AI Extender for Claude

Tell Claude what you want. Use your own words, and AI Extender turns that into a working Claude skill or plugin. It plans the build, writes the files, checks them, and hands you something you can install, plus a companion prompt you can import into Prompt Builder to start it from a short form. You don't need to know what a manifest or an MCP server is.

Version 1.0.0 · Author @wasimness · Company SyncWin · Contact support@syncwin.com · License MIT

What you can build

• Skills that teach Claude a task the same way every time: a checklist, a house style, a report format.
• Plugins that bundle several skills with agents, hooks, or connectors.
• Connectors to the services your work already lives in (Notion, Slack, a CRM, your own API).
• Agents and hooks for narrower jobs, such as a reviewer that can only read files or a rule that blocks edits to .env.
• Internal business tools: your team's onboarding steps, proposal checks, or client reporting, packaged so everyone gets the same result.

Every plugin it builds comes with Composio already declared, so its skills can reach Gmail, Notion, Slack, a CRM, and 1,000+ other apps. Each person signs in with their own Composio account, so nothing personal is stored in the plugin. Plugins that read outside content also get a small injection guardrail. Say so in the plan if you don't want either.

It also updates, audits, versions, and packages extensions you already have. One summary of what changed becomes the changelog entry and every version number, each package comes with install and removal steps for the desktop app, the web, the Code tab, and the terminal, and a release report shows how many files were kept, changed, added, or dropped.

Quick start

1. Install the plugin (below).
2. Type /ai-extender, or just describe what you want: "make a skill that turns my meeting notes into action items."
3. Answer the few questions only you can answer, usually who will use it and where.
4. Approve the short plan. AI Extender builds it, checks it, and gives you the file, install steps, and its companion prompt.

More things to try:

• "My team forgets steps when we onboard a client. Can Claude help with that?"
• "Add a Notion connector to my plugin."
• "Audit my plugin before I share it."
• "Bump the version and package it for Cowork."

Install

Claude Code

claude plugin marketplace add syncwin/ai-extender
claude plugin install ai-extender@syncwin

Cowork or claude.ai: download the .plugin file from the latest release and upload it as a custom plugin. Once the plugin is listed in Anthropic's directory, you can add it from there instead.

Companion prompt

Every extension AI Extender builds comes with a companion prompt: a JSON file for Prompt Builder, a Chrome extension that keeps prompts as fill-in forms and inserts them into Claude. Import it under Settings > Import (help). Fill in the form, and the prompt starts the right skill with everything it needs. The source file lives in the extension's prompts/ folder; packaging delivers it as a separate file next to the package and never puts it inside the package, because it belongs to Prompt Builder, not to Claude.

AI Extender has one too: ai-extender-for-claude.json, attached to each release as its own file. Pick a goal, describe what you want, and send.

Requirements

Use Claude Code, Cowork, or claude.ai. Skills must be turned on. The bundled scripts need Python 3.8 or newer and nothing else; CI tests them on Python 3.9 through 3.13 on Linux. On Windows or macOS, run them with python or python3, whichever your system has.

What it runs, reads, and sends

AI Extender is Markdown instructions plus a few Python scripts, all readable in this repository.

• Scripts run on your machine. They make no network requests and read or write only the folders you point them at (plus a temporary folder while packaging). Packaging also writes the companion prompt into the extension's prompts/ folder and delivers it, with an install guide, next to the package (never inside it). They call only other scripts in this plugin, with fixed arguments and no shell.
• Injection screen: screen_text.py scores text with a 67 KB local model that ships in the plugin. It reads only the files or input you give it, downloads nothing, and sends nothing. AI Extender runs it on itself as a hook (hooks/hooks.json), and plugins it builds with connectors get the same hook: it warns Claude when web or connector output contains instructions aimed at the model.
• Guardrail hook (Claude Code and Cowork; chat doesn't run hooks): after each web fetch, web search, or Composio call, and no other tool, screen_text.py --hook reads that tool's output on your machine. If it looks like an injection, Claude gets a warning, treats the content as data, and asks you before acting on it. It never blocks your work and sends nothing; if Python can't start, Claude carries on without it. To turn it off, remove the PostToolUse entry from hooks/hooks.json.
• One exception, opt-in: validate_extension.py --cache keeps its last result in ~/.cache/ai-extender/. scan_secrets.py reads only the folders you name to look for passwords and keys saved in plain text. It masks what it finds and sends nothing.
• Folder checks: the router may run ls -d /mnt/user-data/outputs mnt/.plugins mnt/.local-plugins to tell whether it's inside Cowork. If a script path isn't filled in, it runs one find under ~/.claude/plugins, /mnt, mnt, and the current folder to locate its own scripts. Both read folder and file names only.
• Web pages: if you ask it to refresh its platform facts, Claude fetches Anthropic's official documentation pages with the web tool your session already has. It tells you before fetching.
• Composio connector: the plugin declares one connector, Composio, at its shared address https://connect.composio.dev/mcp. It ships no account and no key: you connect it yourself under Customize > Plugins > AI Extender > Connectors and sign in with your own Composio account. AI Extender uses it to push and release to GitHub when you ask, and to reach another app when a build needs one. Every skill still works without it; GitHub then falls back to your own GitHub connector or gh login. If you already have Composio connected at the same address, Claude uses that one and doesn't load a second copy.
• GitHub: only when you ask it to publish or keep a repo in sync, and only through Composio, your own GitHub connector, or gh login. It asks for the owner and repo, and confirms before the first push.

