I'm Jordan. I work @cloudflare on the agent experience team.

I build small tools for AI agents: memory, sandboxes, proof, control surfaces, browser automation, and Cloudflare infrastructure. I write about it at coey.dev.

• write-judge: your coding agent can't delete the test to make it pass. A Pi extension; a small yes/no model judges every edit before it lands. pi install git:github.com/acoyfellow/write-judge
• coil: one thread, one agent, one computer. Pi Durable and a Cloudflare Container in one Durable Object.
• skin: an app that rewrites its own code while you use it, and can't brick itself.
• tardigrade: an agent you can kill, and it comes back. Pi Durable + Durable Objects + Artifacts.
• hark: a voice you talk to. No chat box. Deepgram Flux hears, Workers AI thinks, and it speaks back.
• odds: ask for odds, not prose. Clef decision models in Pi codemode.
• chat-ax: a shared team chat room with a durable fleet of AI agents, on your own Cloudflare account.

• my-ax: a personal AI agent operating environment you self-host on Cloudflare. Chat, desk, recipes, machines, and a factory that classifies GitHub issues. A human still merges.
• loop: one permanent conversation with an agent that builds your interface as you work. Every reply can compile a real Svelte 5 widget next to the chat, and it remembers what you tell it.
• mote: a programmable, local-first Mac control shell. Edit a Svelte file; the native menu-bar interface changes immediately.
• filepath: your own background agent environment. Workspaces, sandboxed filesystems, harness registry, and your own model keys.
• up: put a company's private web online on its own Cloudflare account. A folder becomes an immutable, Access-protected URL.
• hangar: give an agent a task, not your keys. Ready locked boxes on Cloudflare.
• spawn-agent-cloud: spawn an AI agent in Cloudflare, not on your laptop.
• executor-cloudflare: deploy Executor, the open-source agent integration layer, to your own Cloudflare account in one step.

• terrarium: execution, status, receipts, cancellation, and callbacks for bounded delegated agent tasks. One bounded task → one child run → one inspectable result.
• terraloop-mode: no child agents spawn until a contract is locked and a driver loop exists. A Pi extension.
• hold: a job waits. A runner claims once. There is no poll.
• lease: subscribe a pantry recipe to a peer outbox for a bounded time.
• room: disposable URL. Mint a room.
• trick: two URLs. Write yours. Read theirs. Not a chat.
• cell: invoke a named tenant. Get result and evidence. Do not receive source.
• tip: commits and trees. History is the product.
• pantry: a capability-scoped recipe store on Cloudflare Workers + D1. Stores and hands back scripts; never runs them.
• harness-runtime: a runtime contract for agent context, capability, authority, execution, proof, and continuity.
• agent-patterns: a practical library of bounded, steerable agent execution patterns.
• promotion-atom: a deterministic kernel that decides when successful agent work becomes a reusable artifact.
• ralph: processes stories from a PRD, one story per run, with strict checks.

• deja: repository-scoped memory and handoffs for coding agents. Agents learn from failures; deja remembers.
• wake: sessionless work continuity and handoffs for AI coding agents.
• imprint: immutable commit-bound repository context so an LLM can answer, compare, promote, and roll back against exact code versions.
• cache-layer: verified recipe routing for AI agents on Cloudflare.
• rolling-context-lab: a reproducible lab for measuring rolling context compaction in long-running agents.
• pi-fresh-session-handoff: context-threshold hooks, Deja auto-memory, and safe fresh-session handoff for Pi.

