One loop from docs to deploy, on servers you own.

Documentation, tickets, CI runners, and deployment to your own machines in one self-hosted platform, with an MCP server over the whole domain.

Read the docs

git clone https://github.com/otal-labs/nexul.git && cd nexul && ./install.sh

The topology canvas: a hostname pill, a gateway row, and services inside one docker network

Docs as source of truth

Tickets come from documents, and deploys link back to both.

Documents are versioned and are where work starts: a doc spawns scoped tickets, each ticket links to its branch, pull request, and deploy, and all of it is indexed so the next piece of work starts with the full trail instead of a blank page.

docpayments-spec, version 4

ticketPAY-12, derived from payments-spec

pr#88, linked to PAY-12

deployapi on vps-1, linked to #88

MCP server

Every use case is an MCP tool, because the web app calls the same layer.

The domain layer is the single source of behavior. The HTTP gateway for the web app and the MCP server are two thin adapters over it, so an LLM agent gets doc search, ticket creation, topology, deploys, run logs, and dead-letter replay by construction. There is no shadow API to keep in sync.

docs

doc_createdoc_getdoc_searchdoc_update

tickets

ticket_createticket_update_statusticket_link_prticket_search

deploy

stack_createstack_deploystack_rollbackdeploy_list_by_status

topology

topology_gettopology_add_nodetopology_add_edge

runners

runner_listrunner_queuemachine_discover

events

list_dead_lettersreplay_dead_letter

Ordinary servers

SSH, Docker, and a runner binary. No Kubernetes.

The server stack is one docker compose file. Runners are Go host binaries that connect out to the server over a WebSocket, mount the Docker socket, and build and deploy on the machine they run on: a VPS, bare metal, or the same host as the server. SQLite is the only database, and it lives on the data volume.

serverAPI, web app, and MCP endpoint; SQLite with WAL and FTS5, no external database

runnerHost agent that builds and deploys, connected to the server over a WebSocket

mcpstdioProxies stdin/stdout MCP JSON-RPC to a running server for local agent clients

How it fits together

docs→tickets→code + pull requests→review→deploy→index everything→next loop

Install

install.sh

$ ./install.sh
Password for the logs UI (user nexul@nexul.local), blank to generate one:
Nexul is starting.
  Setup wizard:  http://10.0.0.12/
  Logs UI:       http://10.0.0.12:5080  (user nexul@nexul.local)
  Logs password: mUq3zYx8ZbWnP1kQ  (generated; also in .env)
Put a tunnel or a TLS proxy in front for a public domain, then enter that https:// address as the instance URL in the wizard.

The script installs Docker if it is missing, asks for a logs password, starts the stack with one runner on the same host, and prints the setup wizard URL. Installation guide

Open core, one repository, two licenses. Everything outside ee/ is AGPL-3.0-or-later: run it, modify it, and self-host it at any company size.

Everything inside ee/ is under a commercial license; the code is public to read and patch, running it in production needs a subscription. Licensing