Agent CLI Integrations

Agent CLI Integrations

Concrete wiring for SpecRoute into each supported agent CLI. Copy commands per vendor.

For the canonical reference, see agentic-docs/agent-cli-integrations.md. For the supported matrix, see Vendor Matrix.

Claude Code

# 1. Drop the runtime layout
cp -R runtimes/.claude/ /path/to/your/repo/.claude/

# 2. Rename templates
cd /path/to/your/repo/.claude
mv settings.template.json              settings.json
mv settings.local.template.json        settings.local.json     # customize per-machine
mv mcp.template.json                   ../.mcp.json            # Claude Code CLI reads .mcp.json at repo root
mv hooks/hooks.template.json           hooks/hooks.json        # authoring artifact only - see step 3

# 3. Merge the hooks into settings.json - the file the CLI actually reads.
#    A bare project-level .claude/hooks/hooks.json is NOT a hook source.
/path/to/specroute/tools/sync-hooks-to-settings.sh \
  /path/to/your/repo/.claude/hooks/hooks.json \
  /path/to/your/repo/.claude/settings.json

# 4. Install hook scripts
cp /path/to/specroute/.claude/hooks/*.sh /path/to/your/repo/.claude/hooks/scripts/
chmod +x /path/to/your/repo/.claude/hooks/scripts/*.sh

# 5. Create the gitignored sanitization wordlist
touch /path/to/your/repo/.claude/.forbidden-strings.txt

# 6. Add to .gitignore
echo ".claude/settings.local.json"        >> /path/to/your/repo/.gitignore
echo ".claude/.forbidden-strings.txt"     >> /path/to/your/repo/.gitignore

# 7. Populate agents/, skills/, commands/ as your project requires.

Claude Code reads hooks from settings.json (user / project / local), managed policy settings, a plugin's hooks/hooks.json, and skill or agent frontmatter - not from a project-level .claude/hooks/hooks.json. SpecRoute keeps hooks.json as the annotated authoring artifact and syncs its hooks key into settings.json. Claude Code has 30 hook events and 5 handler types (command, http, mcp_tool, prompt, agent).

Per-runtime details: runtimes/.claude/README.md.

Codex

cp -R runtimes/.codex/ /path/to/your/repo/.codex/
cd /path/to/your/repo/.codex
mv config.template.toml config.toml

# Mirror skills from your Claude runtime (skills only - agent formats differ)
python3 /path/to/specroute/tools/sync-skills.py --source claude --apply

Codex consumes the same folder-per-skill SKILL.md shape as Claude Code, so skills cross-mirror directly. Agents do not: Codex agents are standalone TOML (.codex/agents/<name>.toml with name, description, developer_instructions) where Claude's are flat Markdown, which is why sync-skills.py syncs skills only. The config.toml carries Codex-specific MCP config and approval policy.

Codex hooks are enabled by default. The canonical feature key is [features] hooks; codex_hooks remains a deprecated alias. Its 11 events reuse Claude Code's names verbatim (SessionStart, SessionEnd, PreToolUse, PermissionRequest, PostToolUse, UserPromptSubmit, Stop, PreCompact, PostCompact, SubagentStart, SubagentStop), but the overlap is close, not total: no PostToolUseFailure, and only type: "command" handlers execute.

.agents/ is the emerging vendor-neutral location. Verified against an installed Codex 0.145.0 binary, not documentation: it carries .agents/skills as a repo-level skills root plus .agents/plugins/marketplace.json. It still carries .codex/skills and $CODEX_HOME/skills too - both work today, so do not migrate off .codex/skills yet.

Per-runtime details: runtimes/.codex/README.md.

Gemini CLI

cp -R runtimes/.gemini/ /path/to/your/repo/.gemini/
cd /path/to/your/repo/.gemini
mv settings.template.json            settings.json
# commands are TOML files under .gemini/commands/ — copy and edit the example(s)

Hooks landed in Gemini CLI v0.26.0; every release since consumes the full stack. Check gemini --version against the upstream changelog rather than trusting a pinned "current stable" number here.

