An assistant that runs your apps with you. A platform agents can drive.
The assistant is in beta and is rolling out. Ask what broke and get an answer from the real logs, metrics and builds, with the fix ready to confirm. Hand the CLI to an agent and it gets JSON, stable exit codes and errors that say what to do next.
Ask about your apps.
Confirm what changes.
Press Ask or ⌘K on any dashboard page, or run deplane ask in a project folder. The assistant reads the same data you see, explains it in plain words, and proposes an action when one would help. Nothing changes until you press Confirm.
The build of v9 failed in npm ci: package.json asks for react@19.4.0, which does not exist. v8 is still serving, so nothing is down.
Change it to ^19.2.0 and push. If you want to retry the same commit first, I can redeploy it.
my-shop runs 2 small instances. At yesterday's peak, 412 requests a minute, CPU reached 81% and p95 was 640 ms. A sale doubling that would likely slow checkout.
Going to 4 medium instances fits your plan and adds no downtime.
I found acme/notes in the repositories you connected: a FastAPI project on main.
Here is the plan: create the app, connect the repository, create a small Postgres database and attach it as DATABASE_URL, then deploy.
The real state of your team, not a guess.
- Logs from every instance, and build logs line by line.
- Metrics: requests, errors, p95 latency, CPU and memory.
- Releases and builds: what changed, when, and whether it worked.
- Domains, workers, cron jobs, databases and alerts.
- Your connected GitHub repositories, to set up new apps from them.
It never reads the values of your variables or secrets. It sees their names only.
Sixteen actions, each one a Confirm away.
- Releases: scale (with autoscaling and sleep when idle), roll back, redeploy, deploy the latest commit.
- Apps: create an app from a GitHub repository.
- Processes: add a worker or a cron job, scale a worker.
- Domains: add, re-check or remove a custom domain.
- Databases: create, attach or detach Postgres; restore a backup into a new database and switch apps to it.
- Variables: set or unset them. It names the variable; you type the value.
Confirm, always
Every change is a card with exactly what will happen. Nothing runs from the assistant's answer alone.
No deletes
The assistant never deletes an app, a database or a domain's history. Those stay with the Delete button and you.
Your permissions
It acts as you, with your team's limits. It cannot do anything you could not do yourself.
On the record
Every confirmed action is in the activity log, marked as done via the assistant, with who confirmed it.
Deploy from your AI tool.
Connect Claude, ChatGPT or Cursor to Deplane with one URL. Then ask it to build and host an app, and it creates, deploys and runs it for you, in your team.
It never deletes anything, and it never creates tokens or manages your team. Deplane refuses those on the server, whatever the tool asks. You can disconnect a tool at any time, and every change it makes is in your activity log, marked with its name.
- Copy the server URL from the Tokens page of your dashboard:
https://api.deplane.cloud/mcp. - Add it to your tool. In Claude: Settings, Connectors, Add custom connector. In ChatGPT: Settings, Connectors, Create. In Cursor and VS Code: add a remote MCP server.
- Sign in and press Allow. A browser window opens, you confirm with your passkey, and you are connected.
The CLI an agent can actually use.
Coding agents already run commands. Deplane makes every command's result something a program can read and act on, without scraping text.
--jsonon every command. One JSON document on stdout; progress and build logs on stderr.- Stable exit codes that say what kind of failure happened.
- Every error carries a fix: the next step, usually as the exact command.
- CI tokens scoped to one team, 90 days by default, revocable at once.
- Docs agents can read at
/llms.txt, with a Markdown twin of every page. - Never prompted. Without a terminal the CLI never waits for input; it fails with the flag to pass.
One number tells an agent what to do next.
| Code | Meaning | What an agent should do |
|---|---|---|
| 0 | Success. | Continue. |
| 1 | User error. | Read fix, change the command or the project, and try again. |
| 2 | Authentication error. | Check DEPLANE_TOKEN, or ask a person to log in or confirm. |
| 3 | The build failed. | Read the build log and fix, change the code, and deploy again. |
| 4 | The deploy failed. | Run deplane logs --json and deplane status --json, fix the app, deploy again. The previous release keeps serving. |
| 5 | Platform error. | Wait and retry. If it persists, report it. |
| 6 | Quota exceeded. | A plan limit was reached. Free resources or ask a person to raise the limit. |
AI-written code needs a stronger wall.
More code reaches production without anyone reading every line, and agents install packages nobody chose by hand. Most of the time that is fine. The question is what happens when it is not.
- Containers share one kernel. On most platforms, a single kernel bug lets code in one container reach the others on the same server.
- Deplane gives every app, build and database its own microVM, with its own kernel. Getting out means breaking a virtual machine, which is much harder.
- Network rules sit outside the VM, on the host. Code inside an app cannot reach our servers, cloud metadata or other teams, whatever it does.
- Honestly: no isolation is perfect. That is why we also watch the host side of every sandbox and run 185 isolation checks after every upgrade.
What comes next.
These are planned, not shipped. The dates are our targets.
A platform for agents
- Agent sandboxes. microVMs that start in about a second, for agents that need to run code they just wrote.
- AI gateway. One key for Claude, OpenAI and others, with usage and spend limits per app.
- Self-healing deploys. When a deploy fails, the assistant opens a pull request with the fix.
Deplane AI
- Open models per token. Llama, Qwen, Mistral and DeepSeek on our own GPU servers, priced per million tokens.
- An OpenAI-compatible API. Point existing code at Deplane by changing one line.
- Private AI for Enterprise. Reserved GPUs and fine-tuned models, with data that never leaves Deplane.
Ask your apps a question.
The assistant is rolling out in the beta. It works once we switch it on for the platform, so it may not answer yet in your account.