Are AI agent plugins just prompt packs?

No. A prompt pack may be one ingredient, but a plugin should carry ownership, components, permissions, versioning, telemetry, and release evidence.

Do AI agent plugins require MCP?

No. MCP is useful for tool and context integration, but a plugin can also contain Skills, hooks, commands, assets, or harness-specific config.

How do you measure an AI agent plugin?

Track component invocation, errors, latency, outcomes, feedback, eval results, and release comparisons using metadata-only events.

Guide

AI Agent Plugins: What They Are, How They Work, and How to Ship Them

An AI agent plugin is an installable capability package. It lets an agent do useful work with instructions, tools, system access, and enough telemetry for the owner to improve it.

Start with Telvine Read the plugin guide

Publish the product

Use Telvine when the plugin is ready to be treated like software.

The first useful milestone is simple: one plugin directory, one owner, one release record, and one measurement model.

npm i -g @telvine/cli
telvine login
telvine publish ./my-plugin

The practical definition

An AI agent plugin bundles the pieces an agent needs to perform a task: task instructions, callable tools, connectors, runtime config, hooks, assets, and sometimes one or more agents. The exact folder shape changes by harness. The product shape should not.

That distinction matters. Teams do not want seven different products because seven agent surfaces exist. They want one managed plugin with adapters for the places where work happens.

What belongs inside

A plugin can start small. A single SKILL.md capability may be enough for a useful first version. As the work gets more serious, the plugin usually gains connectors, MCP config, scripts, test assets, evals, and release notes.

Telvine treats those pieces as components of one product, then records metadata-only events as the plugin is invoked, fails, recovers, and improves.

Skills for task instructions and capability boundaries.
Connectors for approved access to customer systems.
MCP servers and tools for callable actions.
Hooks and commands for runtime behavior.
Eval suites and telemetry for release evidence.

How to publish a first plugin

Start with a workflow that is repeated, rule-heavy, and valuable enough to justify a maintained product. Then package the workflow as a plugin, add at least one observable capability, and publish it with Telvine.

The first release does not need to be broad. It needs to be measurable.

Where it applies

Use cases that deserve a maintained plugin

Finance operations

Month-end checks, invoice collection review, VAT categorisation readiness, and cashflow exception handling.

Support operations

Ticket triage, escalation prep, customer context gathering, and resolution quality checks.

Sales operations

Account research, quote readiness, CRM hygiene, and renewal risk review.

FAQ

Common questions

Are AI agent plugins just prompt packs?: No. A prompt pack may be one ingredient, but a plugin should carry ownership, components, permissions, versioning, telemetry, and release evidence.
Do AI agent plugins require MCP?: No. MCP is useful for tool and context integration, but a plugin can also contain Skills, hooks, commands, assets, or harness-specific config.
How do you measure an AI agent plugin?: Track component invocation, errors, latency, outcomes, feedback, eval results, and release comparisons using metadata-only events.