• settings.json mcpServers (JSON object) and the hooks block (11 lifecycle events: BeforeTool / AfterTool, BeforeAgent / AfterAgent, BeforeModel / AfterModel, BeforeToolSelection, SessionStart / SessionEnd, Notification, PreCompress - Gemini's own vocabulary, not Claude's).
• .gemini/commands/*.toml — custom commands (required prompt, optional description; subdirectories namespace as /parent:child). This replaces the old gemini_cli_config.json JSON command map.
• .gemini/skills/<slug>/SKILL.md — Agent Skills (open standard), activated via the activate_skill tool.
• .gemini/agents/<name>.md — subagents (frontmatter name, description, kind, tools, model, …).
• GEMINI.md (delegation shim → AGENTS.md) and rules/gemini-rules.md document conventions.

Transition note: On 2026-06-18 Google moved free and Google AI Pro / Ultra consumer access to Antigravity CLI. Paid Gemini Code Assist Standard / Enterprise and qualifying API-key users retain Gemini CLI access.

Per-runtime details: runtimes/.gemini/README.md.

Kiro

cp -R runtimes/.kiro/ /path/to/your/repo/.kiro/

Kiro consumes:

• .kiro/steering/<name>.md — rule files with inclusion: always or inclusion: fileMatch frontmatter.
• .kiro/specs/<feature>/{requirements,design,tasks}.md — the spec triplet (same shape as the framework-wide specs/ triplet).
• .kiro/hooks/<name>.json - JSON hook configurations, root "version": "v1".
• .kiro/skills/<slug>/SKILL.md and .kiro/agents/<name>.md - Agent Skills and subagents.
• .kiro/settings/mcp.json - MCP servers (mcpServers JSON).

Mirror your project's spec triplet content from specs/<feature>/ into .kiro/specs/<feature>/ (or symlink, where supported).

Breaking change - Kiro IDE 1.0 (2026-06-25): Kiro replaced *.kiro.hook with .kiro/hooks/<name>.json, moved to ten Claude-compatible trigger names (SessionStart, Stop, PreToolUse, PostToolUse, PreTaskExec, PostTaskExec, UserPromptSubmit, PostFileCreate, PostFileSave, PostFileDelete), and swapped askAgent / runCommand for action.type: "command" or "agent". Manual was retired; use a manual steering file. Legacy files show an upgrade badge and do not execute until migrated. Shape and migration checklist: hooks/kiro/README.md.

Per-runtime details: runtimes/.kiro/README.md.

Cursor

cp -R runtimes/.cursor/ /path/to/your/repo/.cursor/

Cursor consumes far more than rules - earlier SpecRoute matrices understated it, and Skills, Subagents, Hooks and Plugins are all recent additions:

• .cursor/rules/*.mdc - Markdown with frontmatter (description, globs, alwaysApply); AGENTS.md is also natively read. .cursorrules is gone - removed from Cursor's documentation entirely and reported non-functional in current versions. Treat it as removed, not legacy-but-supported.
• .cursor/mcp.json — MCP servers (mcpServers JSON).
• .cursor/skills/<slug>/SKILL.md - Skills, the same open-standard shape as Claude Code and Codex.
• .cursor/agents/<name>.md - Subagents.
• .cursor/commands/*.md - custom slash commands.
• .cursor/hooks.json - lifecycle Hooks, version: 1. 21 events, camelCase (beforeShellExecution, afterFileEdit, …), handler types command and prompt. Exit code 2 denies the action; hooks fail open on error unless the entry sets failClosed: true.
• Plugins / Marketplace - Cursor distributes rules, skills, and hooks as installable plugins.

For each rule file in rules/, create a corresponding .mdc in .cursor/rules/:

---
description: <one-line>
alwaysApply: true             # or
globs: ["src/**/*.tsx"]       # for context-aware loading
---

(rule body — copy from rules/<topic>.md)

Per-runtime details: runtimes/.cursor/README.md.

