AICode

AICode is a spec-driven, offline-first AI coding assistant, provided as a VS Code extension. Its core idea is to move human control from the post-code-generation review stage forward to the spec design stage: before generating any code, design, revise, and lock the change spec in natural language; after code generation, the AI audits and corrects the implementation according to that spec, thereby replacing manual review of generated code. Officially, it is said to result in industrial-grade code quality, with speed up to 10x to 200x.

Main capabilities

• Spec-driven workflow: follows the process of "ideation -> spec -> refinement -> coding -> verification," with control occurring at the stage that is 100x cheaper.
• Five-dimensional codebase index: simultaneously provides five views: project map, lexical, vector, symbol, and Git history, enabling the agent to deeply understand code even in poorly documented codebases.
• Enforced quality assurance: built-in out-of-the-box quality assurance mechanisms covering sound architecture, no silent regressions, constraint memory, reviewable patches, and tests that can truly catch bad code.
• Designed for large legacy codebases: designed for large codebases where a single uncontrolled change could cause regressions in dozens of modules.
• Data sovereignty: no AICode servers, no database; local indexing, use your own API key, direct connection to a private Azure tenant, and code does not pass through its servers.
• Strict execution model: the AI only reads and proposes; it cannot write files, commit, or run commands; every change lands as a patch and is accepted file by file by a human.
• EasyAgents: can turn a spreadsheet into a controlled cluster of autonomous AI coding sessions, scaling execution of tests, documentation, migrations, and transpilation on your own infrastructure.

How to use AICode is a standard VS Code extension, offline-first, with an initial installation taking about 2-3 minutes, no account registration required, no cloud sync, and code remains local. Installation steps:

1. Open VS Code, go to Extensions, search for aicode-ja, and install it.
2. Click the AICode icon in the sidebar, open Settings to configure the LLM provider.
3. Enter an OpenAI or Azure key; the key is stored locally and will not be displayed again afterward.
4. Wait for project indexing to complete, about 20 minutes per 100,000 lines of code. Click "+ New Chat" to describe the task, and AICode handles the rest. Officially labeled as 4 steps, 2-3 minBell, no account required.

Use cases

• Maintenance and refactoring of large legacy codebases, reducing the understanding gaps caused by staff turnover.
• Delivery processes that require strict control of change quality and avoidance of regressions and technical debt.
• Enterprises with data sovereignty requirements that want code to stay local or in a private cloud.
• Teams that need to execute testing, documentation, migration, and transpilation at scale (via EasyAgents).

Limitations and pricing

• Strict execution means AI cannot directly write files, commit, or run commands; all changes must be accepted manually file by file.
• Initial indexing takes a long time (about 20 minutes per 100,000 lines of code).
• Pricing: 30-day free evaluation; after that, self-service licensing is billed monthly at $10 per user per month (tax included), with no commitment period, and can be canceled at any time with immediate effect. Individual professionals and organizations with fewer than 10 employees or registered volunteers can use it for free; schools, universities, and nonprofit organizations can also use it for free. The Business plan provides full product access and ticket support for teams of any size, can be purchased online via Stripe, and the number of licenses can be adjusted at any time. There are also AI coding training and project hands-on support services, which include everything in Business and train teams to use AI coding effectively, collaborating with developers in real projects and codebases.

The official real-world case: in a 717,000-line codebase, 189 files were changed within 3 days, 201 errors were caught, 31 rounds of verification were run, and the final delivery was code with zero defects, zero regressions, and zero technical debt.