Vendor Matrix

Vendor Matrix

The supported-CLI contract. Adding a new tool means a new column, not a fork. This matrix is the single most important table in SpecRoute - it is mirrored in four files (README.md, AGENTS.md, agentic-docs/agent-cli-integrations.md, and this page) and must stay in lock-step across all four.

Vendor	Runtime dir	Root context file	Skills	Agents	Commands	Hooks	MCP config
Claude Code	.claude/	CLAUDE.md	folder-per-skill SKILL.md	flat <name>.md + frontmatter	commands/<name>.md (merged into skills)	settings.json hooks (30 events; 5 handler types)	.mcp.json / ~/.claude.json
Codex	.codex/	AGENTS.md	folder-per-skill SKILL.md	agents/<name>.toml	skills (/skills, $mention)	hooks.json or config.toml [hooks] (11 events; Claude-compatible names)	config.toml [mcp_servers]
Gemini CLI / Antigravity	.gemini/	GEMINI.md (delegation shim)	skills/<slug>/SKILL.md	agents/<name>.md	TOML in .gemini/commands/	settings.json hooks (11 events; Before* / After* names)	settings.json [mcpServers]
Kiro	.kiro/	steering/ files	skills/<slug>/SKILL.md	agents/<name>.md	/skill + manual steering	.kiro/hooks/<name>.json v1 (10 triggers)	.kiro/settings/mcp.json
Cursor	.cursor/	.cursor/rules/*.mdc / AGENTS.md	skills/<slug>/SKILL.md	agents/<name>.md	commands/*.md	.cursor/hooks.json v1 (21 events, camelCase)	.cursor/mcp.json
Devin Desktop	.devin/	AGENTS.md	.agents/skills/ (recommended) / .devin/skills/	.devin/agents/<name>/AGENT.md (experimental)	skills (/skill-name); Cascade workflows	.devin/hooks.v1.json (8 events); Cascade hooks	.devin/config.json [mcpServers]; Cascade MCP

Vendor notes and sources

• Root context files. Claude Code reads CLAUDE.md, not AGENTS.md - the official documentation says so directly. Gemini CLI reads GEMINI.md and only picks up AGENTS.md when you opt in via context.fileName; the upstream issue asking for default support was closed as not planned. Bridge either with an @AGENTS.md import inside the vendor file, or a symlink. Codex, Cursor, and Devin Desktop read AGENTS.md natively.
• Claude Code hooks. Claude Code loads hooks from settings.json (user / project / local), managed policy settings, a plugin's hooks/hooks.json, and skill or agent frontmatter. A bare project-level .claude/hooks/hooks.json is not a hook source - keep it as the authoring artifact and merge its hooks key into .claude/settings.json. The five handler types are command, http, mcp_tool, prompt, and agent.
• Codex hooks. Verified against an installed codex-cli 0.145.0 binary: SessionStart, SessionEnd, PreToolUse, PermissionRequest, PostToolUse, UserPromptSubmit, Stop, PreCompact, PostCompact, SubagentStart, SubagentStop. Codex adopted Claude Code's event names and handler JSON (hookSpecificOutput, permissionDecision, decision: "block") almost verbatim - but the match is close, not total: Claude's PostToolUseFailure has no Codex counterpart, and only type: "command" actually executes in Codex (prompt and agent handlers are parsed, then skipped). Hooks are enabled by default; the canonical [features] key is hooks, with codex_hooks retained as a deprecated alias. Disable with [features] hooks = false.
• Gemini CLI / Antigravity. Google redirected consumer access to Antigravity CLI on 2026-06-18 for free and Google AI Pro / Ultra tiers. Paid Gemini Code Assist Standard / Enterprise and qualifying API-key users retain Gemini CLI access. The .gemini/ integration layout remains valid across both.
• Kiro hooks. Kiro IDE 1.0 (2026-06-25) replaced the *.kiro.hook format with .kiro/hooks/<name>.json (root "version": "v1"). Triggers (10): SessionStart, Stop, PreToolUse, PostToolUse, PreTaskExec, PostTaskExec, UserPromptSubmit, PostFileCreate, PostFileSave, PostFileDelete. The 0.x Manual trigger was retired - manual invocation is now a steering file. Actions are action.type: "command" or "agent", replacing runCommand / askAgent. Legacy *.kiro.hook files show an upgrade badge and no longer execute.
• Cursor rules. .cursorrules has been removed from Cursor's documentation entirely and is reported non-functional in current versions. Treat it as removed rather than legacy-but-supported; use .cursor/rules/*.mdc. Cursor also ships Skills, Subagents, Hooks, and a Plugins marketplace.
• Devin Desktop. Devin Desktop is the new name for Windsurf. Devin Local uses the .devin/ paths shown above and is intended to become the primary local agent. Cascade remains available and still uses .windsurf/workflows/*.md, .windsurf/hooks.json, and ~/.codeium/windsurf/mcp_config.json; .windsurf/rules/ and .windsurf/skills/ remain accepted compatibility locations. These are Devin Desktop configuration namespaces, not a separate vendor or runtime. Do not infer .devin/workflows/, .devin/hooks.json, or .devin/mcp.json.

