Instructions and browser tools solve different jobs.
A skill can explain the procedure. An AgentScript can supply a page action that the procedure calls.
By Customaise · Reviewed
Compares Claude Code skills and host permission controls with Customaise AgentScripts. Other skill hosts have different policies; neither mechanism is a universal sandbox. Anthropic donated the Agent Skills standard to the Linux Foundation’s Agentic AI Foundation on 22 October 2026; more than 70 agent platforms have integrated it.
Guide: Reusable instructions
Skills package guidance and supporting resources for an agent to use when relevant.
Act: Registered browser tools
AgentScripts define named operations for installed, matching pages.
Review: Layered controls
Host permissions, tool declarations and source review address different risks.
Side-by-side comparison
| Capability | Customaise | Claude Skills |
|---|---|---|
| What it contains [5] [1] [9] | JavaScript and metadata for a browser script, including declared agent tools. | Instructions, optional scripts and supporting files that guide the agent. |
| When it runs [5] [1] | An enabled script loads on matching pages; the agent discovers and calls registered tools. | The agent or user invokes a skill under the host’s loading and invocation rules. |
| Reusable behaviour [5] [1] | Named browser operations with inputs and script-defined results. | Reusable procedures that can invoke code, tools and other resources. |
| Permissions [5] [1] [2] | Declared Allow, Prompt and Deny tools combine with approval mode and saved decisions. A script an agent creates over MCP or the CLI cannot grant its own tools Allow, and an agent cannot switch on Full access. | A skill’s allowed-tools pre-approves the tools it lists for that turn; it does not restrict others. Claude Code’s own permission settings still apply. |
| Prompt injection [5] [8] [1] [2] | Untrusted page text, tool descriptions, results and source can influence an agent. Review what is exposed. | Untrusted instructions and executable resources can influence work. Use host controls and review sources. |
| Neither side is immune to malicious code or instructions. JS comments not executing as code does not eliminate an agent’s prompt-injection risk. | ||
| Source review [4] [5] [1] [3] | Inspect grants, network access and tools. Community Marketplace review is an additional layer, not a safety guarantee. | Inspect the skill and its dependencies, including scripts and any dynamic context it loads. |
| Review boundaries [5] [1] [3] | Marketplace review of a published version does not cover every imported file or Send copy. | Review and distribution practices vary by repository and host; do not generalise one marketplace’s research to all skills. |
| Browser access [5] [1] | MCP or CLI connects the agent to the extension’s installed tools. | A skill needs an available browser tool or integration to perform browser actions. |
| AI inside execution [6] [1] [2] [10] | Free on-device AI with CM_promptAI (Gemini Nano, no key, Chrome 148 or later). Power User adds Script Decisions with Jev, OpenAI Decisions API or Gemini Decisions API. | A skill guides model behaviour and can use whatever tools its host permits. |
| Sharing [5] [7] [1] | Publish subscriptions or Send independent copies with Power User; recipients control enablement. | Distribute skill files through the host’s supported project, personal or plugin mechanisms. |
| Maintenance [5] [1] | Update and retest script assumptions as pages and services change. | Update instructions, dependencies and tool assumptions as the workflow changes. |
What is the real difference?
Skills and AgentScripts are complementary abstractions. Skills organise instructions and may include scripts or other resources. AgentScripts execute in a browser context and expose declared tools. A skill can teach an agent how to use those tools through MCP or CLI.
Claude Code has host permission policies and skill controls. Customaise enforces approval per tool when each one is called. As with a skill, read a script’s code before you enable it.
Declared tools are one control layer
Declared tools use Allow, Prompt or Deny. The selected approval mode and saved decisions affect whether a Prompt tool asks again; Full access permits eligible calls. Review source and permissions: tool-call controls do not isolate arbitrary script code or an agent with access to the whole computer. [5]
Review executable resources
Read the code and scripts a skill can invoke, not just its introductory prose. Datadog’s dynamic-context research describes a specific execution path and mitigation; it is not evidence that all skills bypass every host permission setting. [1] [2] [3]
Keep distribution boundaries clear
Power User can publish subscriptions or create a 24-hour Send link to an independent editable copy. A signed-in Free user can receive a Send copy. Copies arrive disabled; expiry stops link access but does not recall copies already received. Send copies do not receive publisher updates. [5] [7]
AI in your scripts
CM_promptAI runs Gemini Nano on the device, free and with no key, in Chrome 148 or later. Power User adds Script Decisions: focused classification or scoring with Jev, OpenAI Decisions API or Gemini Decisions API, declared in the script with @grant CM_decisions and run with a saved provider key on an authorised device. Provider charges are separate. [6] [10]
When should you use which?
Customaise
You need persistent page controls or named browser operations that people can install, enable and inspect.
Claude Skills
You want to teach Claude a procedure, conventions or domain knowledge that can use the host’s existing tools.
A skill can describe how to operate Customaise through MCP or CLI. The installed AgentScripts supply the browser actions; both remain subject to their respective controls.
Sources and scope
Compares Claude Code skills and host permission controls with Customaise AgentScripts. Other skill hosts have different policies; neither mechanism is a universal sandbox. Anthropic donated the Agent Skills standard to the Linux Foundation’s Agentic AI Foundation on 22 October 2026; more than 70 agent platforms have integrated it.
By Customaise. Product capabilities are based on documentation and our implementation. Recommendations are workflow judgments, not a head-to-head performance or security benchmark.
- [1] Anthropic: Claude Code skills
- [2] Anthropic: Claude Code permissions
- [3] Datadog: dynamic-context research and mitigation
- [4] Customaise: UserScripts
- [5] Customaise: AgentScripts and permissions
- [6] Customaise: Script Decisions
- [7] Customaise: plans and allowances
- [8] Customaise: privacy and data flows
- [9] Anthropic: Agent Skills donated to the Agentic AI Foundation
- [10] Customaise: on-device AI and your own models
Related comparisons
- Customaise vs Tampermonkey: Compare a userscript manager with Customaise’s integrated workflow for building, testing and reusing UserScripts and AgentScripts.
- Customaise vs Playwright: Playwright automates browsers from code. Customaise adds named tools and page controls to the Chrome you already use.
- Customaise vs Browser Use: Browser Use offers local and cloud paths, including Browser Harness. Customaise centres the workflow on an installed library of scripts and tools.