🦥 lazyagent

A lazygit-style TUI for skills, subagents, MCP servers, prompts and memory across Claude Code, Codex and Gemini CLI — one tree, one hotkey to share between tools.

 

_{Split-pane tree on the left, glamour-rendered detail on the right. j/k to navigate, tab to drill in, ? for help.}

TL;DR — One TUI for every Claude Code, Codex and Gemini CLI artifact in your home directory. Browse, edit, place a single canonical version into the lazyagent library and project it into all three tools, install skills/agents from any public GitHub repo. Local-first, no telemetry, brew-installable.

---

📣 News

• 2026-05-01 — PRI-26 Hooks are now a first-class Kind — Claude PreToolUse / PostToolUse / SessionStart entries in your settings.json show up in the tree with a "⚠ runs shell" warning so you see what will execute before it does.
• 2026-05-01 — PRI-10 ~/.lazyagent/ignore keeps work / corporate trees out of the global indexer. Manage it from the CLI: lazyagent ignore add '~/work/'.
• 2026-05-01 — PRI-19 Weekly GitHub releases check + amber "↑ vX.Y.Z available" footer when your build is behind. Disable with [updates] notify = false or --no-update-check.
• 2026-05-01 — PRI-4 Global project indexer + A toggle to fold every discovered project's local items into the same tree. Cache lives at ~/.lazyagent/index.json.
• 2026-05-01 — PRI-20 First-run empty-state with an ASCII logo + per-section "no <kind> yet" placeholders so a fresh install no longer paints a blank panel.
• 2026-04-30 — PRI-3 Install skills / agents / prompts from any public GitHub repo: i in the TUI or lazyagent install <url> from the shell. U updates an installed item to the origin's latest sha.
• 2026-04-30 — PRI-18 Broken frontmatter is now surfaced with an (invalid) badge instead of being silently skipped, with the diagnostic shown in the detail panel.
• 2026-04-30 — PRI-21 Shell completions for bash / zsh / fish; Homebrew installs them automatically, otherwise lazyagent completion <shell> prints the script.
• 2026-04-30 — PRI-17 Structured logging via slog + lumberjack. lazyagent logs tail is the new way to grab context for bug reports.
• 2026-04-30 — PRI-16 Optional ~/.lazyagent/config.toml for tool toggles, search roots, logging level and more — managed via lazyagent config {init,show,edit,validate}.
• 2026-04-29 — PRI-5 Resume Claude / Codex / Gemini sessions straight from the tree with R.
• 2026-04-29 — PRI-2 Shared canonical store with symlink projection and drift detection (s / R).

---

Why

Three CLIs, three home directories, three ways of doing the same thing. If you use Claude Code, Codex and Gemini CLI in the same week you end up with:

• Three parallel hierarchies — ~/.claude/, ~/.codex/ (+ ~/.agents/), ~/.gemini/ — each with its own skills/, agents/, prompts/, memory file and MCP config.
• Duplicates by hand: the only way to use a skill in all three tools is to cp -r it, which forks the moment you edit one copy.
• No common view: there's no way to see at a glance which prompts you have, which MCP servers are wired up, or which agent is project-local vs global.

lazyagent puts all of that in a single tree and — through a single unified place action — keeps one canonical copy of each item in ~/.lazyagent/library/ and projects it into every tool you pick via symlinks (or copies on cloud-synced volumes), so editing one place updates everywhere.

Install

Homebrew (macOS arm64 / amd64)

brew install mi-subbotin/tap/lazyagent

First launch on macOS triggers Gatekeeper. Right-click the binary → Open, or xattr -d com.apple.quarantine $(which lazyagent). Codesigning / notarization is on the roadmap but not yet shipped.

go install

go install github.com/mi-subbotin/lazyagent/cmd/lazyagent@latest

Pre-built binaries

Download from the latest release — darwin-arm64 and darwin-amd64 archives include a single lazyagent binary. Linux and Windows builds are tracked in the roadmap.

Quickstart

cd ~/Projects/your-project   # or anywhere
lazyagent

The tree is grouped Origin → Kind → Scope:

Claude
  Skills
    Global  (3)
    Local   (1)
  Agents
  ...
Codex
  ...
Gemini
  ...
Shared              ← appears once you press `p` to put an item into the library

Local-scope rows show only when lazyagent is launched from a directory that contains tool markers (.claude/, .codex/, .gemini/, AGENTS.md, CLAUDE.md, GEMINI.md, .mcp.json). Otherwise local sections are hidden — there's no project to attach to.

