Configuration

AccessAgent is configured via environment variables and optional adapter plugins.

Environment variables

Required

Variable	Description
DATABASE_URL	PostgreSQL connection string. Example: postgresql://user:pass@host:5432/panelai
OPENAI_API_KEY	OpenAI API key. Get one from platform.openai.com.
AUTH_SECRET	Secret used to sign session tokens. Use a long random string.

Optional

Variable	Default	Description
PORT	3000	Port the HTTP server listens on
PROJECTS_DIR	./projects	Where per-project widget files are stored
MAX_PROJECTS_PER_TEAM	unlimited	Limit projects per team (set a number to cap)
AGENT_MODEL	gpt-4o	OpenAI model to use for the agent
AGENT_MAX_TOKENS	4096	Max tokens per agent response

Auth adapter

By default, AccessAgent uses local username/password authentication. Users register with an email and password stored (hashed) in your database.

To replace auth with a different provider, implement the AuthAdapter interface and pass it to the server config:

src/adapters/auth/my-auth.ts

import type { AuthAdapter } from '../interface';

export const myAuthAdapter: AuthAdapter = {
  async createUser(email, password) { ... },
  async verifyUser(email, password) { ... },
  async getUserById(id) { ... },
};

Note

The SaaS layer implements a Google OAuth adapter using this interface. If you’re building a custom deployment and want Google OAuth, the SaaS adapter is a good reference.

Storage adapter

By default, widget files are stored on the local filesystem at ./projects/{projectId}/. For multi-server deployments or cloud storage, implement the StorageAdapter interface:

src/adapters/storage/my-storage.ts

import type { StorageAdapter } from '../interface';

export const myStorageAdapter: StorageAdapter = {
  async readFile(projectId, path) { ... },
  async writeFile(projectId, path, content) { ... },
  async deleteFile(projectId, path) { ... },
  async listFiles(projectId) { ... },
};

Agent model

The agent uses gpt-4o by default. To use a different model, set AGENT_MODEL:

AGENT_MODEL=gpt-4o-mini

Any model accessible via your OpenAI API key will work. gpt-4o produces the best results for code generation. gpt-4o-mini is faster and cheaper but less capable at complex widgets.