Software architect focused on AI agent systems and developer tooling. Open source enthusiast building the infrastructure behind AI Skill Market.
Editorial note: Articles under this byline are researched and drafted with AI assistance under Duke's editorial direction, then reviewed before publishing. See our editorial standards for details.
One project replaces a prose approval reviewer with typed probabilities; the other screens Agent Skills and MCP configs before install. Both report real numbers, both call themselves early.
Three small projects push Jev outside its usual routing/gating role — scoring video transcripts, turning prompts into pixels, and composing UI from a fixed shadcn/ui component catalog.
jev-align tunes AI function definitions with Jev and human feedback via GEPA. effect-agent wires Jev into TypeScript's Effect framework. Two takes on 'typed.'
A quant-research project that actually built its own model instead of using Jev — plus a look back at what 517 Jev projects, most of them tiny, tell us about the ecosystem.
Two unrelated repos share the name jev-code; a third adds Jev-based recommendation to agent harnesses. What each does, and why matching names aren't matching projects.
Two small plugins put Jev inside Codex and Claude Code review workflows — one for general decision consultation and failure triage, one narrowly for Rust.
A dataset-to-LoRA workbench, a pandas/Polars scoring library, and a plain-Postgres semantic query tool: three ways typed Jev judgments are entering data pipelines, with their stated limits.
A MuJoCo quadrotor simulation splits flight into three frequencies: 500Hz control, 50Hz safety, and ~2.5Hz tactical calls from Jev. It's sim-only, and the author says so plainly.
Two repos share a name and a model. One is game demos testing Jev's reflexes; the other, from developer-educator Nader Dabit, is a pre-commit security scanner with real benchmark numbers.
One project routes social-media research through Jev, and two unrelated tools named jev-seo audit pages against dozens of rules. What their evidence actually supports, and what it doesn't.
Two extensions push typed Jev judgments into SQL itself, for Postgres and DuckDB. Batching cuts one benchmark 27x — but both are pre-1.0 and send row data to an external API.
Three projects put typed judgments at three different layers: a Discord bot, an nginx edge filter, and a coding-agent permission gate. What each one measures, and what it doesn't.
Two open-source teams rebuilt Jev's typed decision head from scratch. Kev ships three model sizes with published accuracy numbers; NanoJev claims to beat Jev on ViZDoom at 0.6B params.
Two unrelated projects share the name jev-search, plus a semantic grep called jev-semgrep. Three ways typed Jev scoring is replacing keyword matching in search and code tools.
jev-trader posts limit orders on a Monad spot pair every ~300ms and defaults to a dry run. Its README reports 81ms model latency and no capital at risk unless you set a private key.
jev-voice pairs local whisper.cpp with one Jev call per spoken command; JevBrowserExt ports jev-ultrafast to a Chrome extension; jev-macos-loop drives Finder with local OCR. All three keep Jev to a single narrow decision.
Six small repos, all filed under Game Development on the Jev radar. Two are literal head-to-head matches, one is a local decision server, three are solo tick-by-tick demos.
jevals runs 40+ typed evaluation questions per agent trace for a fraction of a cent, using Jev instead of an LLM judge. Here's the architecture, the cost claims, and where it says it falls short.
Jevbridge speaks both MCP and the Agent Client Protocol, and can swap Jev for another LLM when no TypeSafe key is present. Here's what that trade-off actually costs.
Five separately-authored projects named jev-browser or jev-browser-use, from a 415-star Codex speedup tool to a 4-star Playwright CLI. Here's what each one actually does.
A small n8n community node lets no-code workflows call Jev for classification, routing and scoring. Here's what it ships, what it needs, and where it's still rough.
Three separate extensions wire Jev into the Pi coding agent, two of them sharing a name. Here is what each actually does, and how to tell them apart before you install.
Five projects now put Jev between a coding agent and its tools: two built specifically for Pi, three all named jev-guard. Here's what each one actually checks, and how much evidence backs it.
simple-jev turns open LLMs into Jev-compatible classifiers, jev-skill packages 60+ scenario demos, jev-shell-history ranks your zsh history. Three small tools, three bets.
jevbench, typesafe-ai-benchmark, jev-benchmarks and jev-rerank-bench measure Jev differently. What each one tests, what it found, and how much to trust it.
