English · 简体中文

Every MCP server and Skill your AI agents need — on demand, from one place.
One catalog, one secret store and lazy MCP servers for Claude Code, Codex, DeepSeek Harness, Claude Desktop and Hermes.

Install a hundred Skills and a dozen MCP servers and every agent session pays for all of them: descriptions fill the context window, MCP processes sit running, and each agent keeps its own copy of the config and the API keys.

Fusebox gives every MCP server and Skill one of two states:

• Resident — loaded in every session, because you asked for it.
• On demand — everything else. Agents see the name, find it when a task needs it, load it, and the MCP process stops again when idle.

It writes each agent's native config from one registry, keeps secrets in one file that only the MCP that declares a key can read, and measures from your own sessions what gets used.

# macOS: CLI + app
brew tap risingjerry/tap && brew trust --tap risingjerry/tap
brew install risingjerry/tap/fusebox
brew install --cask risingjerry/tap/fusebox

# any platform with Rust (CLI; the app needs macOS + Xcode)
git clone https://github.com/risingjerry/fusebox && fusebox/scripts/install.sh

Then:

fusebox doctor                     # what it found and what it would write (writes nothing)
fusebox reconcile                  # write every agent's config (backups first)
fusebox install skill owner/repo   # or a GitHub URL / local folder; cloned at a locked commit
fusebox uninstall skill NAME       # gone from every profile, agent, registry, lock and vendor/
fusebox install mcp github --command npx --arg -y --arg @example/github-mcp@1.2.3 --secret GITHUB_TOKEN
printf %s "$TOKEN" | fusebox secret set --stdin GITHUB_TOKEN   # the value never appears in argv or logs

Every command that changes the setup syncs all agents when it finishes, and each agent's instructions tell it to install through Fusebox: ask any agent to "install this skill: " and it lands here, not in one agent's folder.

New MCP servers and Skills are on demand by default. Make one resident with fusebox mcp direct ID or fusebox skills enable global NAME.

Install the bundled skill so any agent knows how to drive Fusebox:

npx skills add risingjerry/fusebox        # or copy skills/fusebox to ~/.claude/skills/

Then ask:

• "Find a skill for turning this outline into slides."
• "Which of my Skills and MCP servers are worth keeping resident? Show me, then apply what I approve."
• "Add the Brave Search MCP on demand; I'll paste the key."

Agents call fusebox find, fusebox advise --json and the hub's fusebox_servers / fusebox_call tools. AGENTS.md is the guide for agents working on this repository.

Every read command takes --json. Full help: fusebox --help.

             fusebox reconcile
config.json ───────────────────▶ ~/.claude.json, ~/.claude/settings.json, ~/.claude/skills
secrets.env                      ~/.codex/config.toml + AGENTS.md
Skill Hub                        ~/.dsh/cordis.patch.yml + AGENTS.md, Claude Desktop, Hermes

agent ──▶ agent-hub (one stdio MCP) ──first call──▶ wrapper ──▶ MCP server (own secrets only)
                                     ◀──idle 5 min── stop

• Fusebox owns only agent-* MCP entries and its own marked instruction blocks; everything else in those files is preserved.
• Skills live in a Skill Hub (default ~/.config/fusebox/skill-hub; set skillHub in config.json or SKILL_HUB_HOME to use another) with locked git sources and per-project profiles.

Not in our measurement. On 112 real prompts where an agent loaded a Skill, a model shown only names picked the right Skill in its top 3 as often as one shown every full description; names plus a one-line description for the 30 most used did better (53% vs 47–48%, two runs each) at a third of the size. Run fusebox eval to measure find on your own sessions.

How do I stop MCP servers from filling Claude Code's context? Put them behind agent-hub (fusebox mcp hub ID, the default for new ones). Claude Code sees the hub's tools by name and loads definitions only when used.

How do I share one set of MCP servers between Claude Code and Codex? Add each server once with fusebox mcp add; fusebox reconcile writes both configs. Both reach every on-demand server through the hub.

Where are my API keys stored? Can an agent read them? In ~/.config/fusebox/secrets.env (mode 0600). Native configs only reference a wrapper; the wrapper exports only the keys that server declares. Agents see key names, never values.

How do I find which Skills I never use? fusebox skills usage --days 60 and fusebox advise.

Does it work on Linux or Windows? The CLI builds and passes CI on Linux; the app is macOS only. Windows is untested.

Is the app signed? Ad-hoc signed, not notarized (no Apple Developer ID yet). The Homebrew cask removes the quarantine flag and says so. Build from source if you prefer.

brew uninstall --cask fusebox; brew uninstall fusebox   # or delete ~/.local/bin/fusebox

Your agents' configs keep working until you remove the agent-* entries; backups of every file Fusebox changed are in ~/.config/fusebox/backups.

• Remote (HTTP/SSE) MCP servers are not managed yet; Fusebox leaves them untouched.
• Codex has no names-only Skill mode, so Fusebox lists on-demand names in its instructions.
• The hub does not forward server-to-client requests (sampling, elicitation).

If Fusebox saved your context window, a �� helps others find it.

MIT © risingjerry · Previously named Agent Switch.