Works with the extensions, keybindings and settings you already have
AI code editor built in Nepal.
Move through your backlog faster. It reads the whole repository, plans the change, writes it, runs your tests, and hands you a diff to approve.
No card required. Set up in about a minute.
- macOS, Apple silicon, macOS 12 or later. Download for macOS
- Linux, x64. Download .deb .rpm | .tar.gz
- Windows, Coming soon. Tell us you want it and we will email you the day it ships.
- Your extensions, keybindings and settings carry over
- Every change arrives as a diff you approve
- A checkpoint before every turn, so a bad run costs nothing
- Your code is never trained on
Describe the change. Read the diff.
Ask in plain language. AstraCode's powerful coding agents find the files, plan the edit, make it and run your tests. What comes back is a change set, not a wall of chat you have to translate into code yourself.
- Works across the repository, not just the open tab
- Runs your test suite before it claims to be done
- Every step names the model that served it
Nothing lands until you say so.
Agent edits arrive as hunks you accept one at a time. Every turn takes a checkpoint first, so any run can be undone completely, including files it deleted and files it created.
- Accept, reject or edit each hunk
- A checkpoint before every turn
- Restore brings deleted files back
One mode that finishes the job.
Other editors hand you a list of models and let you guess, then leave you to notice when the guess was wrong. AstraOne chooses, runs your own test suite before it says it is done, and if the tests go red it brings in a stronger model and keeps working. You read a finished change instead of a half-finished one.
- Nothing to choose, so no wrong choice to make
- Your tests decide when it is done, not the model
- A failed check brings in a stronger model, on its own
It can use the tools your team already runs.
Point it at your Model Context Protocol servers and the agent treats them like any other tool: your issue tracker, your database, your internal services. Servers your organisation has approved arrive already configured, rather than being hunted down and pasted into a JSON file by every engineer separately.
- Any MCP server, local or over HTTP
- Approved servers arrive configured, not hunted for
- Server output is fenced as untrusted before the model reads it
It never becomes training data.
Privacy is enforced at the gateway, not promised by the client. The agent works in your checkout on your machine, and what it sends to a model is used to answer you, not to train anything.
- Never trained on, by us or a provider
- No copy of your repository on anyone else's disk
- Bring your own key, used once and never stored
The agent
It knows every caller
Before it changes a signature it asks the language server who uses it, so it does not break a file it never opened.
Stop choosing models
AstraOne chooses, and a failed check brings in a stronger model. Every step names the model that served it.
It tests in a real browser
It opens your app, clicks through it, reads the console, fixes what broke and checks again.
While it works
It asks before it guesses
An ambiguous request gets questions with concrete options, rather than an afternoon spent building the wrong thing.
Big jobs split up
Subagents and parallel agents coordinate larger work, each with its own context, so the task you asked for stays in focus.
Your dev server keeps running
Long commands run in the background and it reads their output as it goes, instead of waiting on something that never exits.
Your control
Any run undone completely
A checkpoint before every turn. Restoring brings back files it deleted and removes files it created.
It follows your house rules
Project rules and engineering principles ride with every request and live in the repository, so the whole team gets the same agent.
Commands rated by blast radius
Classified by what they can destroy rather than by a keyword list, so rm -rf and ls are not the same question.
Working with it
Infrastructure it will not wreck
Terraform, Kubernetes, Helm and Docker. No apply without a plan you can read, and the approval names everything it would destroy.
Highlight it, attach it
Lines in the editor, output in the terminal, a sentence in the chat: select it and it joins your next message, no pasting.
It says what is left for you
Every run ends with what changed, what it checked after its last edit, and the steps that are yours, each with its command.
Your setup
It remembers what you told it
One durable fact at a time, saved only when you agree, so the next session starts knowing what this one learned.
It runs on your machine
Your checkout, your hardware, your tools. No copy of your repository on anyone else's disk.
Bring your own key
Point it at your own provider account. The key is used for the request and never stored.
Pricing
Hobby, Rs 0/month
The whole editor, free. Enough headroom to finish something real.
- Free forever, no card
- AstraOne picks the model for each task
- Unlimited completions as you type
Pro, Rs 1,999/month
For one developer using it all day.
- 7x the headroom of Free
- AstraOne, escalating on a failed check
- Any model, with your own API key
Teams Standard, Rs 2,999/month per user
Shared billing, and an allowance per seat.
- 11x the headroom of Free, per seat
- Org-wide usage and spend
- Shared project rules
Prices exclude 13% VAT, which is added at checkout.
Questions
Will my extensions and settings work?
Yes. The first time you open AstraCode it brings the extensions, keybindings and settings from the editor you use today across, and tells you what it took. It only ever fills in blanks, so anything you have already chosen here is left alone, and you can skip it or run it again later from Settings. Editor updates reach you on the normal schedule too, so you are never waiting on us to catch up.
Is it AstraCode or Astra Code?
One word: AstraCode. Enough people write Astra Code with a space that search engines treat the two as separate things, so to be clear, both mean this, and the site is astracode.io. It is also not Astrocade, which is a games platform, and not the several unrelated products that share the name.
How is AstraCode different from Cursor and GitHub Copilot?
All three put an AI agent on your code. AstraCode is an AI-native coding environment whose agents plan, implement, test and review their own work: they run your tests, prove them by undoing the change underneath, and hand you a diff. AstraOne chooses the model for every task and brings in a stronger one when a check fails, so there is no model list to manage. And the same agent runs in your terminal, your CI, your pull requests and inside JetBrains IDEs and Zed.
Can AstraCode take on bigger, autonomous tasks?
Yes. Plan mode agrees the approach and can hand each step to an agent of its own. Several agents run in parallel, each in its own git worktree. /goal keeps working in rounds until a goal is met on evidence, /loop re-runs a prompt on a schedule, and the CLI automates the same agent from scripts and CI.
Which AI model does AstraCode use?
AstraOne chooses it for you. It runs the model measured to finish real repository work, and when a check fails it brings in a stronger model and keeps going. Every step names the model that served it. If you bring your own provider key, you can pick any model yourself.
What happens when I run out?
Requests stop and nothing is charged. Going past your allowance takes two deliberate acts: switching overage on, and having a card on file. Until you do both, an agent left running overnight cannot cost you a penny.
Can I cancel whenever I want?
Yes, from your dashboard, in one click, with no call and no retention script. You keep the plan you paid for until the end of the period you already paid, and then it moves to the free tier by itself.
Can I use my own API key?
Yes. It is used for the request and never stored, and those requests go straight to your own provider account. Usage is still logged so you can see where it went.
Does my code get uploaded anywhere?
No. The agent runs in your checkout, on your machine, using the tools you already have. What leaves is the part of a file it needs to answer the request you made, and nothing is copied to a build environment somewhere to be worked on. The tradeoff is that a task runs while your laptop is on; if you need work to continue after you close the lid, talk to us.
Do you train on my code?
No, and neither do the providers we route to. That is enforced at the gateway rather than promised by the client, because a client that promises its own privacy is promising nothing.
Point it at something that actually matters.
Free to start, no card. AstraOne picks the model for each task.