Capability convergence

Earlier SpecRoute releases sorted vendors into capability tiers (Full / Near-full / Partial / Specialized / Rules-only). As of mid-2026 that tiering no longer holds - every supported tool ships every capability class: the Agent Skills open standard (SKILL.md), subagents, custom commands, lifecycle hooks, and MCP.

Two different claims used to be collapsed into one table here, and that conflation is what produced the drift. They are now separated.

1. Does the vendor support the capability?

Vendor	Skills	Agents	Commands	Hooks	MCP
Claude Code	✅	✅	✅	✅	✅
Codex	✅	✅	✅	✅	✅
Gemini CLI / Antigravity	✅	✅	✅	✅	✅
Kiro	✅	✅	✅	✅	✅
Cursor	✅	✅	✅	✅	✅
Devin Desktop	✅	✅	✅	✅	✅

2. Does SpecRoute ship a runtime layout for it?

Measured against runtimes/ on disk, not against vendor capability. A dash means that artifact is absent from the SpecRoute runtime layout; the note below distinguishes intentional omissions from open gaps.

Runtime layout	Skills	Agents	Commands	Hooks	MCP
Claude Code - runtimes/.claude	✅	✅	✅ commands/	✅ hooks/	✅ mcp.template.json
Codex - runtimes/.codex	✅	✅	- (skills serve as commands)	✅ hooks/	✅ config.template.toml
Gemini CLI - runtimes/.gemini	✅	✅	✅ commands/	✅ hooks/ (snippet)	✅ settings.template.json
Kiro - runtimes/.kiro	✅	✅	- (steering + /skill)	✅ hooks/	✅ settings/mcp.template.json
Cursor - runtimes/.cursor	✅	✅	✅ commands/	✅ hooks/	✅ mcp.template.json
Devin Desktop - runtimes/.devin	✅	✅	skills (/name)	✅ hooks/	✅ config.template.json

All six runtime layouts ship hook assets. Gemini is the one exception in shape: its hooks live under the hooks key inside .gemini/settings.json, so the layout ships a documented merge snippet rather than a standalone file.

What has actually converged

Vendors differ in file format and conventions - Codex agents are TOML where the others are Markdown; Gemini commands are TOML where Claude / Cursor are Markdown; MCP config is JSON for most but TOML for Codex.

• Skills: converged. The SKILL.md shape is portable across all six with little more than a path change. This is the one near-total convergence.
• Hooks: not converged. See below.
• MCP: converged on the payload, split on the container - one mcpServers JSON object everywhere except Codex's TOML.

Hook system depth

All six vendors ship hooks; the systems have not converged. Claude Code, Codex, Kiro, and Devin Local share several core event names, but their payload and decision schemas still differ. Cursor and Gemini CLI retain distinct vocabularies.

Vendor	Event count	Naming vocabulary	Blocking semantics	Handler types
Claude Code	30	PreToolUse / PostToolUse / …	rich (per-event)	command, http, mcp_tool, prompt, agent
Codex	11	Claude-compatible	per-event JSON	command (only type that executes)
Gemini CLI	11	BeforeTool / AfterModel / BeforeToolSelection	per-event	command
Kiro	10	Claude-compatible since IDE 1.0	pre-hooks block	command, agent
Cursor	21	camelCase (beforeShellExecution, …)	exit 2 = deny; fail-open unless failClosed: true	command, prompt
Devin Desktop	8	PreToolUse / SessionStart style	exit 2 blocks	command, prompt

See Hooks for the per-vendor event matrix.

Root context files (load-bearing)

Every vendor auto-loads at least one context file at session start. Codex, Cursor, and Devin Desktop auto-load AGENTS.md. Claude Code's documentation states plainly that it reads CLAUDE.md, not AGENTS.md; Gemini CLI reads GEMINI.md and treats AGENTS.md as opt-in via context.fileName. The sanctioned bridge for both is an @AGENTS.md import inside the vendor file, or a symlink.

