For engineering teams on their own infrastructure
Self-hosted AI coding agents for your team
RemoteHangar installs on your Linux server. Each agent works in its own sandbox, takes tickets from Jira, YouTrack or Notion, and hands changes back for review. Repositories, session history and keys stay on your hardware.
Coding agents work for one developer. A team needs more.
Agents already write code, run tests and open merge requests. The trouble starts when five or fifty people use them.
-
An agent on every laptop
Each developer sets up skills, plugins and MCP servers alone. Nobody sees what the agents are doing, and a good setup stays on one machine.
-
“Self-hosted” that is only half yours
Vendor offers move code execution to your machines, but the agent loop and the session transcripts stay in the vendor’s cloud.
-
Agents run with personal keys
An agent on a laptop acts with the developer’s git tokens and cloud credentials, often with nothing between the agent and the rest of the machine.
From install to reviewed changes in three steps
-
Install on your server
One install.sh sets up a systemd service with an SQLite database on Ubuntu 22.04. During the pilot we install it together with you.
-
Connect trackers and git
Link YouTrack, Jira or Notion and your GitHub, GitLab or Gitea. Each person adds their own SSH key or access token.
-
Agents take tasks, people review
Queue tickets for an agent’s bay, its own workspace. The agent works through them one by one, asks for approval where your rules require it, and hands the changes back.
What your team gets
-
A sandbox per agent
Every agent process runs under bubblewrap with its own working directory and HOME, no sudo and a hardened systemd unit.
-
A task queue per bay
Queue typed prompts or tickets from YouTrack, Jira or Notion. The agent marks each one done or blocked; a blocked task pauses the queue.
-
Approvals from your phone
Approve or deny tool calls in the browser or the installable web app. Web Push and Telegram tell you when an agent asks a question or finishes a turn.
-
One place for skills, plugins and MCP
Admins install plugins, add marketplaces, configure MCP servers and curate bundles. The whole team works from the same setup.
-
A Public API for pipelines
Call agents from CI: streaming responses, JSON-schema output, idempotency keys, per-API-token limits and one-shot throwaway bays.
-
Agent tools on your own hosts
A small Go connector on your staging or production host links it to RemoteHangar over mTLS, so agents can run file and shell tools there.
A task queue per bay
A sandbox per agent
Approvals from your phone
Agent tools on your own hosts
What stays on your server, what leaves it
RemoteHangar keeps the control plane on your hardware. The model still runs at your provider, and we say so up front.
Stays on your server
- Repositories and working copies
- Session transcripts and task history
- Git and tracker credentials, encrypted at rest
- Skills, plugins, MCP configuration and bundles
- The audit log of tool-permission decisions
- Application logs and metrics
Leaves your server
- Prompts and code context, sent to the model provider you configure, optionally through your HTTPS proxy
- Calls to cloud trackers and git hosts you connect, such as Jira Cloud, Notion or GitHub
- Calls made by MCP servers you enable that talk to outside services
- Downloads of signed releases and checksum-pinned components, on the schedule you set
We have no remote access to your installation. Updates are signed releases that your server pulls inside your maintenance window.
Runtimes and models
RemoteHangar is built so your team is not locked to one agent vendor; today it runs one runtime. Model calls go to the provider account you configure, billed to you directly.
- Claude Code (Anthropic)
- Status: Supported
- OpenAI Codex
- Status: Planned
Ready on day one
A fresh install arrives with a curated agent stack that admins can trim or extend from the web console.
- 22 curated bundles
- 58 built-in skills
- 19 plugins
- 7 marketplaces
- Trackers
-
- YouTrack
- Jira
- Notion
- Git hosts
-
- GitHub
- GitLab
- Gitea
Built to be run by a small team
Self-hosting means you run it. We made that part boring on purpose.
-
One installer, one service
install.sh sets up a systemd service with SQLite. No Kubernetes, no external database, Docker optional.
-
Signed releases
Every release manifest is signed with minisign, and the signature and checksum are checked before a release is installed. If the host cannot run the verifier, the installer only warns;
--require-signaturemakes the check mandatory. -
Updates when agents are idle
Updates run only inside your maintenance window and only when no agent is busy. You can pin a version or turn updates off.
-
Automatic rollback
The database is snapshotted before every update. If the health check fails, code and schema roll back on their own.
Questions teams ask
What leaves our server?
Prompts and the code context the agent sends to the model provider you configure, optionally through your corporate HTTPS proxy. Also calls to cloud services you connect yourself: Jira Cloud, Notion, GitHub and similar, and MCP servers that talk to outside services. Your server downloads signed releases on your schedule.
Repositories, session transcripts, task history, credentials and logs stay on your server. We have no remote access to it.
Which agent runtimes and models does it support?
Claude Code (Anthropic): supported. OpenAI Codex: planned.
Model calls go to the provider account you configure, with your own API key or accounts. Admins choose which models the team may use.
What does the server need?
Ubuntu 22.04 on x86_64 with systemd and root access for the installer. Plan memory at about 350 MB per concurrent agent turn, on top of the base service; the default 3 GB limit fits four concurrent turns. A release is about 250 MB to download and more once unpacked, and earlier versions are kept for rollback. Optional components such as code search and browsers for UI tests add several hundred MB.
How much does it cost?
There is no public price yet. We agree pilot and license terms on the first call, before anything is installed. Model usage is billed by your provider directly; we add no markup.
How is this different from an AI assistant in the IDE?
An IDE assistant runs on one developer’s laptop with that developer’s credentials and setup. RemoteHangar runs agents on a shared server your team controls: each agent in its own sandbox, with a task queue, shared skills and plugins, approvals from any device and one audit log of tool-permission decisions. Developers work with it from the browser or phone.
How do updates and rollback work?
A systemd timer checks for a new release inside your maintenance window and installs it only when no agent is busy. The release manifest is signed with minisign and verified first; if the host cannot run the verifier, the installer only warns unless --require-signature is set. The database is snapshotted before every update; if the health check fails, code and schema roll back automatically.
If a release turns out to be bad, we point the release channel back to the previous version, so servers that have not updated yet never take it.
How long does the pilot last?
About three months. We agree the start date, goals and a weekly call slot up front. In return we ask for active use, regular feedback, a weekly call and a public case study after launch.
What happens to our data if we stop?
It stays on your disk. The database is a single SQLite file, workspaces are ordinary directories with git repositories, and session transcripts are plain files on the same disk. The built-in backup command writes a database snapshot and an archive of the data directory. Nothing needs to be exported from us, because we never held it.
Who can use it, and how do people sign in?
Admins create accounts in the web console, or you turn on OIDC single sign-on and people sign in with your identity provider. There are two roles: admin and user. Passwords are hashed with Argon2id.
Do we have to install it ourselves?
No. During the pilot we install RemoteHangar together with you on a server you provide, connect your tracker and git host, and set up the first bays.
See it on your own server
Book a 30-minute demo or apply for the pilot. We answer every application personally, by email.