Plexus

Plexus Agent Config - one local dashboard for sharing rules, MCP servers, and skills across Claude Code, Cursor, Codex, Gemini CLI, Qwen Code, DeepSeek Harness, and OpenCode.

Import what you already have in Claude Code, Cursor, Codex, Gemini CLI, Qwen Code, DeepSeek Harness, OpenCode, and then share it with the other agents on your machine.

Latest Release · Homepage · 简体中文 · English

npx -y plexus-agent-config@latest start

Use-case guides: Guides · Claude Code / Cursor Config · Claude Code / Codex Skills + MCP · CLAUDE.md / AGENTS.md

Best fit for developers already using Claude Code plus Cursor, Codex, Gemini CLI, Qwen Code, DeepSeek Harness, or OpenCode and feeling config drift.

local-first · snapshots before native writes · partial-writes shared config files instead of replacing them

If Plexus saves you from maintaining the same agent config in five places, a GitHub star helps other multi-agent developers find it.

---

Why Plexus?

Modern AI coding work is multi-agent. One person might use Claude Code for planning, Cursor for editing, Codex for automation, then jump into Gemini CLI, Qwen Code, DeepSeek Harness, OpenCode, Windsurf, or Kiro for specific workflows. Each tool has its own config files, MCP format, skills folder, and instruction file.

That means every useful update turns into busywork:

• paste the same MCP server into multiple native config files
• keep CLAUDE.md and AGENTS.md in sync
• duplicate skills or prompts across agent-specific folders
• remember what changed when an agent breaks
• undo a bad sync without losing unrelated auth or history

Plexus gives those tools one local source of truth.

30-Second Pitch

Use Plexus if you run Claude Code plus Cursor, Codex, Gemini CLI, Qwen Code, DeepSeek Harness, or OpenCode and are tired of hand-editing the same config in five places.

Plexus treats ~/.config/plexus/ as the local canonical store, scans each agent's native config, and merges the Plexus store plus native-only entries as a union. You only choose a version when the same ID has different content; normal sync no longer asks you to choose a source. Native files are snapshotted before writes, so you can undo from the Backups page.

What It Does

Capability	What Plexus manages
Global Rules	One baseline in ~/.config/plexus/personal/rules/global.md, projected to CLAUDE.md and AGENTS.md
MCP Servers	Team + personal MCP servers synced into each agent's native format
Skills	Markdown skill bundles linked or copied into each agent's skill directory
Skill Market	Browse community skills ranked by GitHub stars and install them with one click
Mirror	Copy effective config from one agent to other agents with a preview
Backups	Snapshot native files before writes, then restore from the dashboard
Team Layer	Subscribe to a Git repo for team-approved MCPs and skills

Plexus does not run your MCP servers. It is a local dashboard for safely editing and projecting configuration.

The Sync Model In One Image

Plexus does not force every AI tool into one invented format. It keeps one canonical config on the local machine, then chooses the safest projection mechanism for each file type: symlinks for Rules and Skills, cache symlinks for exclusive MCP files, and partial writes for shared native config files.

You can open the same map from the local dashboard at /architecture/config-sharing-map.html.

The manual entry point for new tools is Settings -> Agent Catalog. If an AI agent is not built in yet, click Add agent to register its instruction file; Plexus can view, edit, and back it up now, while MCP / Skills adapters can be added later.

Quick Start

Requires Node 20.

Plexus ships as one npm package. The internal core and web dashboard are included, so users do not need to install multiple packages.

npm install -g plexus-agent-config
plexus

Open http://localhost:7777.

Or run it without a global install:

npx -y plexus-agent-config@latest start

You can also check which agents Plexus detects first:

plexus detect

Run from source

git clone https://github.com/miniLV/Plexus.git
cd Plexus
npm ci
npm run dev

Open http://localhost:7777.

On first run, click Share config everywhere in the dashboard:

1. Plexus detects installed agents and merges the Plexus store, native Rules, MCP servers, and Skills as a union.
2. It shows a smart-merge preview; only same-ID content conflicts need a version choice.
3. It applies config to enabled agents and snapshots native files before writing.

For a linked local CLI:

npm run link
plexus

To remove the linked CLI:

npm run unlink

Supported Agents

Agent	Rules target	MCP target	Skills target	MCP write mode
Claude Code	~/.claude/CLAUDE.md	~/.claude.json	~/.claude/skills/	partial write
Cursor	~/.cursor/AGENTS.md	~/.cursor/mcp.json	~/.cursor/skills/	symlink or copy
Codex	~/.codex/AGENTS.md	~/.codex/config.toml	~/.codex/skills/	partial write
Gemini CLI	~/.gemini/GEMINI.md	~/.gemini/settings.json	~/.gemini/skills/	partial write
Qwen Code	~/.qwen/QWEN.md	~/.qwen/settings.json	~/.qwen/skills/	partial write
Factory Droid	~/.factory/AGENTS.md	~/.factory/mcp.json	~/.factory/skills/	symlink or copy
DeepSeek Harness	~/.dsh/AGENTS.md	~/.dsh/mcp.json	~/.dsh/skills/	partial write
OpenCode	~/.config/opencode/AGENTS.md	~/.config/opencode/opencode.json	~/.config/opencode/skills/	partial write

Partial write means Plexus rewrites only the MCP section and preserves the agent-owned auth, history, profile, and settings data in the same file.

