Claude Code Plugin

The SWEny plugin for Claude Code gives you slash commands, MCP tools, a startup hook, and an isolated agent — all from a single install. Instead of switching to your terminal, type /sweny:triage and Claude delegates the work to SWEny’s DAG executor.

Install

First add the SWEny marketplace, then install the plugin:

/plugin marketplace add swenyai/sweny
/plugin install sweny@sweny-official

Skills

Skill	Description
/sweny:triage	Investigate production alerts, create issues, implement fixes
/sweny:implement	Pick up an issue and implement a fix
/sweny:e2e-init	Generate browser E2E test workflows
/sweny:e2e-run	Run E2E test workflows
/sweny:workflow-create	Generate a workflow from natural language
/sweny:workflow-edit	Edit a workflow with natural language
/sweny:workflow-run	Run any workflow from a YAML file
/sweny:workflow-diagram	Visualize a workflow as a Mermaid diagram
/sweny:init	Set up SWEny in the current project
/sweny:check	Verify provider credentials
/sweny:setup	Create labels in your issue tracker

MCP tools

The plugin also registers an MCP server that gives Claude two tools for autonomous use — Claude can list and run workflows without you typing a slash command:

• sweny_list_workflows — list available workflows
• sweny_run_workflow — execute triage or implement programmatically

Hooks

• SessionStart — if a .sweny.yml exists, runs sweny check on startup to surface credential issues early

Agent

• sweny-workflow — long-running workflows (triage, implement, e2e-run, workflow-run) execute in a forked context using this agent, keeping your main conversation clean

Manual MCP setup

If you prefer to configure just the MCP server without the full plugin:

Why use SWEny as an MCP server?

Claude Code runs everything in a single conversation context. For complex tasks, that context gets bloated and unfocused. SWEny’s DAG executor solves this by splitting work into nodes — each node gets a focused context window with only the inputs it needs.

Task type	Best approach
Simple, single-step tasks	Claude Code handles directly
Complex multi-step workflows	Claude delegates to SWEny via MCP

SWEny handles the structured orchestration — conditional routing, scoped tool access, structured output schemas — that a single conversation can’t do reliably.

Setup

Claude Code

Add to .claude/settings.json in your project:

{
  "mcpServers": {
    "sweny": {
      "command": "npx",
      "args": ["-y", "@sweny-ai/mcp"]
    }
  }
}

Claude Desktop

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "sweny": {
      "command": "npx",
      "args": ["-y", "@sweny-ai/mcp"]
    }
  }
}

Prerequisites

• A .sweny.yml config file in the working directory (run sweny new to create one)
• Credentials for your configured providers set via environment variables or .env
• @sweny-ai/core installed (provides the sweny CLI binary)

Available tools

sweny_list_workflows

Lists built-in workflows (triage, implement, seed-content) and any custom workflows found in .sweny/workflows/*.yml.

Parameter	Type	Required	Description
cwd	string	no	Working directory to search. Defaults to the server’s working directory.

Example response:

[
  {
    "id": "triage",
    "name": "Triage",
    "description": "Investigate alerts and create issues or PRs",
    "nodeCount": 8,
    "source": "builtin"
  },
  {
    "id": "implement",
    "name": "Implement",
    "description": "Pick up an issue and implement a fix",
    "nodeCount": 5,
    "source": "builtin"
  }
]

sweny_run_workflow

Executes a triage or implement workflow by spawning the sweny CLI. Returns structured JSON results when the workflow completes.

Parameter	Type	Required	Description
workflow	"triage" or "implement"	yes	Which workflow to run
input	string	for implement	Issue ID or URL (required for implement, ignored for triage)
cwd	string	no	Working directory containing .sweny.yml
dryRun	boolean	no	Skip side effects — investigate but don’t create issues/PRs

Execution time

Workflows run a full DAG with multiple Claude invocations under the hood. Expect triage to take 2-5 minutes and implement to take 3-10 minutes depending on complexity. The MCP server enforces a 10-minute timeout.

How it works

When Claude calls sweny_run_workflow, the MCP server:

1. Spawns a separate sweny CLI process — this avoids recursion (the MCP server runs inside Claude Code, but the workflow spawns its own Claude Code instance)
2. Passes --json --stream — the CLI outputs NDJSON events as the DAG executes
3. Streams progress in real time — parses NDJSON events (node:enter, node:progress, node:exit) and forwards them to Claude via MCP logging notifications
4. Returns the final result — the last JSON object from the stream, with success/failure status

The MCP server itself is stateless — each tool call is independent.

Relationship to MCP server auto-injection

SWEny has two MCP integration points — don’t confuse them:

Concept	What it does	Docs
SWEny as MCP server (this page)	Exposes SWEny workflows as tools for Claude Code/Desktop
MCP server auto-injection	SWEny injects external MCP servers (GitHub, Linear, etc.) into its own workflows	MCP Servers

They complement each other: Claude Code calls SWEny via MCP, and SWEny’s workflows internally use other MCP servers to access external services.