Skills RCE

Skills RCE is the code execution service that runs skill scripts. When an agent activates a skill with executable scripts, the runtime sends the code to Skills RCE for sandboxed execution.

How It Works

sequenceDiagram
    participant A as Agent
    participant R as Runtime
    participant RCE as Skills RCE

    R->>RCE: Upload skill (zip, on startup)
    RCE-->>R: Cached

    A->>R: run_skill("python scripts/generate.py")
    R->>RCE: POST /skill/{id}/run
    RCE->>RCE: Execute in subprocess
    RCE-->>R: stdout, artifacts
    R-->>A: Results

sequenceDiagram
    participant A as Agent
    participant R as Runtime
    participant RCE as Skills RCE

    R->>RCE: Upload skill (zip, on startup)
    RCE-->>R: Cached

    A->>R: run_skill("python scripts/generate.py")
    R->>RCE: POST /skill/{id}/run
    RCE->>RCE: Execute in subprocess
    RCE-->>R: stdout, artifacts
    R-->>A: Results

1. Startup: Runtime zips skill directories and uploads them to RCE (hash-based, skips unchanged skills)
2. Execution: Agent calls run_skill, runtime sends the command to RCE
3. Response: RCE returns stdout, stderr, exit code, and any generated files as base64 artifacts

Two Ways to Use RCE

Built-in (server-managed)

When running formations via MUXI Server, Skills RCE is included. Formations use it automatically with no configuration needed.

Custom instance

Run your own RCE service and point your formation at it:

# formation.afs
rce:
  url: "http://localhost:7891"
  token: "${{ secrets.RCE_TOKEN }}"

This is useful when you need:

• A specific set of packages or runtimes
• Dedicated resources for heavy workloads
• Network isolation or custom security policies

Running Your Own Instance

Docker

docker run -d -p 7891:7891 \
  -e RCE_AUTH_TOKEN=my-secret \
  ghcr.io/muxi-ai/skills-rce:latest

SIF (Linux)

Download the .sif file from the Releases page:

apptainer run skills-rce.sif

From source

cd src && go build -o skills-rce ./cmd/rce
RCE_AUTH_TOKEN=my-secret ./skills-rce

Configuration

All via environment variables:

Authentication

Set RCE_AUTH_TOKEN to require a bearer token on all endpoints except /health and /status:

docker run -d -p 7891:7891 -e RCE_AUTH_TOKEN=my-secret ghcr.io/muxi-ai/skills-rce:latest

When configured, all requests must include Authorization: Bearer .

Available Runtimes

The Docker image bundles runtimes commonly used by agent skills:

Python packages

Data & analysis: numpy, pandas, scipy, scikit-learn, statsmodels, sympy

Visualization: matplotlib, seaborn, plotly, bokeh, altair

Documents: reportlab, fpdf2, python-docx, openpyxl, python-pptx, xlsxwriter

Images: pillow, pytesseract, pdf2image, qrcode

HTTP: requests, httpx

General: pyyaml, jinja2, tabulate, orjson

JS/TS packages

lodash, axios, cheerio, sharp, csv-parse, date-fns, zod, marked, uuid, yaml

System tools

curl, wget, git, ffmpeg, imagemagick, poppler-utils, tesseract-ocr

API

See the full OpenAPI spec for request/response schemas.

Security

• Each job runs in an isolated subprocess with resource limits
• Cached skill directories are read-only during execution
• Working directories cleaned up after each job
• Output truncated at 100KB
• Zip uploads validated against path traversal attacks
• No host filesystem access beyond mounted volumes

Troubleshooting

curl http://localhost:7891/health

curl http://localhost:7891/status | jq '.packages'

Next Steps

Skills Concepts - How skills work
Skills Reference - SKILL.md syntax and config
Add Skills Guide - Step-by-step tutorial
GitHub: skills-rce - Source and releases