AICA — API Integration & Code Intelligence

Takes an API, an intent, and the codebase you already have, and works out what to change — from evidence rather than from a guess.

Most of what it does needs no model at all.

What it does

Reads your codebase properly. Builds a symbol and reference index with the TypeScript compiler, then a dependency graph on top of it. "What depends on this?" is answered from the AST, and the answer says what it could not see — a reference reached through a property access needs a type to resolve, and a report that quietly omitted those would be confidently wrong.

Imports an API and tells you if you already call it. OpenAPI, Swagger, Postman collections, a pasted cURL command, or a Postman workspace. Every source lands in the same intermediate representation, and each endpoint is matched against the URLs actually present in your code — including when the base path lives in a BASE_URL constant and the specification's path carries it instead. "Three call sites" and "not called yet" are different tasks, and the catalog knows which one you have.

Plans deterministically before a model is involved. The plan names files, quotes signatures the index actually holds, states which files not to touch, and lists what it could not determine. A plan the evidence does not support arrives as a question rather than as work.

Proposes changes; you decide. Nothing is written until you say so. The diff you review is against the file as it is now, not a snapshot from when the run started. A revert restores content captured at apply time rather than inverting a diff.

Validates what it wrote, and repairs within a budget. Typecheck, lint, tests, build. A check with no configured command is reported as skipped, never folded into "passed" — and a run that changed nothing reports "no changes" rather than a tick.

Connects to MCP servers, with per-tool risk classification. A server's claim that its own tool is read-only buys one step down, never a free pass; only you can mark a specific tool trusted.

Getting started

1. Open a folder. The extension starts a local agent server for it.
2. AICA: Index Workspace — this is what everything else reads.
3. AICA: Import API Specification — a file, the clipboard, or Postman.
4. Open the chat and describe what you want.

Steps 1–3 need no API key. Step 4 needs a model.

Configuration

An optional agent.config.json in your project root. Every default is the cautious one: the agent proposes rather than writes, no command runs that is not allowlisted, and nothing is sent anywhere you have not configured.

{
  "model": { "provider": "openrouter", "model": "anthropic/claude-sonnet-4.5" },
  "permissions": { "approvalMode": "askAlways" },
  "validation": { "typecheck": ["pnpm", "typecheck"], "test": ["pnpm", "test"] },
  "limits": { "maxFilesChanged": 25 }
}

A key never appears in configuration. It holds a reference — env:OPENROUTER_API_KEY, keychain:postman — and a literal credential there is a validation failure rather than a leak. Set the value with an environment variable, or for Postman use AICA: Set Postman API Key, which puts it in the OS keychain through VS Code's SecretStorage.

Set "model": { "provider": "scripted" } for a dry run: the plan is real, the writing is not, and nothing reaches a network.

What it will not do

Worth knowing before you install it.

• It will not write without permission in the default modes. It is not given a tool that writes.
• It will not disable a test to make validation pass. That is stated in its instructions and the validation layer would report the change regardless.
• It will not put a credential in code. Authentication is described as a scheme and a reference; a value never enters the model's context.
• It will not read outside your project. Every path is resolved, symlinks included, and checked for containment.
• It will not claim a check passed that did not run.

Privacy

The agent server is a local process. Your code is read locally and indexed locally. What leaves the machine is what you configure: prompts to your chosen model provider, and requests to APIs you point it at. Both are recorded in an egress ledger — destinations and volumes, never payloads — and "privacy": { "localOnly": true } refuses egress entirely.

Optional history is stored in Postgres, off by default, and metadata only: file paths, symbol names, signatures, and the graph between them. Source text, doc comments, prompts, and model output are never stored, because the schema has no column for them.

Requirements

• VS Code 1.96 or later
• Node.js 22 or later
• A TypeScript or JavaScript project. Other languages are not yet indexed.

Known limits

• Indexing covers TypeScript and JavaScript only.
• Symbol resolution is syntactic. A reference through a property access is counted but not attributed, and impact reports say so rather than guessing.
• The bundled agent server is large, because it carries the TypeScript compiler.

Licence

MIT.