Nothing else leaves your machine. Apart from the Composio connector and the guardrail hook, the plugin has no MCP servers or executables of its own, and it collects no data. Claude asks before any step that deletes, sends, or publishes, and you can report a vulnerability privately by following the steps in SECURITY.md.

Skills and agents

Start with /ai-extender. It reads your request and calls the right role, so most people never call the others directly.

Skill	What it does
ai-extender	Router. Asks what you want, picks the steps
ai-extender:claude-planner	Scopes the request and designs the structure
ai-extender:claude-developer	Writes skills, connectors, agents, hooks, MCP servers, and tests
ai-extender:claude-maintainer	Updates existing extensions, versions, changelogs
ai-extender:claude-reviewer	Validates and audits before anything ships
ai-extender:claude-packager	Builds .plugin and .zip files, companion prompts, and marketplace entries

Agents: ai-extender:claude-grader grades test runs against their assertions, and ai-extender:claude-comparator compares two outputs blind.

Connector	Used by	For	Without it
Composio (https://connect.composio.dev/mcp)	ai-extender:claude-packager, and any role when a build needs an outside app	Pushing, tagging, and releasing on GitHub; reaching other apps a build depends on	Skills work as usual; GitHub goes through your own GitHub connector or gh

Naming. ai- marks parts that aren't tied to one platform; claude- marks Claude-specific parts. ChatGPT or Gemini layers could sit beside the Claude one later (skills/ai-extender/references/platforms.md). Only the Claude layer exists today. Because the role skills contain the word claude, claude.ai and the API reject them as individual skill uploads, so install the whole plugin instead.

Security and verification

Each release .plugin is built from its tagged commit on GitHub's runners and signed with a build-provenance attestation, and the release lists its SHA-256 checksums. To check a download before you install it:

gh attestation verify "AI.Extender.for.Claude.v1.0.0.plugin" --repo syncwin/ai-extender
sha256sum -c SHA256SUMS

Every change also runs CodeQL, a secret scan, the injection screen, and the full self-test suite, with GitHub Actions pinned to exact commits. Report a vulnerability privately as described in SECURITY.md. The plugin collects no data (PRIVACY.md).

Troubleshooting

• A skill doesn't trigger. Say what you want in plain words ("audit my plugin") or call /ai-extender directly. To check the descriptions offline: python skills/claude-reviewer/scripts/simulate.py . triggers.
• Validation fails. Run python skills/claude-reviewer/scripts/validate_extension.py <dir> --strict and fix the listed codes. Each message names the file and the rule.
• A skill upload is rejected on claude.ai. Skill names that contain claude can't be uploaded alone. Upload the .plugin file instead.
• A plugin's connector shows "Not added" though you already use that service. The plugin points at a different address than your own connector, so Claude treats them as two. Run the validator with --known-connectors "Name=https://your-connector-url" and use the address it reports. Claude Code skips a plugin connector only when the address matches exactly.
• A key or password ended up in a plugin. Run python skills/claude-reviewer/scripts/scan_secrets.py <plugin-folder>, move each value it finds into the plugin's settings, and rotate the key, since anyone who received a copy still has it.

Status

Version 1.0.0, released after AI Extender audited itself from a real install. Every push to main and every pull request runs the 117 self-tests, the directory validator, the secret scan, the injection screen, and the trigger simulation on Python 3.9 to 3.13, and claude plugin validate --strict passed before release. The evals/ folder holds five live cases for claude plugin eval; the last full run (2026-10-12, Sonnet as judge) scored 0.93, 0.73 above the same tasks without the plugin; six-run repeats scored 1.00 on audits, intake, and scaffolding. Run them yourself with claude plugin eval . --allow-tools Bash Write Edit --judge-model sonnet (the shell cases need the sandbox tools bubblewrap and socat on Linux; the default small judge misreads long, table-heavy replies). Platform facts were checked against Anthropic's documentation on 2026-10-11. Found a mismatch? Please open an issue.

Contributing and credits

Issues and pull requests are welcome; see CONTRIBUTING.md. Credits are in CREDITS.md. AI Extender is an independent project by SyncWin, and it is not affiliated with or endorsed by Anthropic in any way.