Skills, tools, MCP servers, browser extensions, and UI themes for AI agents. Built for developers who need practical, tested components instead of hello-world examples. Provider-agnostic except for the hermes/ directory.
All items follow a simple format: the problem that needed solving, what was built, and the outcome. Detail lives in each artifact's own docs.
Daily agent workflows.
|
|
| Problem |
Files grow unchecked and silently degrade agent performance until you notice instructions being ignored. |
| Built |
A skill that enforces file size discipline, audits bloated files, and encodes four techniques from Andrew's OpenClaw Blueprint video on the 18,000 character trap. |
| Outcome |
Your agent actually reads every instruction because nothing gets truncated or pushed out of context. |
|
|
| Problem |
Setting up recurring tasks on macOS means wrestling with launchd plist syntax and remembering the right load/unload commands. |
| Built |
A skill that handles launchd configuration, scheduling syntax, and task lifecycle (load, unload, start, stop) so you don't have to. |
| Outcome |
Schedule scripts without touching XML or digging through man pages. |
SDLC pipeline in 7 ordered phases.
|
|
| Problem |
Most bugs come from unclear requirements. |
| Built |
A phase that forces you to write down objective, assumptions, and boundaries before any code gets written. |
| Outcome |
Catches ambiguity early when it's cheap to fix. |
|
|
| Problem |
A spec without a plan leads to messy execution. |
| Built |
Breaks specs into ordered, dependency-aware tasks. |
| Outcome |
Nothing gets built before its prerequisites are ready. |
|
|
| Problem |
Big bang merges break things and hide who broke what. |
| Built |
Thin, testable vertical slices. |
| Outcome |
Each piece works before the next one starts. |
|
|
| Problem |
Writing tests after the code means tests that pass against broken logic. |
| Built |
Red-Green-Refactor and reproduction-first bug fixes. |
| Outcome |
The test suite stays honest and the code stays correct. |
|
|
| Problem |
Merging without review lets bad patterns compound. |
| Built |
Five-axis review gates (correctness, security, performance, style, test coverage). |
| Outcome |
Catches what the author missed. |
|
|
| Problem |
Code that works but is hard to understand is technical debt waiting to compound. |
| Built |
Systematic complexity reduction without behavior changes. |
| Outcome |
The next person can reason about the code. |
|
|
| Problem |
A clean merge doesn't mean a clean deploy. |
| Built |
Rollout, monitoring, and rollback discipline. |
| Outcome |
The launch doesn't become the incident. |
|
|
| Problem |
Standard headless browsers get blocked by Cloudflare and LinkedIn immediately. |
| Built |
agent-browser CLI (Vercel Labs) combined with a stealth Chrome extension that patches browser fingerprint vectors at document_start, before any page scripts run. |
| Outcome |
Gets through WAFs that normally stop automated browsing cold. |
|
|
| Problem |
Remote coding agents lose state when sessions break or auth expires. |
| Built |
A skill covering session lifecycle, RPC control, response extraction, and failure recovery for pi --mode rpc. |
| Outcome |
Your coding agent survives terminal restarts, auth reconnects, and network blips without losing context. |
Standalone CLIs, each with their own docs.
|
|
| Problem |
Most bookmarking tools are either SaaS or require heavy setup with databases and config. |
| Built |
A single Python file with zero dependencies that gives you full CRUD on Karakeep and optionally syncs to Obsidian. |
| Outcome |
Ready to use the second you download it. |
|
|
| Problem |
OpenCode costs add up fast and the CLI doesn't tell you what you're actually spending. |
| Built |
Usage breakdowns by day, week, and month with comparisons to OpenRouter equivalents. |
| Outcome |
You see budget drift before it becomes a surprise bill. |
|
|
| Problem |
Gmail piles up noise faster than you can triage it. |
| Built |
A two-layer pipeline: a zero-LLM archiver that filters noise instantly without spending tokens, and an LLM-powered reviewer that surfaces only the signal. |
| Outcome |
Noise is gone before you even see it. |
MCP servers for agent platforms.
|
|
| Problem |
Browser automation MCP servers hit the same WAF blocks as everything else. |
| Built |
Wraps agent-browser CLI (Vercel Labs) with anti-detection flags and a stealth extension bundled inline with no external dependency. Exposes 10 tools (navigate, click, eval, screenshot, and more). |
| Outcome |
Works through Cloudflare and LinkedIn from any MCP-compatible agent. |
|
|
| Problem |
Most stealth extensions load too late when fingerprint signals are already exposed. |
| Built |
A Manifest V3 extension that runs at document_start, before any page script executes, and patches navigator.webdriver, plugins, languages, and Permissions API. |
| Outcome |
By the time the page loads, your browser looks native. |
|
|
| Problem |
Staring at a bright dashboard in a dark room gets old fast. |
| Built |
Two themes: Clean WebUI for a bright, minimal workspace and Gruvbox Dark for warm-toned low-light sessions. |
| Outcome |
Drop them in and the dashboard matches your environment instead of fighting it. |
Platform-specific tooling for Hermes Agent.
|
|
| Problem |
Using pi RPC from Hermes meant manually managing sessions and parsing raw responses. |
| Built |
A plugin that wraps the full lifecycle (start, send, poll, stop) as native Hermes tools, tested against the actual pi protocol. |
| Outcome |
Handles edge cases and auth failures cleanly without manual session wrangling. |
Shared checklists.
git clone git@github.com:ShaggyD/agent-toolkit.git ~/agent-toolkit
# For Hermes users: symlink for skill discovery
ln -sf ~/agent-toolkit ~/.hermes/skills
# Load a skill (hermes -s <name> or /skill <name>)
hermes -s browser-automation
# Register an MCP server
hermes mcp add stealth-browser --command "python3 path/to/mcp/stealth-browser/assets/mcp_server.py"
# Install a CLI tool
cp tools/karakeep/kk ~/.local/bin/
MIT - use freely, adapt openly, attribute where meaningful.
Attribution:
Maintained by Dustin "Dusty" Chadwick