• argus: a sourced web research agent that gathers evidence, follows weak spots, and returns a clear report with links.
• orbit: bounded public-source monitor that turns releases and posts into evidence-linked briefs.
• pai-agent: research and analysis agent as a Cloudflare Durable Object with shell access, tool use, and WebSocket streaming.
• swarm: parallel research agents on your own Cloudflare account, across Workers, Durable Objects, and Workers AI.
• effect-agents: five agent patterns proven with Effect v4 — parallelism, streaming cancellation, approvals, typed errors, and MCP exposure.
• parley: two AI models debate until they agree on a plan.
• molt: an agent writes a tool, the loop injects a fault, and the agent repairs it. A separate verifier decides whether it passed — the agent does not.
• ditto: run multiple Cloudflare AI models simultaneously. Merge every output with consensus.
• speedrun: three Workers AI models race to fix the same bug in containers; a referee reruns the tests.
• jido-cartographer: map a public GitHub repo with one Jido agent per source file, on the BEAM.
• benchmarks: tool-calling reliability and latency tracking for Workers AI models.
• lower: one floor lower each experiment, from JavaScript to the math inside models. Claims and receipts.

• echo: drive a browser tab you're already signed into over MCP, without handing an agent your password or session token.
• unsurf: turn any website into a typed API. Scout, replay, auto-heal.
• agentcast: live browser sessions for AI agents. Watch them work in real time.
• machinectl: MCP server that lets AI control your machine from any device.
• chomp: OpenAI-compatible proxy for free LLM providers. Bring your keys. We route the rest.
• mcp-code-mode: wrap a Node/Bun MCP server with fail-closed search and sandboxed multi-tool execution.
• cmux-webchat: drive authenticated web chat UIs through cmux without exporting browser credentials.
• vidframes: cost-aware video frame extraction, vision analysis, and audio transcription via Workers AI.
• image-production-kit: small, deterministic tools for producing high-resolution web image variants and image lightrooms.
• anytool: give your AI agent infinite tools with one API call.
• inbox.dog: Gmail OAuth for AI agents. Read, write, search, and send with one integration.
• firestore-mcp-kit: build secure, typed MCP tools backed by Firestore with explicit Zod schemas.
• getlee: turn any prompt into a link that opens Cloudflare's Agent Lee, ready to send.
• lee: a standalone open-source Agent Lee client built with Svelte, Hono, and Effect 4.
• lee-model: Agent Lee as a model name. An OpenAI-compatible Worker with server-side tools.
• artifacts-lee-skills: team-editable agent skills on Cloudflare Artifacts. git push to update.
• mulder: add selected read-only API tools to a website with WebMCP.
• predict: a small browser signal for the next useful step on a website.
• jev-code-mode: typed Jev judgments behind two MCP tools, search and execute.
• homerun: turn your laptop into a local HTTP proxy for web automation.
• ralphwiggums: Effect-first browser automation on Cloudflare Workers. Give it a prompt, get a completed task.
• effect-first: a resource for agents that want to write proper Effect programs.

Extensions for Pi, the coding agent — the tools I run my own agents on every day.

• witness: a coding agent shouldn't decide its own work is done. Runs a real check against what the agent produced and blocks the action if it fails.
• hashline-pi: line-anchored, tag-verified file edits with in-memory preflight and 3-way-merge recovery.
• mutex-pi: cross-process file lock and optimistic stale-read guard for Pi's read, edit, and write tools.
• loops-yaml: session-scoped recurring prompts for Pi and OpenCode, plus tiny scheduled or on-demand command loops.
• auto-thinking-pi: classifies prompt difficulty with a local heuristic and sets the reasoning effort for the next turn — no extra model call.
• edit: a small, approval-first Pi tool for safe, checked code changes.
• pi-ext-durable: a stock Pi extension running inside a Durable Object agent on pi-durable.

