# Labs: Open Source

> The tools we build for our own work, published under the MIT license. Each one handles part of the context problem: what a system knows about itself, what it knows about the world, and how its actions are recorded.

Canonical: https://www.adjective.us/labs

All repositories: https://github.com/adjective-rob

## 01. Internal context

*What an agent knows about the environment it works in.*

### [Prelude](https://github.com/adjective-rob/prelude)

TypeScript · MIT · `npm i -g prelude-context`

Machine-readable context for codebases. Prelude writes a project’s stack, architecture, decisions and module map into a .context/ folder and serves it to agents over MCP. Every session starts already knowing the code.

### [Snap](https://github.com/adjective-rob/snap)

Rust · JavaScript · Python · MIT

Visual feedback for coding agents. Press a hotkey, circle the problem, type a label. The annotated screenshot and its structured metadata land in an inbox your agent reads over MCP.

## 02. External context

*What an agent knows about the world outside it.*

### [Sonde](https://github.com/adjective-rob/sonde)

Python · MIT

A registry for collection intent. Declare topics, queries, sources and scoring rules for GitHub, arXiv, Hugging Face, RSS and web monitors in versioned files. Simulate before you collect, and trace every artifact back to the topic version that produced it.

### [Gerolamo MCP](https://github.com/adjective-rob/gerolamo-mcp)

TypeScript · MIT · `npx gerolamo-mcp setup`

Connects any MCP agent to our technology feed, a scored corpus of open-source projects, research papers and ML models. Search it, distill reusable patterns, profile creators and trace lineage.

### [Gerolamo Workflows](https://github.com/adjective-rob/gerolamo-workflows)

Markdown · MIT

Multi-step prompt templates that chain the feed’s tools into complete intelligence runs. The repo is the contribution surface. Any MCP agent reads the steps and executes them.

### [USASpending MCP](https://github.com/adjective-rob/usaspending-mcp)

TypeScript · MIT · `npx -y federal-spending-mcp`

Federal contract data for AI assistants. 18 tools over USASpending.gov for searching awards, profiling agencies and contractors, and running business development workflows. Public data, no API key.

## 03. Governed execution

*How an agent’s own actions are carried out and recorded.*

### [Honeymoon](https://github.com/adjective-rob/honeymoon)

Python · MIT

A local-first, multi-agent engine for development and security work. Agents build code, investigate codebases and simulate attacks in parallel swarms tuned for small models. Every action and report is cryptographically signed.