Autonomous agents shouldn't be given blind sudo or unbounded context windows. I engineer the deterministic layer underneath: terminal firewalls that intercept destructive shell commands, AST skeletonizers that strip function bodies to save 60–80% of prompt tokens, local circuit breakers that survive 429 Too Many Requests, and zero-knowledge vaults for local secrets. Everything runs locally, offline-first, without cloud vendor lock-in.

Multi-account OAuth authentication provider and intelligent model routing plugin for OpenCode. Automates credential refresh, manages Google Antigravity quotas (Gemini 3.8 Flash, Gemini 3 Pro, Claude 4.6), supports thinking-model variants, and provides zero-downtime fallback failovers across distributed account pools.

npm i -g opencode-antigravity

Zero-overhead terminal firewall and blast-radius containment engine for autonomous coding agents. Evaluates shell commands against strict AST/pattern policies before execution, blocking destructive filesystem modifications (rm -rf, mkfs, raw partition writes), unauthorized network pipes (curl ... | bash), and unvetted git force pushes.

npx agent-jail -- bash -c "npm test"

Local HTTP reverse-proxy and circuit-breaker for LLM client tools and agent loops. Intercepts 429 Too Many Requests and upstream rate-limits in real-time, streaming seamless failover traffic to backup models or local inference engines (Ollama, vLLM) without crashing multi-step agent runs.

npx quota-guard --upstream https://api.openai.com --fallback http://127.0.0.1:11434

Complete verification and discovery suite for Agent Skills (SKILL.md). skill-lint enforces schema rules, detects broken links, and calculates context overhead before agents burn tokens. skillmaxx deterministically discovers project topologies and provisions verified skills without hallucinated instructions.

npx skill-lint ./skills --strict
npx skillmaxx

🛡️ AI Agent Security & Runtime Infrastructure

Tool	Status / Distribution	Description	Quick Execution
opencode-antigravity		OAuth multi-account gateway & quota router for OpenCode	npm i -g opencode-antigravity
agent-jail		Terminal firewall & command sandbox against dangerous agent actions	npx agent-jail -- <command>
quota-guard		Local reverse proxy & circuit-breaker for HTTP 429 failovers	npx quota-guard --upstream ...
env-seal		AES-256-GCM local .env vault & compile-time process.env typings	npx env-seal --seal

⚡ Token Economy, AST & Agent Protocols

Tool	Status / Distribution	Description	Focus
context-trimmer	[TypeScript / AST]	Code skeletonizer stripping function bodies to preserve 60–80% prompt tokens	AST Compiler API
skill-lint		Linter & token-budget estimator for agent capability specifications	AgentOps / SKILL.md
skillmaxx		Topology discovery engine injecting agent instructions without hallucination	Project Topology
prompt-diff	[TypeScript / Diff]	Semantic diff & prompt-smell regression detector (pleading, weak negations)	Prompt Engineering
mcp-git-chronicle	[TypeScript / MCP]	Model Context Protocol server for Git archaeology, churn, & co-change coupling	MCP / Stdio
ts-deadcode	[TypeScript / AST]	Static import-graph analyzer detecting orphaned files & dead code	Static Analysis

🛠️ Local-First Developer Utilities & Zero-Dependency Tooling

Tool	Status / Distribution	Description	Quick Execution
git-pulse		Zero-API, offline Git activity card & SVG generator	npx git-pulse --days 90
schema-craft		Instant JSON Schema & TypeScript d.ts generator from JSON / stdin	curl ... | npx schema-craft --ts
fast-mock		Zero-config directory-based REST mock server with latency injection	npx fast-mock --dir ./mocks
http-replay		Record upstream HTTP responses for deterministic offline CI replay	npx http-replay --port 8080
markdown-deck		Single-file Markdown to offline interactive HTML slide deck presenter	npx markdown-deck slides.md
bench-diff	[TypeScript / CLI]	Micro-benchmark runner with median, p90, & terminal ASCII bar charts	Micro-benchmarking
sqlite-lens	[TypeScript / SQLite]	Zero-dependency terminal schema visualizer & index inspector	Terminal DB Inspector

1. Determinism over prompt optimism — Never ask an LLM to guess what can be extracted reliably via AST parsing, ripgrep, or filesystem checks. If code structure can be verified deterministically, keep it out of the prompt.
2. Context windows are not garbage dumps — Massive contexts increase hallucination drift and burn budget exponentially. Skeletonize code, prune orphaned branches, and enforce strict token budgets at compile time.
3. Local-first state, zero blast radius — Agent runtimes belong in local sandboxes (~/.config, SQLite, terminal firewalls), never blind sudo access or opaque cloud lock-in.
4. Zero-dependency discipline — Core developer utilities should install in milliseconds, run offline, and execute without fragile supply-chain bloat.

• Languages & Runtimes: TypeScript, Node.js (>=20), Bun, Go, Rust
• Agent Protocols & APIs: Model Context Protocol (MCP), OpenCode Extensibility API, HTTP SSE Streaming, JSON-RPC 2.0
• Security & Persistence: AES-256-GCM, Scrypt, SQLite, Linux Process Isolation, JSON Schema

• Found a tool useful? Star the repository — it directly drives open-source discoverability.
• Encountered a bug or want a new feature? Open an issue on the respective repository.
• Building autonomous AI agents or local developer runtimes? Connect on GitHub.