Hosts: Cursor, Claude Code, Copilot, Antigravity
Lifecycle rules live in AGENTS.md. wk init and wk export-rules copy one thin stub (templates/host-pointer.md) to each host filename. wk mcp writes MCP config in the file that host actually reads. wk model resolve --host … maps the same capability class to that host’s model slug.
Windsurf is rules-only forever. wk export-rules / wk align keep .windsurfrules. wk mcp does not accept --host windsurf and does not write a Windsurf MCP file. Cascade’s own MCP UI is outside this kit.
What each host gets
| Surface | Cursor | Claude Code | GitHub Copilot / VS Code | Antigravity / Gemini CLI |
|---|---|---|---|---|
| Rules | .cursorrules | CLAUDE.md | .github/copilot-instructions.md | GEMINI.md |
| TDD hooks | .cursor/hooks.json (wk tdd-guard install) | .claude/settings.json hooks | not installed | not installed |
| Project MCP | .cursor/mcp.json | .mcp.json | .vscode/mcp.json (servers) and shared .mcp.json | .agents/mcp_config.json |
| User MCP | ~/.cursor/mcp.json | ~/.claude.json (mcpServers) | ~/.copilot/mcp-config.json | ~/.gemini/config/mcp_config.json |
| Model overlay | models/hosts/cursor.yaml | models/hosts/claude.yaml | models/hosts/copilot.yaml | models/hosts/antigravity.yaml |
External skills (wk sync) | ~/.cursor/skills | symlink ~/.claude/skills | not mirrored (use repo .mcp.json) | symlink ~/.gemini/skills |
Canonical content is still AGENTS.md, skills/, and mcps/. Host files are adapters: the same pointer, not a symlink to AGENTS.md (Cursor loads AGENTS.md plus host files in one session).
wk init . --mcp default --hook
wk align .
wk mcp default --install
wk mcp default --install --host claude
wk mcp default --project
wk mcp restore --project
wk model resolve --skill agent-tdd --spec-complete --host claude
wk model resolve --skill agent-tdd --spec-complete --host copilot
wk model resolve --skill agent-tdd --spec-complete --host antigravity
--host accepts cursor, claude, copilot, antigravity (aliases: gemini, agy), a comma list, or all (default).
What is still host-shaped
Cursor’s in-product skill picker and Task model slugs are the deepest. Claude Code and Antigravity pick up Agent Skills from their user skill dirs after wk sync. Copilot Chat reads .vscode/mcp.json; Copilot’s Agent Host also reads workspace .mcp.json.
Kit-authored skills/agent-* stay in the Waykit clone (~/.agents/skills). Hosts that do not walk that tree still need the handshake: read ~/.agents/AGENTS.md, then load the skill file.
Which roles may become Cursor/Claude subagents (fresh window) vs stay skills: subagent allowlist. Profiles never become agents.
wk agents generate
wk agents install
wk agents install (also wk align . --write and wk sync --install) copies kit stubs into user ~/.cursor/agents and ~/.claude/agents. It does not write the product repo’s .cursor/agents, and it does not invent Copilot or Antigravity agent dirs. Launch contract: subagent launch. Run wk agents status for launch vs skills-only. WK_SUBAGENTS=0 in the host’s start shell keeps specialists as skills in the parent without uninstalling stubs. Print the Task body with wk agents launch-prompt --skill <id>.
OAuth popups differ (Cursor vs Claude vs VS Code). Stdio servers that need env vars still need those vars in the process that launches the host.
Coding-agent token / cost / trace export is host OTLP into community SigNoz Docker (desktop) or SigNoz Cloud ingest (Cloud Agents), not an extra MCP on default. Claude Code and Copilot have native exporters; Cursor Enterprise is metrics/logs only (hooks for traces). Procedure: coding-agent-observability. Profile: wk mcp signoz --install.
Cloud Agent and Automations
A Cloud Agent or Cursor Automation does not read the local ~/.cursor/mcp.json that wk mcp --install writes. Connect GitHub, Linear, PostHog, Cloudflare Observability, and SonarQube on the Cursor dashboard if unattended loops need them. For Cloud Agent traces, also connect SigNoz Cloud hosted MCP and point OTLP at Cloud ingest — localhost SigNoz is unreachable. Missing tools → stop BLOCKED; do not invent counts. Operator CLI: wk loops status / wk loops setup --write. quality-loops. MCP library. coding-agent observability.
EDD vs hosts
The eval runner is wk, not the IDE chat. --style cli --cli cursor-agent|claude|agy can drive a live host binary. That path is part of EDD (alpha), not a complete eval platform.