Skip to main content

FAQ

Why does AIFactory exist when Cursor / Devin / Aider exist?​

Different category. Cursor is a code editor with AI. Devin is a hosted agent service. Aider is a CLI pair-programmer. AIFactory is a task management + agent orchestration platform — it sits one level above the agent and treats it like a build runner.

Concretely, AIFactory is what you reach for when:

  • You have a backlog of GitHub issues and you want to triage them into agent-shippable units without writing prompts
  • You want spec-first development: a reviewable artifact before code is written
  • You want to mix providers by phase (Claude planning, cheap local coding)
  • You need audit trail + SOC2 evidence for AI-assisted work

Does it work offline?​

The portal does. The agent only works offline if you point it at a local Ollama instance — see Multi-Provider.

How does authentication work?​

Web-server auth: a token (rotating, printed to console at boot) that the frontend stores. Optionally OIDC SSO (Keycloak / Okta / Azure AD) via the OIDC_ISSUER_URL env var.

Claude auth: the OAuth token from claude setup-token. Never the direct API key.

Can I run multiple tasks in parallel?​

Yes. Each task runs in its own subprocess and its own worktree. Resource limits are per task; no shared mutable state.

What if the agent gets stuck?​

The runtime detects stuck subtasks (no log output for N minutes) and invokes the Coder Recovery agent (prompts/coder_recovery.md). After three failed recoveries, the task is paused and surfaced to the UI.

You can also click Stop at any time — that sends SIGTERM to the agent process and stops the rmux session if one exists.

How do I rotate the OAuth token?​

claude setup-token
# Replace CLAUDE_CODE_OAUTH_TOKEN in apps/backend/.env
# Restart the web-server

No tasks need to be cancelled — running tasks already hold a session.

Where does my code go?​

Nowhere it wasn't already. Agent edits land in a git worktree under .aifactory/worktrees/tasks/<spec-id>/. Nothing leaves your machine unless you push.

LLM providers see only the file contents the agent sends them — typically the files relevant to the task (chosen via the codebase_analyzer sub-agent) plus the spec.

How does the CI fixture seeding work?​

The E2E tests use AIFACTORY_TEST_AGENT_CMD=sleep 300 to replace the agent subprocess with a no-op while the rmux integration hook still creates the live session. This lets the Playwright suite exercise the Live Console without burning LLM tokens. See Troubleshooting for details.