Two small Chrome extensions send page elements to Jev for a yes/no or best-match call: one blocks ads by asking 'is this an ad?', the other fills forms by asking 'which note matches this field?'
Two Jev-powered computer-use agents, two different tradeoffs: OCR-and-accessibility screen reading versus text-only Codex integration. Here's what each README actually proves.
typesafe-sdk-js and typesafe-sdk-python are TypeSafe's own client libraries for Jev. What each ships, how their APIs differ, and what's still undocumented.
A WeChat reply assistant with offline OCR and a Home Assistant integration both use Jev for narrow, typed judgments — never auto-sending messages or controlling safety-critical devices.
Nine game and sim projects landed in the Jev ecosystem's first weeks — two open reimplementations, two duels, a local decision server, and four solo tick-by-tick demos. Here's what each proves.
Typed decisions come with probabilities, but a threshold is only useful if you've measured it. jevcal, an eval-first jev-mcp and LightJev show how to calibrate and catch drift.
The same Jev pruning idea now exists for Claude Code, Codex, Pi and omp. How each implementation differs, what the READMEs measure, and how thin the evidence still is.
Two open-source tools use Jev to decide which old tool output an agent can forget, keeping originals verbatim. How the pruning pattern works and where it can go wrong.
166 projects use Jev, and most are tiny and unbenchmarked. Here is how to read pinned vs unreviewed entries, star counts and author-tested claims before you adopt one.
jev-desktop, mobile-jev, jev-voice-browser and ego-jev use Jev to pick the next UI action. Here is what each README proves, what it discloses, and where it stops.
Semantic filters inside SQL, a LlamaIndex reranker and router, and a dataset sifter: three Jev projects that put typed decisions into data pipelines, with their real limits.
Typed probabilities make cheap guardrails possible: injection screens, risk scores on tool calls, gateways. What jev-guard and Safer with Jev show, and what they don't.
Scores, not prose: how two Jev Review tools, a GitHub Action and a flash-review engine turn code review into typed signals a coding agent can act on.
browser-use's jev-ultrafast lets Jev pick the next DOM action while a small model types. The author test shows 9.45s to 7.09s, from three run pairs. Here is what that does and doesn't show.
jevpilot, jev-gomoku, jevarena and two trading repos push Jev into tight decision loops. What they demonstrate, what they don't prove, and why trading needs extra care.
Three open-source routers use Jev to decide where a request goes. One reports a 60% list-price saving from a replay of 237 turns. The fine print matters as much as the number.
Blink walks file trees and neo4jev walks Neo4j graphs by asking Jev which branch to take next. How index-free navigation works, and what the demos do not prove.
Agents parse free text to make routing and gating calls. A typed decision layer returns a choice plus probabilities instead. See how two small Jev projects show the difference.
Jev is TypeSafe AI's System One model: it returns typed choices, scores and yes/no answers with probabilities instead of text. Here is what that means and who is building on it.
typesafe-mcp, jev-mcp, jevwire and Jevbridge all expose Jev as a decision tool. Here is how they differ, what each ships, and which claims are still unproven.
Beyond flights and hotels: AI travel skills for free campsites, Dubai yacht rentals, vegan food itineraries, and live flight tracking — the long tail of travel automation.
Use AI skills to find cruise deals on Carnival and Expedia Cruises, then research ships with Cruise Critic reviews — promotions, sailings, and ratings as structured data.
How two AI hotel-booking skills drive Booking.com and Airbnb autonomously — comparing hotels versus short-term rentals on filters, pricing and structured JSON results.
Agent37 is Nous's hosted Hermes from $3.99/mo. Self-host runs $4-8/mo on a VPS plus ops. Decide with cost, compliance, data locality, and ops capacity in mind.
Your Claude Code CLAUDE.md is static project context. Hermes memory is dynamic and agent-curated. Here is how to move between them without losing signal.
