MINA LABS
Start creating free
GitHub has released Agent Plugins 1.0 for VS Code, Copilot CLI, and the Copilot app, letting a single plugin package run across all compatible agent clients. The specification stabilizes interfaces and manifest rules so tools can declare capabilities once and be invoked consistently. Details and timelines are posted in the changelog.
The release was coordinated with AWS, Anysphere, Microsoft, OpenAI, and Vercel, and the plugin format is now treated as a stable contract rather than an experiment. Clients agree on how tools describe inputs, outputs, and execution boundaries, which reduces the drift between editing, terminal, and conversational surfaces. Early adopters can migrate existing integrations without rewriting core logic, while new plugins can target the full ecosystem from day one.
For someone building software, this changes how tools are bundled and maintained. Instead of maintaining separate extensions for the editor, the command line, and the assistant app, you ship one package and let each client enforce its own policy surface. Dependencies are declared up front, permissions are scoped, and failures are surfaced in a standard shape, which makes it easier to reason about what a tool can do when it runs.
The manifest requires explicit capability claims and environment constraints, so you stop guessing whether a plugin will work in a headless session or a full desktop context. Inputs and outputs follow a strict schema, which means generated plans can be validated before execution and results can be consumed by downstream steps without custom parsing. This is useful when you chain tools that touch files, shells, or language models, because the contract reduces the amount of brittle glue code you write to normalize responses.
Reliability improves because the runtime can reject mismatched invocations instead of failing mid-task. You can version plugins independently of the clients, roll back a single package, and keep workflows stable across updates. For teams that build internal tooling, this lowers the cost of keeping assistants useful as languages, frameworks, and cloud environments evolve.
What it costs and whether it is on Mina Labs is the part you can use today. The plugin specification itself is free to implement, but image generation through the Mina Labs pipeline is priced at 2 per image for MAI Image 2.5 Pro (Microsoft). This pricing applies when a plugin triggers generation via the Mina Labs model, and it is available here without additional subscriptions or seat fees.
We would use this to stitch together editor actions, terminal commands, and conversational planning without rewriting the same integration three times. A plugin that formats and runs database migrations could be invoked from a Copilot prompt, validated in VS Code, and rerun from the CLI with identical inputs and outputs. Another plugin could generate and insert diagrams into documentation by calling MAI Image 2.5 Pro (Microsoft) at 2 per image, then commit the assets and update links in one step.
We would also use it to enforce policy at the boundary between agents and production systems. By declaring environment constraints and required approvals in the manifest, a plugin can refuse to run in unapproved regions or without explicit user consent, and every client respects that refusal in the same way. This makes it safer to delegate work to agents without losing oversight of what touches state, secrets, or build artifacts.
Over time, we would treat plugins as durable building blocks rather than disposable scripts. When a new language or cloud service appears, we update the plugin once and all clients gain the capability. When a tool needs stricter isolation, we tighten the manifest and let the runtime enforce it everywhere. The result is less maintenance, clearer expectations, and fewer surprises when an agent tries to act on our behalf.