• lab: a sandbox that hands you the receipt. Isolated execution on Cloudflare Worker Loaders; every step becomes a shareable trace.
• cloudbox: durable Cloudflare computers for agents — run repos, steer live workspaces, verify results, and return receipts.
• cloudshell: browser-based terminal and dev environment in the cloud. Instant access to a full shell from anywhere.
• cloudterm: DOM-rendered web terminal emulator with custom ANSI parsing and latency prediction.
• mcpu: an MCP control surface for a Cloudflare Artifacts repo. GitHub is only the bootstrap seed; after first deploy, commits live in Artifacts and deploy back to Workers.
• promptlog: Dynamic Worker Loader with sandboxed code execution.
• ffmpeg-container: tiny Cloudflare Containers hello world proving ffmpeg runs at the edge.
• termlab: side-by-side benchmark of web terminal emulators.
• tsc-rs-worker: the TypeScript 7 type checker (tsc-rs) running in memory inside a Cloudflare Worker.
• nix-cf-proof: one devenv.nix gives a git that cannot force-push, on your Mac and in a Cloudflare Container.
• zero-cloudflare-hello: run a Zero native executable on Cloudflare Containers.
• tldraw-sandbox: an infinite canvas with AI-powered code terminals.

• glance: see your projects move. A local-first dashboard that observes work by people and agents as a live feed, color, sound, and a responsive 3D orb.
• desk: a personal app store for tiny edge devices. git push installs apps; MCP lets agents use the device as an out-of-band human I/O surface.
• t2t: local push-to-talk dictation for macOS. Hold Fn to type with on-device Whisper; hold Fn+Ctrl to talk to a local Pi agent.
• living-artifact: physical and digital artifacts and tiny devices, with Cloudflare-hosted firmware OTA and health-gated rollback.
• anvil: animations agents can write. Lottie + three.js in one JSON dialect, on Cloudflare Workers.
• blackbox: replay a Pi agent log as a timeline and see which step Clef scores as the fault.
• lookout: private, view-only live screen streaming on Cloudflare.
• entropy-c: a portable physical-entropy experiment for Cloudflare's NYC office.

• gateproof: a local proof loop. Put the claim in a typed plan; it runs gates — observe, act, assert — until they pass or stop. Workers may change implementation. They may not rewrite plan.ts.
• tollgate: a local HTTP policy gate. Every request meets a rule before it leaves your machine.
• airlock: push a candidate version, run its tests, and make it live only if a signed proof of those tests passes.
• agent-admit: the model proposes. An admitter it does not own decides. Map of where that is live.
• mutant: mutation testing for TypeScript. Finds code your tests run but never actually check.
• keel: a provider-agnostic control plane for verified self-update. A candidate is identified by content and only promoted when its proof checks against trusted keys.
• cloudeval: run model evals, compare results, and ship shareable reports.
• corrections-to-gates: mine your own AI agent-correction exhaust into deterministic $0 gates for the mechanical slice of your corrections.
• visual-diff: framework-agnostic 3-tier visual-diff cascade — structure, computed style, and pixels — for a fail-closed 1:1 verdict.
• semantic-diff: extends visual-diff to a fail-closed 5-tier verdict with accessibility/semantic tiers.
• vitest-visual-diff: compare two live elements through structure, computed style, pixels, and accessibility in Vitest Browser Mode.
• agent-human-bridge: auditable agent and human software-building protocol designed to pass a human safety review.
• repo-audit: deterministic repo scoring across 11 quality dimensions.
• cf-tutorial: 95 questions to prove you know Cloudflare.
• vet: CASA Tier 2 / OWASP ASVS security scanner.
• deadlint: find dead public methods on Workers DurableObject / RpcTarget / Agent classes, plus structural clones. The linter layer above knip/oxlint.
• ascii-smuggler: Unicode smuggling and steganography toolkit for security research.
• gate-review: red-team your gates before implementing them.
• nightglass: an owned, deterministic classifier for whether agent claims are supported by evidence.
• killjoy: crash-test Durable Object agents. Kill at every save point and check what breaks.
• basic: a conformance suite for agent chat. The behaviors every chat app must keep, as Gherkin.
• capability-frontier: a reproducible benchmark for useful coding-agent autonomy under enforced capability limits.
• agent-adversary: fail-closed Rust/Wasm validation for adversarial agent operations.
• self-healing-ci: a bounded reference implementation for authenticated CI repair.
• mutant-nightly: a nightly agent that finds proven test gaps with mutation testing and opens a PR to close them.
• flue-gitlab-test-improver: a nightly Flue agent that adds missing tests to a GitLab repo and opens a draft MR.
• greenlight: gates go in red. Code comes out. Gates turn green.
• exhaust: turn the review comments you wrote into an evidence-linked checklist.
• goodput: evidence-adjusted efficiency for AI-assisted work.
• doclint: returns what an agent understood from your documentation, so you can see what it missed.

