Everything an AI client needs to use Homebase, the asset library for AI agents: the Homebase skill (how to use the product), a plugin that packages the skill with the MCP server and slash commands, and marketplace files so Claude Code and Codex can install it with one command. The command-line tool is homebase-cli on npm.
skills/homebase/ the skill
SKILL.md what Homebase is, when to use it, search/upload/bucket workflows, safety rules, errors
references/cli.md every CLI command and its JSON output
references/mcp-tools.md the MCP tools, mapped to the CLI
references/setup.md installing, logging in, connecting each client
agents/openai.yaml Codex display metadata (ignored elsewhere)
assets/, LICENSE icons and license
plugins/homebase/ one plugin folder for Claude Code, Codex and Cursor
.claude-plugin/plugin.json Claude Code manifest (MCP server inline, type http)
plugin.json Agent Plugins 1.0 manifest (Codex, Cursor) + com.openai interface
mcp.json Agent Plugins MCP config (streamable-http)
commands/{find,upload,connect,status}.md /homebase:* slash commands (Claude Code)
skills/homebase/ the skill, plus scripts/homebase.mjs (the bundled CLI)
assets/ logo.png (512), icon.png (192), logo.svg
README.md, LICENSE the plugin's README and MIT license
.claude-plugin/marketplace.json Claude Code marketplace (Cursor reads it too)
.agents/plugins/marketplace.json Codex marketplace
CHANGELOG.md, LICENSE one changelog for the CLI and plugin; MIT
Without a shell (claude.ai, ChatGPT) the skill uses the Homebase MCP tools. With a shell (Claude Code, Codex) it uses whichever takes fewer calls: the MCP tools or the homebase CLI for reads, and always the CLI for uploads (one call, no curl) and downloads. The CLI prints one JSON document per command. No global install is needed: the plugin ships the CLI inside the skill folder. Claude Code commands run it as node "${CLAUDE_PLUGIN_ROOT}/skills/homebase/scripts/homebase.mjs", and Codex resolves scripts/homebase.mjs relative to SKILL.md. (Not a top-level bin/: Claude Code would put that on the PATH, but claude.ai and Cowork refuse to install plugins that have one.) Anywhere else with Node 20 or newer, npx -y homebase-cli <command> runs the same CLI.
| Client | How |
|---|---|
| Claude Code | claude plugin marketplace add feebleai/homebase-agent, then claude plugin install homebase@homebase (or /plugin marketplace add feebleai/homebase-agent and /plugin install homebase@homebase inside a session). You get the skill, the bundled CLI, the MCP server and /homebase:find, /homebase:upload, /homebase:connect and /homebase:status. To try it for one session from a clone of this repo: claude --plugin-dir plugins/homebase. Authenticate the MCP server with /mcp → homebase → Authenticate; over SSH, claude mcp login plugin:homebase:homebase --no-browser and paste the redirect URL back. For the skill alone, copy skills/homebase to ~/.claude/skills/ or <project>/.claude/skills/ |
| Codex (CLI and app) | codex plugin marketplace add feebleai/homebase-agent, then codex plugin add homebase@homebase. That installs the skill, the bundled CLI and the MCP server; then codex mcp login homebase (OAuth; --no-browser over SSH). For the skill alone, copy skills/homebase to ~/.agents/skills/ (or ~/.codex/skills/). For the MCP server alone, use npx -y homebase-cli connect codex |
| Cursor | cursor-agent plugin marketplace add https://github.com/feebleai/homebase-agent, or put the plugins/homebase folder in ~/.cursor/plugins/local/. Cursor reads the Agent Plugins plugin.json. For the MCP server alone, run npx -y homebase-cli connect cursor (deeplink + mcp.json snippet) |
| claude.ai / Claude Desktop | Zip the skill folder so homebase/SKILL.md is at the top (from a clone: cd skills && zip -r ../homebase-skill.zip homebase), then upload it in Customize → Skills → + → Create skill → Upload a skill (needs Settings → Capabilities → code execution). Add the MCP server under Customize → Connectors → Add custom connector with <api>/mcp; npx -y homebase-cli connect claude-ai prints the steps and the URL |
| ChatGPT | There is no package to install. Go to chatgpt.com/plugins → + → Create custom MCP server with <api>/mcp and OAuth (npx -y homebase-cli connect chatgpt). Custom MCP servers need Developer mode: OpenAI's help center lists it for Business, Enterprise and Edu (an admin allows it; members turn it on in Settings → Apps → Advanced settings). If Create custom MCP server is missing, turn Developer mode on first. Optionally, upload the same skill zip as for claude.ai under Skills → Create → Upload from your computer |
- Claude Code
-p: allow the Skill tool as well as specific commands, e.g.--allowedTools "Skill" "Bash(homebase search *)" "Bash(homebase get *)", rather than all ofhomebase. IfSkillisn't allowed, the skill is denied and the model may make up a reason. Never pre-allowhomebase deleteorhomebase login. - Codex
exec:- Its approval policy is
never, so MCP tools that would ask (import_from_url,remove_from_bucket,delete_assets) are refused. Approve specific tools in config: for the plugin's server under[plugins."homebase@homebase".mcp_servers.homebase.tools.<tool>] approval_mode = "approve"(not[mcp_servers.homebase…], which only applies to a server added withhomebase connect codex). Details inskills/homebase/references/setup.md. - Shell commands (the CLI) need
--sandbox workspace-writeandsandbox_workspace_write.network_access = true. - Close stdin (
< /dev/null) in scripts.
- Its approval policy is
- An MCP connector that shows "needs auth" doesn't affect the CLI. The skill says so, because a model once took the connector state as the reason the CLI couldn't work.
This repository is published from Homebase's main source on every release (v<version> tags), so changes made here directly would be overwritten. Report problems and suggestions in Issues.