Settings also includes an Agent Catalog for common tools such as Windsurf, Kiro, VS Code Copilot, Cline, Roo Code, Kilo Code, Continue, Aider, Amp, OpenHands, and Zed AI. Tools without a native Plexus adapter are shown as manual presets so users can register a custom instruction file quickly.

The manual entry point is Settings -> Agent Catalog -> Add agent. Use it when a new tool is not in the built-in list yet, or when your local path differs from the preset.

How It Works

Plexus stores canonical config under ~/.config/plexus/:

~/.config/plexus/
├── config.yaml
├── team/
├── personal/
│   ├── mcp/servers.yaml
│   ├── rules/global.md
│   └── skills/<id>/SKILL.md
├── .cache/mcp/
└── backups/

The team/ layer is intended to come from a shared Git repo. The personal/ layer belongs to the local user and overrides team entries with the same ID.

For single-purpose native MCP files such as Cursor and Factory Droid, Plexus uses symlinks when possible. For shared native files such as ~/.claude.json, ~/.codex/config.toml, ~/.gemini/settings.json, ~/.qwen/settings.json, ~/.dsh/mcp.json, and ~/.config/opencode/opencode.json, Plexus partial-writes only the MCP section.

Team Starter Repo

You can use miniLV/agent-primer as a first team baseline. It includes:

• rules/global.md: shared agent behavior rules
• skills/<id>/SKILL.md: reusable team skills
• mcp/servers.yaml: team-approved MCP servers

plexus join https://github.com/miniLV/agent-primer.git
plexus pull
plexus sync

You can also do this from the dashboard: open Team, click Use agent-primer, then click Join. If the repo is private, make sure your local GitHub credential can access it, or use the SSH URL: git@github.com:miniLV/agent-primer.git.

This clones agent-primer into ~/.config/plexus/team/. Plexus then treats it as the team layer and overlays your local ~/.config/plexus/personal/ layer:

agent-primer repo
  rules/global.md        -> Plexus Team Rules
  skills/*/SKILL.md      -> Plexus Team Skills
  mcp/servers.yaml       -> Plexus Team MCP

local machine
  ~/.config/plexus/team/     # git clone of agent-primer
  ~/.config/plexus/personal/ # your overrides, secrets, local paths

Example: your team adds the code-review skill to agent-primer. Every member runs plexus pull && plexus sync, and that skill is projected into Claude Code, Cursor, Codex, and other enabled agents. The team can keep the github MCP as an optional placeholder; each person enables it from the Plexus MCP page and adds their own token in the personal layer.

Note: MCP servers that depend on personal tokens, local paths, or database connection strings should stay as placeholders / optional entries in the team repo, then be completed or overridden from the user's local ~/.config/plexus/personal/ layer.

Skill Market

Open Skill Market to browse community agent skills from GitHub, ranked by stars, with free-text search and pagination (up to 5 pages). Click Install on a skill and Plexus will:

1. Download the repo tarball and extract it to a temp directory.
2. Locate SKILL.md (preferring the repo root) and read its frontmatter.
3. Write SKILL.md plus sibling resources (scripts/, references/, assets/, …) into ~/.config/plexus/personal/skills/<repo>/.
4. Snapshot native files, then sync to all enabled agents.

One install works across every supported agent — Claude Code, Cursor, Codex, Gemini CLI, Qwen Code, Factory Droid, DeepSeek Harness, and OpenCode.

Already-installed skills show Installed. Repos that are not single-skill (no SKILL.md, or multiple skills) are rejected with a clear message rather than writing broken config.

Market data comes from the GitHub Search API, sorted by stars descending. It is rate-limited without a token; set GITHUB_TOKEN (or keep gh signed in) to raise the limit automatically.

CLI

plexus              start the dashboard
plexus start -p 7777
plexus detect       list detected agents
plexus join <url>   clone a team config repo into ~/.config/plexus/team
plexus pull         pull the configured team repo
plexus sync         union Plexus + native config and apply to enabled agents
plexus sync --prefer codex   use Codex only to resolve native-native conflicts
plexus status       show team subscription and sync status
plexus help

Safety Model

• Plexus is local-first.
• Plexus does not execute MCP servers.
• Native files are snapshotted before writes.
• Shared native config files use partial writes: only the Plexus-managed MCP section changes.
• Dedicated MCP files use symlink/copy mode, with the previous file quarantined first.
• Debug snapshots return metadata only, not file contents.
• Imported MCP env values are stored as plaintext in the local personal store.
• Do not push ~/.config/plexus/personal/ to a shared team repo without reviewing and redacting secrets.

Development

npm ci
npm run verify

Focused commands:

npm run check
npm run test:core
npm run build --workspace=plexus-agent-config-core
npm run build --workspace=plexus-agent-config-web

Current Status

Plexus is alpha software. The local workflows are usable on macOS and Linux, but there are still sharp edges:

• project-scoped MCP files are not managed yet
• dashboard PR proposals for team config are not built yet
• custom agents are instruction-file registry records only
• Rules apply currently targets built-in agents only
• Windows support is unverified

License

Apache-2.0