| --- |
| title: Meeseeks |
| emoji: ⚖️ |
| colorFrom: indigo |
| colorTo: blue |
| sdk: docker |
| app_port: 7860 |
| pinned: true |
| license: mit |
| short_description: Author skills once, run on any agent |
| --- |
| |
| # Meeseeks |
|
|
| Meeseeks turns repeatable computer workflows into structured, **agent-agnostic** skills. The same JSON skill description drops into Claude, GPT, Gemini, or any local model that speaks the open Anthropic Agent Skills format — no translation layer. |
|
|
| This Space is a live demo of the authoring + grounding pipeline: |
|
|
| - **Author** — hand-edit a skill JSON, see the placeholder plan substitute live |
| - **Bundle** — derive `SKILL.md`, `manifest.json`, and `actions.toon` (token-oriented action format) from one source of truth |
| - **Receipts** — see the executor's screenshot trail when the skill walks through a real DOM (Playwright + accessibility-tree grounding) |
|
|
| The full project (CLI, runtime, mock paralegal target, sample skills) lives at the root of the repo. |
|
|
| ## Why this matters |
|
|
| Today, every agent platform reinvents skill capture: ChatGPT custom GPTs, Claude Skills, GLaDOS plugins, OpenAI Operator macros — all incompatible. Meeseeks is the smallest possible spec that keeps a skill portable across all of them while preserving the grounding metadata (bbox, accessibility role/name, success checks, fallbacks) that makes a skill safe to run. |
|
|
| ## Domain |
|
|
| The current sample set is paralegal — discovery objections, meet-and-confer letters — because it's a high-value workflow domain where redacted templates + verified grounding genuinely save hours. The same shape applies to healthcare intake, financial back-office, and code-review workflows. |
|
|