• svelte-hono: Svelte 5 SSR and hydration in a Hono Worker, without SvelteKit.
• kumo-compiler: author a component once in canonical React; compile to genuine native Vue, Svelte, and Solid.
• tuiport: real SSH applications built with OpenTUI and deployed in a Cloudflare Container.
• svelte-edge: agent-generated Svelte 5 UI compiled on Cloudflare Workers into inline edge bundles.
• capa: third-party APIs as proof-carrying Cloudflare service bindings.
• UserDO: per-user Durable Objects as data pods.
• JotDB: durable state without migrations. One Durable Object per user, tenant, room, or actor.
• blaze: real-time documents on Cloudflare. Firestore feel, Cloudflare price.
• fleet-pattern: hierarchical Durable Objects with ordered fan-out.
• workflow-live: real-time workflow monitoring on the edge.
• cache-sync: edge cache invalidation that actually scales.
• tiny: real-time collaborative todos with TinyBase + Durable Objects.
• bio: single-button WebAuthn auth on Cloudflare.
• ironalarm: reliable task scheduling for Cloudflare Durable Objects.
• ai-connect: minimal zero-dependency primitive for configuring and calling AI inference endpoints.
• remote: barebones starter for authenticated apps with persistent state on the edge.
• share: temporary, resumable file shares on Cloudflare Workers, Durable Objects, R2, and Turnstile.
• flue-snippets: real, runnable Flue agents on Cloudflare. Every snippet ships an E2E test that deploys, asserts, and tears down.
• think-snippets: live-proven Cloudflare Project Think contracts. Each example deploys to a real Worker, proves its claim against live state, and tears itself down.
• optkit: email subscriber management for Cloudflare Workers.
• facets-stress-test: how far one Durable Object can go as a parent of facets, measured in production.
• kumo-astro-demo: Astro consuming Kumo design-system components directly. No React.
• phoneline: a minimal voice AI agent on Workers, Twilio, and OpenAI Realtime.
• rank: an agentic SEO loop that runs on your own Cloudflare account.
• tax-agent: an AI tax filing agent on Cloudflare Workers.

• vibe-cdn: browser games on Cloudflare's edge.
• arcade: neon hovercraft survival. A browser game built and deployed on Cloudflare.
• heist: a prompt injection game. Get VaultBot (Llama 3.3 70B) to leak the vault.
• zoo: a shared world where each creature runs JavaScript a model wrote from one sentence.
• terra: a multiplayer 3D terrain playground on Workers, Svelte, three.js, and Durable Objects.
• oven: chat a 3D game into existence on Cloudflare.
• food-court: multiplayer food games on Workers + Durable Objects. One repo, many stalls.

• phonesites: build a website in 5 minutes from any device.
• signedseal: self-custody e-signatures.
• formdrift: voice, photo, and document input for web forms.
• formwing: AI outreach from contact forms.

Some favorites from coey.dev/blog.

• Patience: You've Never Controlled Anything: what Memento taught me about prompting agents.
• Campfire: The Oldest Protocol: every storytelling medium had a context window. We've always been writing prompts, we just called them stories.
• Prompts Are Wishes: every prompt is a hypothesis. The guardrails are somewhere else.
• Liquid Primitives: five parts of your stack to hold loosely.
• Worker Loaders as a Place: Dynamic Workers clicked when I stopped thinking about the API and started thinking about the room.

More at coey.dev/blog.

coey.dev · blog · @acoyfellow