Features

Feature	Claude	Codex	Gemini
Skills (SKILL.md + assets)	✓	✓	✓
Subagents	✓	✓ ¹	✓
MCP servers	✓	✓	✓
Prompts / slash commands	✓	✓	✓ ²
Memory file (CLAUDE.md/AGENTS.md/...)	✓	✓	✓
Hooks (PreToolUse / PostToolUse / …)	✓ ⁴	—	—
Place to library + projections (p)	✓	✓ ³	✓ ³
Inline edit (E) + external editor (e)	✓	✓	✓
Drift detection + resync (R)	✓	✓	✓

¹ Codex agents are stored as [profiles.<name>] entries in config.toml; format conversion to/from agent.md is deferred (PRI-68).
² Gemini commands are TOML; conversion from Claude/Codex markdown is deferred (PRI-68).
³ The place picker greys out cells that need format conversion until PRI-68 lands. MCP/Hook entry-shape items are also deferred from p (PRI-69) — use the legacy CLI flow for now.
⁴ Codex / Gemini hook adapters are deferred — formats need verification on a live install (PRI-57).

_{Press tab on a leaf to zoom into a glamour-rendered detail view.}

Keys

  j/k      navigate up/down (scroll body in detail)
  h/l      collapse / expand group
  tab      open item full-screen / back out (toggle)
  enter    same as tab (drill into item or expand group)
  space    page down in fullscreen detail
  g / G    jump to top / end in fullscreen detail
  /        filter items by name/description
  esc      clear filter / cancel filter editor / cancel confirm
  t        toggle JSON / TOML for MCP entries
  d        delete item (asks y/n)
  p        place item — pick which (Origin × Scope) cells project the
           item from the library; bytes live once in ~/.lazyagent/library
           and project back to each chosen cell (replaces legacy c/m/x/s)
  R        resync drifted library item (canonical / tool wins)
  e        open in $EDITOR (external)
  E        edit in built-in editor (ctrl+s save · esc cancel)
  n        create new Skill / Agent / Prompt
  r        reload all sources
  ?        toggle this help (in-app)
  q        quit

lazyagent — read-only by default for everything it doesn't recognise. Editing actions are explicit and confirm before destructive ops.

How storage works — the library

lazyagent stores every shareable item once, in the library at ~/.lazyagent/library/, and projects it back to each tool that needs it. No duplicate copies, no fork-on-edit. The library is created on first launch — no separate init step. (An older ~/.lazyagent/store/ directory is migrated automatically.)

~/.lazyagent/library/
  skills/<name>/{SKILL.md, manifest.toml, ...}
  agents/<name>/{agent.md, manifest.toml}
  prompts/<name>/{prompt.md, manifest.toml}
  memory/<name>/{memory.md, manifest.toml}
  mcp/...                 (deferred — PRI-69)

Pressing p on any per-tool item opens a single picker:

Place skill foo:

  Library: yes  (canonical bytes — for optimization)

           Global   Local
  Claude   [x]      [ ]
  Codex    [ ]      [ ]
  Gemini   [x]      [ ]

  [arrows] move  [space] toggle  [enter] apply  [esc] cancel

Apply moves the bytes into the library (if not already there) and reconciles every (Origin, Scope) cell — symlink by default, byte-copy on iCloud / Dropbox / OneDrive / Google Drive volumes where symlinks don't sync. Unticking a cell unprojects it; an empty matrix is valid (item lives only in the library, no projections — like git stash). Pressing p again on a placed item lets you reshape the projection set. Pressing R on an item the detector flagged with (drift) opens the canonical-vs-tool resync overlay.

Configuration

Optional. lazyagent works out of the box with no config file at all — every option below has a baked-in default. When you want to override something, run lazyagent config init to seed ~/.lazyagent/config.toml, then edit it (lazyagent config edit is a shortcut for $EDITOR).

[search]
roots            = ["$HOME"]                        # PRI-4 global indexer roots
ignore_paths     = []                               # PRI-10 .gitignore-style excludes
follow_symlinks  = false

[ui]
theme            = "tokyonight"                     # tokyonight | catppuccin | nord
default_mode     = "cwd"                            # cwd | global
display_mode     = "origin"                         # origin | local-grouped
default_origin   = "all"                            # all | claude | codex | gemini

[tools]
claude           = true
codex            = true
gemini           = true
shared           = false

[install]
cache_dir        = "~/.lazyagent/cache"             # PRI-3 GitHub-install cache
gc_after_days    = 30

[updates]
check_interval_days = 7                             # PRI-19 release check cadence
notify              = true

[logging]
level            = "warn"                           # debug | info | warn | error
file             = "~/.lazyagent/logs/lazyagent.log"
format           = "text"                           # text | json

CLI helpers:

Command	What it does
lazyagent config init	Write the defaults above to ~/.lazyagent/config.toml. Refuses if the file exists; pass --force to overwrite.
lazyagent config show	Print the effective config (defaults + your overrides) to stdout.
lazyagent config edit	Open the file in $EDITOR (creates it from defaults first if missing).
lazyagent config validate	Parse the file, list unknown keys and invalid enum values, exit non-zero on any.

Partial files are fine — only put the keys you want to override; everything else stays at the default. Unknown keys and invalid enum values become warnings on stderr and the offender resets to its default rather than aborting startup.