Hermes ships devops and mlops skill categories. Tour representative skills and watch them chain through a typical incident — alert to rollback — with Claude in the loop.
Step-by-step port of a write-changelog skill from Claude Code to Hermes. Original file, target file, line-by-line what changes and why.
Tour the red-teaming skill category that ships with Hermes Agent. What each category of skill does, responsible-use norms, and how skills chain in an engagement.
Step-by-step setup for the hermes-acp entry point in VS Code, Zed, and JetBrains. Get a persistent Hermes agent panel inside your editor.
Two ways Hermes talks to Claude Code: claude -p print mode and tmux-backed PTY sessions. When to use each, what they cost, where they break.
Hermes ships with a bundled skill that teaches it to delegate heavy coding work to Claude Code. Here is how the delegate_task() pattern works in practice.
Claude Code is your IDE agent. Hermes is your server-side agent. Here is how to decide which tool fits which job, and why most teams end up using both.
OpenClaw extends Claude Code inside its runtime. Hermes runs a second, Python-based runtime that calls Claude. Compare philosophies, layers, and where each fits.
Short answer: yes, format-compatible. Long answer: there are a few differences in frontmatter, triggers, and related-skills resolution. Here is what to change.
Model Context Protocol is identical in both runtimes. What differs is the host lifecycle. Compare filesystem, GitHub, and Postgres MCP servers side by side.
Learn how to design skills that guide Claude's tool usage. Master when to invoke tools, how to interpret results, and tool-aware skill patterns.
Build Level 3 AI skills that leverage tools for file access, API integration, and system interaction. The gateway to powerful, practical automation.
Build AI agents that select and invoke external tools. Learn tool definition, parameter extraction, result handling, and error recovery patterns.
Master the router pattern for AI agents. Learn how LLMs decide execution paths, implement conditional logic, and build intelligent routing systems.
Learn to build Level 2 AI skills that use templates and placeholders for structured, consistent output generation across documents, code, and content.
Design AI skill outputs that are consistent, parseable, and useful. Master output format specification for reliable, actionable results.
Design and implement multi-agent systems. Learn orchestration patterns, agent communication, task delegation, and building collaborative AI teams.
Master the multi-agent design pattern. Learn how to build systems where specialized agents collaborate to solve complex problems together.
Learn how to implement effective guardrails and constraints in AI skills to ensure reliable, safe, and predictable behavior in production environments.
Master the art of composable AI skills. Learn patterns for building modular, reusable skill components that combine into powerful workflows.
Learn how to build systems where multiple AI agents collaborate. Patterns for coordination, delegation, and collective intelligence.
Learn to design AI skills with focused, clear purposes. Master the art of scoping skills for maximum effectiveness and minimum overlap.
Understand Level 1 agents - the foundation of agentic systems. Learn when simple responders are enough and how to build them effectively.
Explore fully autonomous AI agents that generate and execute code independently. Learn safety patterns, supervision strategies, and responsible autonomy design.
Master Level 1 AI skills that enhance prompts through tone, style, and vocabulary adjustments. The simplest skills with the fastest path to value.
Master the art of integrating tools into AI agents. Learn tool design, MCP servers, custom tools, and best practices for connecting agents to the real world.
Build Level 4 AI skills that coordinate multiple specialized agents for complex tasks. Master delegation, communication, and result synthesis.
Understand how memory transforms AI agents from stateless responders to intelligent systems that learn and remember. Implement all three memory types.
Learn how to implement guardrails that keep AI agents on track. Prevent harmful outputs, maintain boundaries, and ensure reliable agent behavior.
Master context injection techniques to teach Claude domain-specific knowledge. Learn to inject conventions, rules, and expertise effectively.
Learn how role-playing techniques dramatically improve AI agent performance. Specific roles, personas, and contexts lead to better results.
Build Level 5 AI skills that operate autonomously with self-directed execution, adaptive decision-making, and minimal human intervention.
Why specialized agents outperform general-purpose ones. Learn how to design focused agents that excel at specific tasks.