Vendor	Auto-loaded file	Reads AGENTS.md natively?	What goes there
Claude Code	CLAUDE.md	no - bridge via @AGENTS.md import or symlink	Delegation shim → AGENTS.md + Claude-specific overrides
Codex	AGENTS.md	yes	Canonical, vendor-neutral substance
Gemini CLI	GEMINI.md	no - opt in via context.fileName	Delegation shim → AGENTS.md + Gemini-specific overrides
Cursor	.cursor/rules/*.mdc or AGENTS.md	yes	Rule files, MDC frontmatter; native AGENTS.md support. .cursorrules is removed, not legacy - it is gone from Cursor's docs and reported non-functional
Kiro	.kiro/steering/*.md (inclusion: always)	no	Steering rules
Devin Desktop	AGENTS.md	yes	Canonical rules; .devin/rules/ is an additional rule location

The canonical pattern: AGENTS.md carries substance; CLAUDE.md / GEMINI.md etc. are short delegation shims. See Multi-Vendor Context Files.

The AGENTS.md convention itself

AGENTS.md started as an OpenAI convention and is now stewarded by the Agentic AI Foundation under the Linux Foundation, with 60,000+ repositories using it. It specifies no required fields, no frontmatter, and no schema - plain Markdown, contents entirely up to the project. Nested AGENTS.md files are spec'd behaviour: an agent reads the nearest one up the directory tree.

A vendor-neutral .agents/ directory is emerging as the companion location for runtime artifacts. SpecRoute verified this against an installed Codex 0.145.0 binary rather than documentation, because public docs are inconsistent: the binary carries .agents/skills as a repo-level skills root (error string failed to stat repo skills root), .agents/plugins/marketplace.json, .agents/plugins/api_marketplace.json, and enumerates .agents beside .claude and .cursor when detecting external agent configuration. The same binary still carries .codex/skills and $CODEX_HOME/skills, so both work today - do not migrate off .codex/skills yet.

MCP support

All six vendors now consume MCP server configurations, in different shapes:

Vendor	Shape	Path
Claude Code	mcpServers JSON object	.mcp.json (project) / ~/.claude.json (user)
Codex	[mcp_servers.<name>] TOML sections	.codex/config.toml
Gemini CLI / Antigravity	mcpServers JSON object	.gemini/settings.json
Kiro	mcpServers JSON object	.kiro/settings/mcp.json
Cursor	mcpServers JSON object	.cursor/mcp.json
Devin Desktop	mcpServers JSON object	.devin/config.json

Maintaining these by hand is the failure mode. SpecRoute ships a single source of truth at runtimes/mcp/servers.yaml with six per-vendor renderers: render_claude.py, render_codex.py, render_gemini.py, render_kiro.py, render_cursor.py, and render_devin.py. See MCP Integration.

Spec change in flight: an MCP specification revision dated 2026-07-28 removes sessions and the Mcp-Session-Id header - the protocol goes stateless - and formally deprecates HTTP+SSE in favour of streamable HTTP. Treat this as imminent rather than shipped, and check the revision date before relying on it. SpecRoute's servers.yaml shape is unaffected either way; only transport wiring is.

Note: claude_desktop_config.json is the Claude Desktop app's MCP file, not Claude Code's. The Claude Code CLI reads .mcp.json (project, committed) and ~/.claude.json (user). Earlier SpecRoute releases pointed at the Desktop file in error.

Vendor neutrality is the contract

• Don't fold one vendor into another.
• Don't treat any vendor as the default.
• Every artifact declares which vendors it targets and uses each vendor's native shape.
• Don't drift the matrix between README.md, AGENTS.md, agentic-docs/agent-cli-integrations.md, and this page - when one changes, all four change. The table body is byte-identical across all four by design; copy the whole table rather than editing one cell. Supporting notes live here and in the integration reference instead of the README.
• Don't conflate "the vendor supports it" with "SpecRoute ships a runtime layout for it". Those are the two tables under Capability convergence, and merging them is what caused the last round of drift.

See also

• Agent CLI Integrations — concrete per-vendor wiring (copy commands)
• Cross-Vendor Sync — tools/sync-skills.py + MCP renderers keep things aligned
• Adding a Vendor — process for adding a new column
• Multi-Vendor Context Files — the AGENTS.md + delegation-shim pattern