Work

Available

Build engines and agents

Build engines execute working-tree changes; agents are contextual collaborators inside Iterel surfaces. They are different concepts.

Open beta guideJump to first section

Build engines

Claude Code, Codex, Cursor, GitHub Copilot, and OpenCode have first-class setup rows. Aider and Mistral Vibe remain available through the additional-engine list. Each runs on the Mac using its own supported CLI authentication. Iterel records the actual engine, model, version, transport, and effective run controls while owning worktree isolation, context assembly, Gates, browser verification, and receipts.

Claude Code and Codex expose rich session controls through their verified transports, including steer, resume, plan/review strategies, approvals, or user-input cards when supported by the installed version. Cursor supports durable resume but not live steer. Every control is capability-discovered or conservatively hidden; Iterel never presents an unsupported switch.

Specialist execution

Research, Strategy, Creative Direction, and authored agents default to a connected engine using the user's existing vendor plan. These review-first runs are read-only against bounded context and save cited Chats, documents, Whiteboard results, or structured findings. Managed Claude and GPT are explicit Iterel Cloud choices that use credits; Iterel never switches to them silently.

Agents

An agent is a named, reusable, version-pinned specialist. Use one once with a picker-authorized @handle, retain it in an Agent Chat, or assign it to a Task. The Agents panel manages Iterel, personal, project-shared, and verified engine-native agents; General remains the invisible absence of an agent. Code Quality, Security, UX Audit, UI Audit, Accessibility, and Release Check are optional certified starters rather than permanent front-door modes. Historical agent:* conversations stay readable and continuable.

One task memory

Switching Build engines does not require re-prompting. Iterel freezes prior rounds, conversation, receipts, changed files, remaining plan items, task/mission references, effective rules, recalled memory, design guidance, and workspace state in an engine-neutral Context Pack before the next run.