1:8003
# pull the latest code $ git pull $ uv sync $ uv run docy build-web $ uv run docy run # migrations run when the server starts
Build an agent from real steps, check every value against a rule, pause for a person when something fails, and keep the record. With sound and captions.
Run as many agents as you need. When a rule fails, the run stops at Action Required and waits for someone with the authority to decide.
● Data extractor ⎿ 14 fields · contract.pdf ● Audit · procurement policy ⎿ 5 of 6 rules passed ● Conditional risk == "high" ⎿ checking…
Each node's output is kept, so you can see how the agent got there.
✕ Rule failed install date in window expected ≥ 2026-07-01 found 2026-06-28 ◆ Interaction ⎿ approve | reject | escalate ⎿ waiting for an assessor
The agent shows the rule, the expected value and what it found.
✓ Assessor approved exception ⎿ reason: letter on file ✓ Content generator ⎿ decision-summary.docx ✓ Record saved ⎿ inputs, rules, answer
Who decided, what they chose and why, stored with the run.
One Python app on your own server, with SQLite as the only runtime dependency. Agents, inputs and results stay where you put them.
Each node's output, every rule's expected and found value, a flowchart and a timestamped log for every run.
The Interaction node pauses a run at Action Required and resumes when someone answers. The answer is part of the record.
The canvas, the task runtime, templates and the audit log, with no hosted dependency.
Rules per field: equals, contains, greater or less than, before, after. Each result shows found, expected and reason, linked to where the value came from.
Ask for text, a choice or a file mid-run. Earlier outputs are shown alongside so the reviewer has the context.
Run one node, read its output or error, change it and run again. Drafts save as you work.
Upload DOCX or PDF, draw the fields, set font and borders. The template node fills them from upstream output.
Repeatable inputs and batch nodes process a folder of documents with one agent, grouped the way you upload them.
Who did what to which agent or task, and the result. Filter by action, resource and date, then export CSV.
Change the nodes and the rules, keep the runtime and the record. These are the kinds of agents Docy AI runs today.
Check installation evidence against scheme rules, route exceptions to an assessor, produce the filing.
Read the application pack, test it against lending policy, escalate anything outside appetite.
Match invoices to purchase orders in batches, hold mismatches for approval, export the result.
Abstract leases and reports, flag missing clauses, write the investment summary.
Check contracts and signatures against policy, look up the supplier, record who approved.
Pull named facts from the web, check them against your criteria, generate the briefing document.
Every task is in exactly one state. Run, rerun or cancel from the task list, and open the flowchart and log for any run.
| Node | You configure | Open Docy |
|---|---|---|
| Input | ||
| input_file | File name, description, allowed formats, size limit (50 MB total) | Available |
| input_file.repeatable | Grouped, repeatable uploads with association management | Available |
| Extract | ||
| data_extractor | Profile scan, area select or manual fields; apply rule; grouped results | Available |
| data_extractor.batch | As above, over batch or loop input | Available |
| extract_signature | Returns signature image, signer, role and date signed | Available |
| Check and route | ||
| audit.global | Custom rules over the whole document; value, expected, result, reason | Available |
| audit.field | Rule and expected value per field; compare with reference data | Available |
| conditional | Conditions on fields; true and false branches | Available |
| interaction | Pause for a person: file, text or select | Available |
| Generate | ||
| document_template | Template, upstream source and field mapping | Available |
| web_search | URL and named query items | Available |
| content_generator | Input fields, prompt and output document name | Available |
| Docy AI | ||
| knowledge | Knowledge base and one or more queries | Visible, locked |
| compliance_policy | Policy and version; include in final export | Visible, locked |
| custom_component | Variables and Python code in an editor; ports generated on save | Visible, locked |
What is in the source. The app, canvas and task runtime are open source. The extraction and AI logic inside each node ships compiled: you can configure, chain and run every available node, but its internals are not in the repository.
Agents use the same nodes in both. Start with Open Docy. Move to Docy AI, in our cloud or on your own premises, when you need teams, an API, the policy and knowledge nodes, or agents drafted by chat.
For developers and teams who want to build and run agents on their own servers.
Everything in Open Docy, plus what a team and an integration need.
Requires Python 3.11 to 3.13 and uv. Node.js 18+ with pnpm builds the web UI.
git clone https://github.com/OpenDocy/opendocy.git cd opendocy cp settings.ini.sample settings.ini cp .env.sample .env uv sync uv run docy install-ai uv run docy build-web uv run docy run
Open 127.0.0.1:8003, create an account and an agent. Add nodes and debug each one as you go.
input_file → data_extractor → audit.field → document_template
Create a task from the agent, add its inputs, press Run. Review results, answer any intervention, download outputs.
task procurement-contract-review
run 0412
state Completed
output decision-summary.docx
Issues, discussions and pull requests are welcome.
Open Docy (also written OpenDocy) is the open-source, self-hosted edition of the Docy AI Agent OS. You compose an AI agent from nodes on a canvas, run it as tasks, and get back its decisions with the evidence: extracted values, rule results, a person's answers and generated outputs. It runs on your own server with Docker Compose.
Yes. Open Docy is free to download and self-host. You pay only for your own server and for the AI model usage your agents make. Docy AI, the commercial edition, is a paid product in the cloud or on-premise.
An Agent OS is the layer that builds, runs and governs AI agents: a builder to compose them, a runtime that executes them step by step, and a record of what each agent did and why. Open Docy provides the builder, the runtime, human-in-the-loop steps and an audit log.
Clone the repository, copy settings.ini.sample to settings.ini and .env.sample to .env and replace the change-me values, then run uv sync, uv run docy install-ai, uv run docy build-web and uv run docy run. The app opens on http://127.0.0.1:8003. It needs Python 3.11 to 3.13 and uv, plus Node.js 18+ with pnpm to build the web UI. SQLite is the only runtime dependency: no Docker, Postgres, Redis or message queue.
No. Documents are the most common input, but agents also search the web, generate content, branch on rules and ask people for input. API input and output belong to Docy AI.
Open Docy covers building and running agents in a personal space. Docy AI, in the cloud or on-premise, adds an assistant that builds agents from a chat, sealed decision records and replay, team spaces and roles, OAuth, knowledge bases, versioned compliance policies, the pre-built agent library and usage controls.
The app, canvas and task runtime are. The AI runtime that does the extraction inside each node is a separate compiled plugin, installed with docy install-ai, so you can configure and run every available node but its internals are not in the repository. Knowledge, Compliance policy and Custom component nodes are visible and locked.
On the server you run Open Docy on: agents, tasks, uploaded documents and results.
Yes. Open Docy has a REST API under /api/v1 for tasks, agents, workspaces, files and webhooks, authenticated with user-scoped API keys. Webhooks push HMAC-signed task events (task_status_changed, node_finished, task_finished) with automatic retry, and interactive API docs are served at /docs. OAuth is part of Docy AI.
In Docy AI. Describe the workflow and the assistant drafts the agent on the canvas for you to check and adjust. In Open Docy you build agents on the canvas yourself.
Both editions use the same node types, so an agent designed in Open Docy carries over to Docy AI, cloud or on-premise, without redesign.
The source opens soon. Leave your email and we'll send you the release, plus a few early-access seats on Docy AI for teams who want to try the full edition.