Install items from GitHub

Pull skills, agents and prompts straight from a public repo — either via the TUI (i) or the shell.

# show what's installable in a repo
lazyagent install github.com/anthropics/skills --list

# install everything into Claude (global)
lazyagent install github.com/anthropics/skills --all

# install one item by name into Codex (global)
lazyagent install github.com/foo/bar --target=codex --name=cool-skill

# refresh an installed item to the origin's latest sha
# (TUI: cursor on the item, then U)
# CLI: re-run install with --overwrite

# remove an installed item; --target/--scope to disambiguate when needed
lazyagent uninstall cool-skill --target=codex

# sweep abandoned tarball cache (auto-runs every 30 days at startup)
lazyagent cache gc

Supported URL shapes:

• github.com/<owner>/<repo> — install at the default branch
• github.com/<owner>/<repo>/tree/<ref>[/path] — pin a tag/sha or narrow to one subdir
• github.com/<owner>/<repo>/blob/<ref>/path/file.md — single file
• gist.github.com/<id> (coming soon — currently rejected)

The detection is convention-based: skills/<name>/SKILL.md → Skill, agents/<name>.md → Agent, commands/<name>.md → Prompt. Set GH_TOKEN (or GITHUB_TOKEN) for private repos. Pinned shas land in ~/.lazyagent/installed.toml; tarballs cache under ~/.lazyagent/cache/ and are swept automatically.

Shell completions

lazyagent completion <bash|zsh|fish> prints the completion script for the chosen shell to stdout.

Homebrew installs all three automatically — restart your shell after brew install and tab-completion just works.

For non-brew installs:

# bash (one-shot, current shell)
source <(lazyagent completion bash)

# bash (persistent, system-wide on macOS)
lazyagent completion bash | sudo tee /opt/homebrew/etc/bash_completion.d/lazyagent > /dev/null

# zsh (persistent — pick any directory in $fpath)
lazyagent completion zsh > "${fpath[1]}/_lazyagent"

# fish
lazyagent completion fish > ~/.config/fish/completions/lazyagent.fish

Completions cover all subcommands (config / logs / shared / completion) and the global flags (--mock, --verbose/-v, --log-file, --log-format).

Troubleshooting

lazyagent writes a structured log to ~/.lazyagent/logs/lazyagent.log (rotated daily, kept for a week). Adapter parse failures, edit/copy/share/delete actions and startup metadata land there — nothing ever goes to stdout/stderr while the TUI is running, so the altscreen stays clean.

lazyagent --verbose                  # bump log level to debug for the next run
lazyagent logs path                  # print the resolved log file path
lazyagent logs tail                  # last 50 lines of the active log
lazyagent logs tail -n 200           # more lines
lazyagent logs clean                 # remove the active log + rotated siblings

Override location and format on a per-run basis:

lazyagent --log-file /tmp/lz.log --log-format json

For bug reports, please attach lazyagent logs tail output — the bug_report issue form has a dedicated field for it.

Roadmap

• Read-only TUI MVP across Claude / Codex / Gemini
• Unified place picker (p) — Origin × Scope matrix backed by a single library copy (PRI-65)
• Inline editor with mtime conflict detection (E)
• Shared canonical store + symlink projector + drift detection (s / R)
• Backlog: editing all tool-specific configs from one place (PRI-1)
• Backlog: GitHub install for skills / agents (PRI-3)
• Backlog: Global search across the whole tree (PRI-4)
• Backlog: Browse historical sessions across tools (PRI-5)
• Backlog: One-button sync of every shareable item (PRI-27)

Privacy

lazyagent runs entirely local. It reads files from your home directory and your project, and writes only when you trigger an explicit action (delete / place / resync / edit). No telemetry, no analytics.

The only outbound traffic is an optional weekly check of the GitHub releases API for an "↑ vX.Y.Z available" banner. Disable with [updates] notify = false in ~/.lazyagent/config.toml or --no-update-check on a single launch.

Filtering local projects

The global indexer (A hotkey) walks $HOME to find every directory that looks like a project root. To keep work / corporate / experimental trees out of that view, drop a gitignore-syntax file at ~/.lazyagent/ignore:

# work projects
~/work/
$HOME/Company/

# anything matching a glob
**/private-*

Manage it from the CLI without touching the file directly:

lazyagent ignore add '~/work/'   # append a pattern
lazyagent ignore list            # print the active rules
lazyagent ignore path            # print the file path

~/ and $HOME/ are expanded to your home directory at load time. Anything else passes through to standard gitignore semantics — negations (!), globs, and anchored paths all work.

Contributing

Bug reports and feature requests via GitHub Issues. PRs welcome — please open an issue first to discuss scope. Read VISION.md for what's in scope and what isn't, ARCHITECTURE.md for a codebase tour, and CONTRIBUTING.md for the workflow. Participation is governed by the Code of Conduct.

License

Apache-2.0 — see also NOTICE for attribution requirements when redistributing modified versions.