Content
<p align="center">
<img src="https://raw.githubusercontent.com/echoo19/hive-mcp/main/.github/banner.svg" alt="hive" width="640">
</p>
<p align="center"><b>The context-aware package manager for agent capabilities.</b></p>
<p align="center">
<a href="https://hive-tooling.vercel.app">Website</a> ·
<a href="https://hive-tooling.vercel.app/catalog">Catalog</a> ·
<a href="https://hive-tooling.vercel.app/audit">Audit</a> ·
<a href="#quickstart">Quickstart</a>
</p>
<p align="center">
<a href="https://www.npmjs.com/package/@echoo19/hive-mcp"><img src="https://img.shields.io/npm/v/@echoo19/hive-mcp?color=e7ba54&label=npm" alt="npm"></a>
<img src="https://img.shields.io/node/v/@echoo19/hive-mcp?color=555" alt="node">
<a href="./LICENSE"><img src="https://img.shields.io/badge/license-MIT-blue" alt="license MIT"></a>
</p>
---
Every skill, subagent, plugin, and MCP server you add to a coding agent spends context before it does any work. Tool schemas, system prompts, and always-on descriptions sit in the window whether or not you use them, and it is easy to lose a third of your budget to tools you forgot you installed.
**hive** is an MCP server that treats agent tools like packages with a cost. It finds tools for a task from the [Hive catalog](https://hive-tooling.vercel.app), plans a stack that fits a context budget, installs and records them in a `hive.lock`, and audits the always-on context an existing setup is spending. It runs inside any MCP client, so the agent that uses the tools is the one that manages them.
```
┌────────── hive · one MCP server ───────────┐
│ discover find tools for a task │
│ plan fit a stack to a context budget │
│ install write it into hive.lock │
│ audit price the context you carry │
└────────────────────────────────────────────┘
reads the Hive catalog · writes your hive.lock
```
Nothing runs a model inside hive. It reads a public catalog, writes a lockfile in your project, and reports token costs computed from the same model the [catalog](https://hive-tooling.vercel.app/catalog) uses.
## Install
```bash
npm install -g @echoo19/hive-mcp
```
## Configure
Add hive to your agent's MCP config:
```json
{
"mcpServers": {
"hive": { "command": "hive-mcp" }
}
}
```
That is the whole setup. No API key, no account, no cloud.
## Quickstart
Ask your agent to build something, and let it drive hive:
```
You: I want to ship a Next.js app on Vercel with a Supabase database.
Agent: → plan("Next.js on Vercel with Supabase", budget: "balanced")
Suggested stack (~1.2k always-on tokens):
gh source control cli 0 tok
mcp-supabase database mcp ~760 tok
vercel deployment cli 0 tok
→ install("mcp-supabase")
Recorded in hive.lock.
```
Later, check what your setup is costing:
```
You: How much context are my tools using?
Agent: → audit()
Always-on: ~4.3k tokens across 6 tools.
Lighter swaps: replace mcp-playwright (~1.6k) with the playwright CLI (0 tok).
```
## Tools
| Tool | Signature | What it does |
| --- | --- | --- |
| `discover` | `discover(intent)` | Find catalog tools that match what you want to build. |
| `plan` | `plan(brief, budget?)` | Suggest a small stack for a project, ranked by fit and context cost. `budget` is `lean`, `balanced`, or `capable`. |
| `install` | `install(slug)` | Install a tool and record it in `hive.lock`. |
| `uninstall` | `uninstall(slug)` | Reverse a recorded install. |
| `update` | `update(slug?)` | Reinstall a tool at the catalog's current version; omit `slug` to update all. |
| `sync` | `sync()` | Install everything listed in `hive.lock` that is missing. |
| `audit` | `audit()` | Report lockfile drift and the always-on context cost of your MCP setup, with lighter swaps. |
| `list` | `list()` | Show the tools recorded in `hive.lock` and their total context cost. |
## Why context cost
A tool's real price is the context it occupies on every turn, not its download size. hive scores that always-on cost per tool type: CLIs cost nothing until called, skills and subagents cost their description, MCP servers cost roughly the size of the schemas they expose, and plugins carry a fixed load. `plan`, `audit`, and `list` all read from that one model, so the number an agent sees in the client matches the number on the [website](https://hive-tooling.vercel.app/audit).
## Development
```bash
npm install
npm run build # tsc -> dist/
npm test # vitest
npm run dev # tsc --watch
```
The server talks to the live Hive API (`/api/match`, `/api/plan`, `/api/audit-index.json`) and installs from `/tools/<slug>/install.md`. Requires Node 18 or newer.
## Contributing
Issues and pull requests are welcome. See [CONTRIBUTING.md](./CONTRIBUTING.md) for how the project is laid out and what a good change looks like. To add a tool to the catalog itself, submit it through the [website](https://hive-tooling.vercel.app) rather than this repo.
## License
[MIT](./LICENSE) © Jake Kang
