Mushrooms are just the fruit; the real organism is the mycelium underground — a living network that connects an entire forest and lets every tree share nutrients and signals. Your AI coding agents are that forest: Claude Code, WorkBuddy, Codex CLI, Cursor, Antigravity — each powerful, each completely walled off from the others.
Myco is the mycelium. A native menu-bar / tray app (macOS & Windows) that quietly connects your agents so they can share what they know:
- 🟢 Share skills — write one
SKILL.mdand fan it out to every agent's repo directory, so a skill you teach one tool works in all of them. - 🔵 Hand off conversations — pick up a chat from one agent and continue it in another, in a legitimately new session (no forged IDs, no fake history).
- 🟣 Unify history — read every agent's local transcripts into one neutral, searchable, offline timeline you can browse and back up.
All of it from the menu bar. No command line, nothing to configure.
Install
Windows
Two options on the latest release:
Myco-Setup-x.y.z.exe(recommended) — a standard install wizard: per-user (no admin), Start-Menu shortcut, uninstall via Settings → Apps.Myco-win-x.y.z.zip(portable) — unzip anywhere, double-clickMyco.exe, delete the folder to remove.
Either way, Myco lives in the system tray (bottom-right stacked-tiles icon). SmartScreen may warn on first run (unsigned): More info → Run anyway, once.
Requirements: Windows 10/11. Fully self-contained — .NET runtime and an embedded Python are bundled; nothing else to install.
macOS
- Download
Myco-x.y.z.dmgfrom the latest release (built on GitHub's macOS runners, Apple Silicon). - Open the DMG and drag
Myco.appinto Applications. - First launch: because the app is ad-hoc signed (not notarized), macOS
Gatekeeper will hesitate. Right-click
Myco.app→ Open → Open once, and you're set.
Myco then lives in your menu bar (the stacked-tiles icon at the top-right).
Requirements: macOS 13+. That's it — Myco is fully self-contained and uses the
python3that already ships with macOS. Nothing else to install.
📖 Illustrated user guide (install, every tab explained, FAQ): docs/USER-GUIDE.md · 中文版
What's inside the app
Myco opens to five tabs, each a capability rather than a separate tool:
| Tab | What it does |
|---|---|
| 总览 / Home | At-a-glance view of which agents are installed on this Mac and how many sessions each holds. |
| 共享 / Share | Pick a skill, choose which agents to fan it out to (.claude / .workbuddy / .codex / .agents …), preview, then write. Commit and the whole team is in sync. |
| 接力 / Relay | Pick a past conversation and package it as paste-ready text to continue in a different agent. |
| 历史 / History | The merged, searchable timeline across all detected agents. |
| 设置 / Settings | Theme, agent toggles, work directory. |
Everything is read-only by design: Myco never writes back into any agent's storage. Agent databases (Cursor / Antigravity SQLite) are opened in immutable read-only mode; hand-off and archive operations only ever produce text.
Why this problem exists (the 30-second version)
If you run more than one AI coding tool on the same project, you've hit both walls:
Skills are siloed. There is no shared skills directory across agents. "Install once, everything sees it" is literally false — each product reads skills from a different repo directory:
| Agent | Repo-level dir (travels with Git) | How you invoke it |
|---|---|---|
| Claude Code | .claude/skills/ |
mention the skill by name (some suites add /slash) |
| WorkBuddy | .workbuddy/skills/ |
mention the skill by name |
| Codex CLI | .codex/skills/ and/or .agents/skills/ |
$skill-name, /skills, or name it — not /design |
| Cursor | .cursor/skills/ (also reads .cursor/rules/ .mdc) |
rules auto-inject; also tolerates .agents/ |
| Antigravity | .antigravity/skills/ (VS Code–based, convention) |
mention the skill by name |
.agents/skills/is the emerging cross-agent path (Codex, Gemini CLI and Cursor all read it; Cline tolerates.claude/skills/). When you can maintain only one extra path,.agents/skills/is the safest bet — which is why Myco offers it as a distribution target alongside the five above.