Devin Desktop

cp -R runtimes/.devin/ /path/to/your/repo/.devin/

The runtime targets Devin Local, the next-generation agent harness shared with Devin CLI and intended to become Devin Desktop's primary local agent:

• AGENTS.md — native root and nested project instructions.
• .agents/skills/<name>/SKILL.md (recommended) or .devin/skills/<name>/SKILL.md — reusable skills, invoked as /skill-name.
• .devin/agents/<name>/AGENT.md — experimental custom subagent profiles.
• .devin/hooks.v1.json — eight lifecycle events (PreToolUse, PostToolUse, PermissionRequest, UserPromptSubmit, Stop, PostCompaction, SessionStart, SessionEnd) with command and prompt handlers.
• .devin/config.json — project configuration, including the mcpServers object. Keep personal values in .devin/config.local.json.

Devin Desktop still includes Cascade during the transition. Cascade continues to use .windsurf/workflows/*.md, .windsurf/hooks.json, and ~/.codeium/windsurf/mcp_config.json; .windsurf/rules/ and .windsurf/skills/ remain accepted compatibility locations. These literal paths are Devin Desktop compatibility namespaces, not a separate vendor or runtime. Do not infer .devin/workflows/, .devin/hooks.json, or .devin/mcp.json.

Per-runtime details: runtimes/.devin/README.md.

MCP single source of truth

All six vendors consume MCP configs, in different shapes and paths - Claude Code .mcp.json (project) / ~/.claude.json (user), Codex [mcp_servers] TOML, Gemini .gemini/settings.json, Kiro .kiro/settings/mcp.json, Cursor .cursor/mcp.json, and Devin Desktop .devin/config.json. Apart from Codex's TOML, every vendor uses the same mcpServers JSON object. There are six renderers:

python3 runtimes/mcp/render/render_claude.py   > runtimes/.claude/mcp.template.json
python3 runtimes/mcp/render/render_codex.py    > runtimes/.codex/config.template.toml
python3 runtimes/mcp/render/render_gemini.py   > runtimes/.gemini/settings.template.json
python3 runtimes/mcp/render/render_kiro.py     > runtimes/.kiro/settings/mcp.template.json
python3 runtimes/mcp/render/render_cursor.py   > runtimes/.cursor/mcp.template.json
python3 runtimes/mcp/render/render_devin.py    > runtimes/.devin/config.template.json

Cascade's compatibility path ~/.codeium/windsurf/mcp_config.json accepts the same mcpServers object, so the Devin renderer covers both destinations. Edit runtimes/mcp/servers.yaml, re-render, and commit the canonical source and rendered files together so reviewers see the impact in one diff. See MCP Integration.

Spec change in flight: as of 2026-07-27, the MCP revision dated 2026-07-28 is a release candidate, not a published specification. Re-check the final revision before changing transport wiring. SpecRoute's servers.yaml schema is unaffected.

Note: claude_desktop_config.json is the Claude Desktop app's MCP file, not the Claude Code CLI's — the CLI reads .mcp.json / ~/.claude.json.

Cross-vendor sync

Skills share one shape across every vendor - that is the layer where cross-vendor convergence is near-total. sync-skills.py mirrors skill bodies across all six runtime layouts while preserving each vendor's native frontmatter; it deliberately excludes agents because their formats diverge. To keep skills aligned:

python3 tools/sync-skills.py             # dry-run, reports drift
python3 tools/sync-skills.py --apply     # actually copy
python3 tools/sync-skills.py --source codex --apply   # reverse direction

The /parity command (in this repo's .claude/commands/) is the consumer-facing wrapper. See Cross-Vendor Sync.

See also

• Vendor Matrix — what each vendor supports
• Multi-Vendor Context Files — AGENTS.md + delegation shims
• Adding a Vendor — the process for adding a new column
• MCP Integration — single source of truth for MCP server inventory