AstraCode is an AI-native code editor, positioned as an AI-native code editor. At its core, it lets agents plan, build, test, and review their own work across the full codebase, while AstraOne automatically selects the model for each task, so users do not need to choose manually.
Purpose
- Describe the desired changes in natural language, and the agent finds the relevant files, plans edits, implements the changes, and runs tests, ultimately delivering a changeset (diff) rather than a long chat log that users must translate into code themselves.
- The agent works across the entire repository, not just the currently open tab.
Key capabilities
- Automated testing and completion determination: before claiming completion, the agent runs the user's own test suite; if checks fail (tests go red), it automatically brings in a stronger model to continue working. Each step in the execution process notes the name of the model used.
- Change review and rollback: the agent's edits are presented as hunks, and users can accept, reject, or edit them individually. A checkpoint is created before each round begins, and any run can be fully undone, including files it deleted or created.
- Tool integration: it can connect to the user's Model Context Protocol (MCP) servers, treating issue trackers, databases, internal services, and the like as ordinary tools. Organization-approved servers can be preconfigured directly. Server output is marked as untrusted content before the model reads it.
- Security and privacy: privacy is enforced at the gateway layer. The agent runs in the user's local checkout directory, and content sent to the model is used only to answer the request and is not used for training; no copy of the repository is copied to someone else's disk. It supports using your own API key, with requests going directly to your own provider account, and the key is not stored after use.
- Infrastructure operations: supports Terraform, Kubernetes, Helm, and Docker. It will not execute apply without a readable plan, and approvals list everything that will be destroyed.
- Command classification: commands are classified by the degree of damage they may cause, rather than by a keyword list, so rm -rf and ls are not treated the same.
- Other capabilities: testing in a real browser (opening the app, clicking, reading the console, fixing, and rechecking); asking concrete questions with specific options for ambiguous requests; splitting large tasks among subagents or parallel agents; supporting running long commands in the background and reading the output; project rules and engineering principles taking effect with each request and being stored in the repository; in fixBefore changing a signature, query callers through the language server; supports adding selected lines in the editor, terminal output, or sentences in chat directly to the next message; at the end of each run it reports the changes made, the checks after the last edit, and the steps and commands the user needs to perform; gradually remembers persistent facts told by the user, saving them only with the user's consent.
How to use
- Download and install on macOS (Apple silicon, macOS 12 or later), Linux (x64, .deb / .rpm / .tar.gz), and Windows (x64, Windows 10 or later).
- On first launch, it migrates extensions, shortcuts, and settings from the editor currently in use and informs you what was migrated; it only fills in gaps and does not overwrite existing choices, and it can also be skipped or rerun later. Editor updates are pushed at the normal cadence.
- The agent runs in the local environment and uses the tools the user already has; if a task needs to continue running after the laptop is closed, that must be discussed separately.
- The same agent can also run in the terminal, CI, pull requests, and within JetBrains IDEs and Zed, and the CLI can be operated from scripts and CI automation. Plan mode is supported to agree on an approach and assign steps, /goal keeps working in rounds until the goal is achieved based on evidence, and /loop reruns prompts on a schedule.
Use cases and limitations
- Suitable for development scenarios requiring AI-assisted coding across repositories, automated testing, change review and rollback, and integration with internal toolchains.
- Limitations: tasks run on the local machine, and tasks cannot continue after the laptop lid is closed (the official statement says you can contact them to discuss if needed); when using your own API key, requests connect directly to your own provider account, but usage is still recorded.
Pricing
- Hobby plan: $0/month, includes the full editor, free forever, no credit card required. AstraOne selects the model for each task, provides unlimited completions on input, is aimed at one developer using it all day, has 7 times the quota of Free, AstraOne upgrades the model when checks fail, and you can use your own API key to choose any model.
- Teams Standard plan: $30/month/user, includes shared billing, per-seat quota, 11 times the quota of Free per seat, organization-level usage and spending, and shared project rules.
- After exceeding the quota, requests stop and there is no charge; enabling overage and binding a card require two explicit actions, otherwise no charges will be incurred. You can cancel anytime with one click from the dashboard, and within a paid period you retainthe plan, after which it automatically switches to the free tier.