These are the five agents Myco detects and connects on your Mac. Skill
paths across all these tools drift fast; Myco keeps the mapping in one place
(engine/agents.json) so you never have to.
Conversations are siloed too. Each tool keeps its transcripts in its own place and format — JSONL here (Claude / WorkBuddy / Codex), SQLite blobs there (Cursor / Antigravity) — so context you built up in one agent is invisible to the next.
Myco knows all of these locations and formats, and bridges them for you. That's the whole point: you shouldn't have to memorize this table — the app does.
For contributors — the engine under the hood
Myco's UI is a thin SwiftUI shell. The actual work is done by a small,
pure-Python-stdlib engine (in engine/) that the app calls via
Process. You normally never see it — but if you want to hack on Myco, build it
from source, or script these capabilities headlessly, it's all there:
| Engine module | Powers |
|---|---|
engine/distribute.py |
skill fan-out (Share tab) |
engine/sync_chats.py |
history aggregation (History tab) |
engine/handoff_chat.py |
conversation hand-off (Relay tab) |
engine/chatsync/ |
canonical message model + per-agent readers + exporters |
Build the app from source (Command Line Tools only, no Xcode):
cd app
./build.sh # swift build -c release → assembles self-contained Myco.app
./package_dmg.sh # (optional) produce a distributable .dmg
open Myco.app
Architecture, env-var switches, and source layout: app/README.md.
The portable skill this app ships and distributes lives in
skills/multi-agent-skill-sharing/SKILL.md;
its design notes are in docs/.
Repository layout
multi-agent-skill-sharing/ (the Myco project)
├── README.md
├── LICENSE
├── app/ # Myco — the SwiftUI menu-bar app (macOS)
│ ├── Sources/Myco/ # native UI + PythonBridge
│ ├── build.sh # assemble self-contained Myco.app
│ └── package_dmg.sh # produce the installable .dmg
├── app-windows/ # Myco — the WPF tray app (Windows)
│ ├── *.cs / Views/ # native UI + PythonBridge (0 NuGet deps)
│ └── build.ps1 # assemble self-contained Myco-win zip
├── engine/ # Myco's internal Python engine (pure stdlib)
│ ├── distribute.py # skill fan-out
│ ├── sync_chats.py # history aggregation → archive + HTML
│ ├── handoff_chat.py # package one chat for a legit hand-off
│ └── chatsync/ # canonical model + readers + exporters
├── skills/ # the SKILL.md payload Myco ships & distributes
│ └── multi-agent-skill-sharing/SKILL.md
├── prototype/ # high-fidelity interactive HTML prototype
├── assets/ # brand: logo, wordmark, palette
└── docs/ # install notes + design docs
Related projects
Myco is about connecting the agents you already run, not cataloging skills. If you're looking for large catalogs of ready-made skills, these are excellent:
| Project | Stars | What it is |
|---|---|---|
| VoltAgent/awesome-agent-skills | 20k+ | Cross-agent catalog (Claude, Codex, Gemini, Cursor) — the biggest curated list |
| openai/skills | 9k+ | OpenAI's official Codex skills directory |
| vercel-labs/skills | 6k+ | Vercel's official skills + CLI tooling |
| anthropics/skills | — | Anthropic's official skills for Claude Code |
| agentskills/agentskills | 10k+ | The open SKILL.md specification / standard |
Those tell you what skills exist. Myco connects the tools you actually run so a skill — or a conversation — can move freely between them.
Caveat
Skill-discovery conventions across these tools change quickly. The paths Myco uses were verified 2026-07. If one doesn't resolve, check the tool's own docs — and PRs to keep them current are very welcome.
Contributing
Found a new agent, a changed path, or a better invocation trick? Path updates are the most valuable contribution here — see CONTRIBUTING.md for what to include (tool version, OS, how you verified) and a quick local check.
License
MIT — see LICENSE.
No comments yet
Be the first to share your take.