Supported agents and compatibility

Tokenade supports the coding agents below. tokenade install detects the ones on your machine and wires each through the strongest channel that agent offers. What an agent gets depends on that channel: a hook or plugin that sees every tool call covers more than an MCP-only integration.

How Tokenade attaches to an agent

Each agent also receives a short Tokenade rules section in the global rules file it loads, so it knows the tokenade commands (map, skeleton, query, semantic, read, search…).

Feature matrix

Legend: Yes = covered automatically. Partial = covered, but not automatically or not for all traffic. Partial+ = partial with the binary alone, full once the LLM proxy is set up for that agent. No = not covered.

"Code search" means the tokenade map, skeleton, query, impact and semantic commands (details). "Context" covers context-window optimisation such as the auto-compact threshold and folding history. "Documents & media" is tokenade read on PDF, Office, audio and video files (details).

Where these results come from

The matrix comes from Tokenade's compatibility probes. Each probe ran a real tokenade install in a throwaway home on Linux, macOS (Apple Silicon) and Windows 10, then exercised each feature through the channel that install produced. The last full probe run was on version 1.1.7. The changes since then that affect this table are listed below.

Desktop apps (1.2.1)

• Claude Desktop, Code tab: runs Claude Code itself with your ~/.claude settings, so it gets the full Claude Code row. Since 1.2.1, tokenade install sets it up even when no separate claude command is installed, by checking the copy of Claude Code bundled with the app.
• Claude Desktop, Chat: only MCP optimisation applies. Chat runs no local commands or file reads. tokenade install wraps the MCP servers in claude_desktop_config.json, including the Microsoft Store build on Windows.
• Codex app: runs Codex on your machine with your ~/.codex/config.toml. Since 1.2.1 it gets the Codex coverage through the local LLM proxy (folded tool output, scrubbed credentials, the style note, savings figures), and MCP optimisation through mcp-wrap. The app's own built-in MCP server (browser and computer use) is left untouched.
• Claude Cowork: not supported. It runs in an isolated environment that local tools cannot reach.

MCP savings in desktop apps are credited to the app by name in your savings since 1.2.1, instead of "unknown".

Hooks that need your approval

Some agents only run hooks after a validation step. Tokenade checks, per agent, whether its hooks really run, and lets the LLM proxy cover what they would have done only when they do not:

• Codex: each hook needs a trust entry in ~/.codex/config.toml. tokenade install writes it; tokenade codex-trust re-applies it. See Codex.
• Claude Code: hooks are switched off by disableAllHooks, and do not run in a folder whose trust prompt you have not accepted. tokenade healthcheck reports the second case. See Claude Code.
• Qwen Code: needs hooks.enabled: true in ~/.qwen/settings.json.
• OpenClaw, Hermes: the Tokenade plugin must be enabled in the agent's plugin list (install does it).

Agents no longer supported

Gemini CLI and Reasonix are no longer supported since 1.1.19. tokenade uninstall still removes the hooks and settings older versions wrote for them. Windsurf is now handled as Devin Desktop (same paths). See Antigravity.

Keeping the agent's own web search

To keep web queries inside your AI vendor, set TOKENADE_NO_WEBSEARCH=1 in the agent's environment and run tokenade install once with it set. Tokenade then leaves the agent's web search alone, tokenade search refuses to run, and the rules it writes no longer mention it. Page fetching has its own switch, TOKENADE_NO_WEBFETCH=1. See Environment variables.

Related

• Install Tokenade
• How hooks intercept tool calls
• Wrap MCP servers
• The LLM proxy
• Other agents