Supported agents

Eight coding agents run natively: QualityLayer adds its skills and hooks to each, and the App shows their sessions. Each keeps its own login and provider.

1. Install your agent

   Sign in and start it once from a terminal.

2. Run Setup

   The App's Setup, or qualitylayer agents check, connects every supported agent it finds. If Codex asks you to review hooks, review and allow only QualityLayer's entries before using Conversation.

3. Start a task

   + › Task in the App, or tell your agent "Using QualityLayer, plan a CSV export" (QL works too, in any language).

Agent	Checked with	What QualityLayer adds
Claude Code	2.1.306	Skills, lifecycle hooks and its existing prompt integration
Codex	0.167.1 (macOS)	Skills, implicit invocation policies and lifecycle hooks
Cursor	2026.10.19	Skills and lifecycle hooks in its own hooks file
Grok	1.0.62	Skills and hooks, in its own hooks folder
Antigravity	1.4.12	Skills, a workflow command and lifecycle hooks
Pi	1.2.0	Skills, a lifecycle extension and fresh helper sessions
OpenCode	2.0.29	Skills, a command and a lifecycle plugin
GitHub Copilot	1.0.105	Skills, a hook file and the memory server

Note

Cursor, Grok and Antigravity are set up only after you have started them once from a terminal. GitHub Copilot Business and Enterprise need an organization admin to enable Copilot CLI in the Copilot policies.

The six skills

Ask in plain words; your agent picks the skill. An ordinary coding request does not start a workflow.

ql-workflowPlan, build, verify and review a change. The only skill that starts a task.ql-artifactMockups, diagrams and clickable pages.ql-manageTasks, memory, settings and setup.ql-peerSessions of any of these agents review each other or hand over a task.ql-qualityCode complexity, structure and duplication.ql-chatA Chat without workflow steps.

Models and reviews

Pick the agent, model and effort for each task; the default effort is High.

Agent	Model choices
Codex	The models your Codex account offers, as Codex lists them; a default your plan lacks moves to Codex's own model
Pi / OpenCode	Provider-qualified IDs: openai-codex/gpt-6.1-sol / openai/gpt-6.1-sol; both can use a Copilot subscription
Cursor / Antigravity	Effort is part of the model name, such as Opus 5.5 (High)
GitHub Copilot	Auto works on every plan; paid plans can choose Claude Opus 5.5

Where the agent says which models your account has (Codex, OpenCode, Cursor and Antigravity), model menus show only those; a model already in use that your plan lacks shows as · not offered. Claude Code, Pi, Grok and GitHub Copilot list every model QualityLayer knows.

Reviewers work in a separate context and edit nothing. Cursor and Antigravity review disposable copies and stop at the first tool that is not read-only (choosing reviewers).

Sessions in the App

The Conversation view reads each agent's session record (controls). Some records hold less:

Agent	What its Conversation lacks
Claude Code, Codex, Pi, OpenCode	Nothing
Grok	Context fill, permission questions
Antigravity	Permission questions
Cursor	Thinking, tool output, token counts, context fill, permission questions
GitHub Copilot	Permission questions

Permission waits show in Conversation. A reply that only the Terminal view shows gets Open Terminal. Cursor gets numbered workflow questions; the others ask in their own picker (In your agent).

After a crash, stop the old run and ask "Using QL, continue verification of task retry-webhooks."

Permission modes

New managed sessions start in Bypass unless you saved a global or per-agent mode; Agent settings keeps the agent's own policy. A new task opens with the agent and choices you last planned with, and a build with those you last built with. A build on the planning agent keeps the planning session's choices, including permission changes you accepted.

Agent	Modes
Claude Code	Auto, Ask first, Accept edits, Bypass
Codex	On request, Ask first, Auto review, Full access
Cursor	Ask first, Auto-review, Run everything
Grok	Auto, Ask first, Accept edits, Bypass
Antigravity	Ask first, Auto, Accept edits, Bypass
OpenCode	Agent settings, Auto-approve
Pi	None; its own tool permissions apply
GitHub Copilot	Manual Approval, Assisted Permissions, Allow All; older sessions use the Terminal view's /permissions menu

Warning

Switching a running session to Bypass, Full access or Run everything asks you to confirm. Apply and continue changes permissions at a supported prompt and resumes the conversation. Pi has no bypass switch, and OpenCode keeps explicit denials. Your Plan and final approvals stay separate.

Plan limits

Claude Code, Codex and GitHub Copilot show the usage windows their records report, such as Week · 58 % used; Usage display switches the App and the macOS menu bar to Remaining. Missing or reset readings stay unknown. Agents billed by use show the session's estimated cost instead.

When one of them stops at its limit, the task shows Limit · resets 17:23. With Resume when a limit resets on, QualityLayer continues the session then; Cancel resume stops that. Otherwise copy the command that continues it.

How QualityLayer finds your agents

Not installed means it was not found. Enter the full path in ⋯ › Edit launch… › Command, such as /opt/tools/bin/codex, or press Check again.

Search order

1. The Command you set. If it cannot run, the agent counts as missing.
2. The App's PATH, then your login shell's PATH.
3. ~/.local/bin, ~/.codex/packages/standalone/current/bin, Homebrew, ~/.npm-global/bin, Bun, Volta, pnpm, nvm and fnm.
4. npm's global folder.
5. For Codex, the command inside the ChatGPT or Codex desktop app.

--version must name the agent, such as codex-cli 0.163.0.

Setup and removal

qualitylayer agents list and qualitylayer agents check match Settings › Agents. Pi's files go to ~/.pi/agent, OpenCode's to ~/.config/opencode, GitHub Copilot's to ~/.copilot. Updates replace only unchanged files QualityLayer owns; uninstall keeps your logins, settings, session history and edited files.