agent-hub

CLI + TUI to manage skills and MCP servers across:

Harness	Skills (live)	MCP
Cursor	~/.cursor/skills	~/.cursor/mcp.json
Claude Code	~/.claude/skills	~/.claude.json
Codex	~/.codex/skills	~/.codex/config.toml
pi	~/.pi/agent/skills	~/.pi/agent/mcp.json
OMP	~/.omp/agent/skills	~/.omp/agent/mcp.json

Model: profile = workspace

A profile is an active workspace. Switching profile changes which skills (and MCP flags) harnesses actually see.

~/.agent-hub/
  library/skills/           # all skill packages (SoT contents)
  workspaces/<profile>/skills/  # subset for this profile (symlinks → library)
  profiles/<profile>.yaml   # per-harness enable matrix
  catalog/mcp.yaml
  state.json                # active profile

• Workspace — profile-private skills (workspaces/<profile>/skills/) and MCP (workspaces/<profile>/mcp.yaml).
• Library — migrate leftovers / orphans only (not where new installs go).
• Apply — parks shared ~/.agents/skills (empty), syncs workspace, projects into harness dirs, then prunes harness skills/MCP that are not enabled in the profile (empty profile → clear hub-managed projections).
• Install / remove — only the active profile. Same name in another profile or library → conflict.
• Adopt — claim an orphan library skill / global MCP into the active profile.
• Sync / watch — external installs in ~/.agents/skills and harness dirs (~/.cursor/skills, ~/.codex/skills, ~/.pi/agent/skills, …) move into the active profile. Hub symlinks are ignored. Switching profile keeps those skills only on the previous profile.

~/.agents/skills is parked empty on purpose: pi/Cursor/Codex also auto-load that path, so a shared pool would leak skills across profiles. Live skills live only in harness-private dirs.

Inspired by agent-toggle’s TUI — but the SoT is a workspace + per-harness matrix, not park-in-place only.

No GUI. Config + CLI + TUI.

Agent skill

Bundled skill for coding agents: skills/agent-hub/SKILL.md
(also linked at .agents/skills/agent-hub and .cursor/skills/agent-hub for project discovery).

Install into an agent-hub profile:

# from any directory — relative paths also fall back to the agent-hub package root
agent-hub skill install ./skills/agent-hub --apply
agent-hub skill install agent-hub --apply

Install

git clone git@github.com:servertdot/agent-hub.git
cd agent-hub
bun install
bun link          # puts `agent-hub` / `ahub` on PATH via bun
# or: bun run src/cli.ts …

Requires Node 22+ or Bun.

Quick start

agent-hub init                 # ~/.agent-hub + example profiles
agent-hub migrate              # absorb ~/.agents/skills → library; park shared empty
agent-hub status
agent-hub discover             # matrix + token estimates
agent-hub cost                 # biggest startup costs first
agent-hub import-mcp           # fill catalog/mcp.yaml from Cursor/Claude/OMP JSON
agent-hub scrub-catalog        # replace pasted secrets with $ENV refs
agent-hub scrub-harnesses --write
agent-hub profile save stable  # capture current discovery into a profile
agent-hub profile seed-lab     # clone stable → lab + omp extras
agent-hub diff lab             # dry-run (hides noops; -v for full)
agent-hub apply lab            # activate workspace + project to harnesses
agent-hub skill install ./my-skill --apply
agent-hub skill remove my-skill --apply
agent-hub skill adopt orphan-name --apply
agent-hub skill copy pdf-reader --from stable --to lab --apply
agent-hub skill copy --all --from stable --to lab --apply
agent-hub sync                 # ~/.agents/skills → active profile (after npx skills add)
agent-hub sync --apply
agent-hub watch                # auto-sync ~/.agents/skills into active profile
agent-hub mcp install ./server.yaml --name fff --apply
agent-hub mcp remove fff --apply
agent-hub tui                  # Ink matrix on a TTY
agent-hub tui --menu           # numbered fallback (also used without TTY)
agent-hub tui --dry-run

Layout

~/.agent-hub/
  config.yaml
  catalog/mcp.yaml
  library/skills/
  workspaces/
    lab/skills/
    stable/skills/
  profiles/
    stable.yaml
    lab.yaml
  state.json

Profile shape:

name: lab
description: Experimental
skills:
  tdd:
    cursor: true
    claude: false
    codex: false
    pi: false
    omp: false
mcp:
  fff:
    cursor: true
    claude: false
    codex: false
    pi: true
    omp: false

Only keys you set are applied. Omitted harness cells are left alone.

Skill with all harnesses false → removed from workspace and from harness dirs on apply.

TUI (agent-toggle flavoured)

• ● live / ○ off / - unknown
• + / - before glyph = staged change
• Tab toggles focused cell; ←→ moves harness column; ↑↓ rows
• Top tabs: Skills | MCP | Harnesses (t / 1 / 2 / 3)
• Selecting a profile in the sidebar loads it for editing — it does not apply
• Enter applies (save profile + workspace sync + harness projection); Esc cancels
• w saves profile YAML only
• c copy focused skill / C copy all visible skills → pick target profile in sidebar (Enter)

Safety

• Disabling a skill removes harness symlinks; real dirs are moved into the library (never deleted blindly)
• Shared ~/.agents/skills is emptied/parked on migrate/apply so profiles cannot leak
• MCP edits merge into existing JSON; Codex TOML flips enabled and keeps subtables
• Prefer $ENV_NAME in MCP configs — do not commit secrets
• Always diff / tui --dry-run before the first real apply

Status (0.1)

• workspace library + per-profile skills
• migrate / apply parks shared Agent Skills pool
• skill install → active workspace
• discover / profiles / Ink TUI (stage → apply)
• cost estimates, MCP import/scrub

License

MIT