---
title: "Earthworks — working AI tools and applets"
url: https://setearth.io/earthworks.md
for_agents: /agents.md
updated: 2026-10-02
---

This file backs the /earthworks/ route for llms.txt and the sitemap. The
rendered page builds from content/earthworks.json; this body is route
metadata, not page copy.

## The catalog

- **pi** (pi, open-source, production, open) — The coding-agent CLI my agents run inside: a terminal harness that gives a model file, shell, and subagent tools under operator control, with durable sessions and worktree isolation. Every agent session described on this site runs inside it or the multiplexer above it. Author: Mario Zechner / earendil-works. Recommended uses: Running coding agents in a terminal with durable sessions and operator-held gates; A harness built to extend: the other pi entries in this catalog are extensions of it.
- **WorksGood wg** (workgraph, open-source, production, open) — The work-coordination graph every agent task on this site runs through: tasks with dependencies, dispatcher-spawned workers in isolated worktrees, receipt-backed completion, and evidence fencing for humans and AI agents. This catalog was built by agents dispatched through it. Author: graphwork. Recommended uses: Coordinating multiple coding agents on one graph without stepping on each other; Task work with deterministic completion checks and fenced evidence.
- **Superpowers** (superpowers, skill, production, open) — A software-development methodology for coding agents, delivered as a fourteen-skill set (brainstorming, test-first development, systematic debugging, planning, code review) that make the agent work to a discipline instead of improvising. The full set installs from the upstream repo. Author: obra. Recommended uses: Any team whose AI coding agent produces clever but undisciplined work; Standardizing design-before-code, test-first, and verification habits across engineers.
- **Herdr** (herdr, open-source, production, open) — A pane-based agent multiplexer: a persistent headless server owning real PTYs, with TUI clients, workspaces, tabs and panes, and live agent-session state for auto-resume. Author: herdrdev. Recommended uses: Operating a fleet of coding agents in panes with sessions that survive a restart; Replacing ad-hoc tmux sprawl for agent ops.
- **Obscura** (obscura, open-source, production, open) — Headless browser engine written in Rust for scraping and agent automation. Real V8 JavaScript, Chrome DevTools Protocol, a drop-in for headless-Chrome workflows, and it ships an MCP server. Author: h4ckf0r0day. Recommended uses: Agent web automation without a Chromium install; MCP-driven browsing from any coding agent.
- **Zulip** (zulip, open-source, production, open) — Open-source team chat with threads; I run it as the message channel between agents and people across engagements. Author: Zulip. Recommended uses: Threaded async coordination for distributed teams and agent fleets; Self-hosted messaging where data stays in-house.
- **tmux** (tmux, open-source, production, open) — The terminal multiplexer: persistent sessions, panes, and windows that survive disconnects. It is the substrate I run multi-agent fleets on. Author: tmux. Recommended uses: Multi-agent orchestration: one pane per coding agent, with sessions that outlive the connection; Durable remote sessions on headless hosts, resumable later from desktop or mobile.
- **Tailscale** (tailscale, tool, production, open) — Zero-config WireGuard mesh VPN (open-source clients over a hosted coordination plane) that puts every host I run on one private tailnet with stable 100.x addresses. Author: Tailscale. Recommended uses: Private tailnet dev previews: share a build at a stable 100.x address (e.g. 100.68.255.106:8183) with no public exposure; One flat network across laptops, dev boxes, and cloud hosts, reachable from desktop and mobile anywhere.
- **Termius** (termius, tool, production, open) — Cross-platform desktop SSH client with hosts, keys, and snippets synchronized across devices. Author: Termius. Recommended uses: Day-to-day SSH from the desktop into my hosts and client hosts on the tailnet; Keeping host and key inventory in sync across machines instead of per-device ssh_config sprawl.
- **Moshi: SSH & MOSH Terminal** (moshi, tool, production, open) — iOS terminal over SSH, Mosh, and ET built for driving remote hosts and coding-agent sessions from a phone; tmux-aware, so mobile attaches to the same panes the desktop sees. Author: Moshi. Recommended uses: Attaching to tmux agent-fleet sessions from a phone over Mosh (roaming-friendly; survives network switches); Checking on long-running agent work away from the desk.
- **Speedrift** (speedrift, skill, beta, open) — Workgraph-first autopilot for coding agents: dependency-aware execution graphs with drift checks (code, spec, scope) that keep implementation, spec, and intent synced without hard blocks. Two installed skills: the speedrift driver and rifts, a compatibility alias that routes to the same workflow. Author: Braydon. Recommended uses: Multi-task agent projects where the work falls out of sync with the spec; Turning a design spec into a verifiable execution graph.
- **Agent Designer** (agent-designer, skill, production, client-gated) — Methodology and skill files for designing expert ensembles: domain-expert personas coordinated through a fixed inherited method, for standing analysis panels or one-shot live casts. Author: Braydon. Recommended uses: Building a repeatable analyst panel for a domain you work repeatedly; Casting a small expert ensemble for an immediate decision.
- **PlanForge** (planforge, skill, beta, client-gated) — Planning and Workgraph compilation pipeline: discovery, decomposition, a critic pass, and module selection that turn a goal into an executable graph with tests and drift gates. Author: Braydon. Recommended uses: Front-end planning for agent execution graphs before code starts; Turning strategy work into scheduled, verifiable tasks.
- **Outreach Engine** (outreach-engine, skill, production, client-gated) — Drafting, sequencing, personalizing, staging, and QA discipline for outbound email, LinkedIn, investor, and BD outreach; process craft for outreach that reads human. Author: Braydon. Recommended uses: Founders and BD leads running multi-touch outbound with an agent; QA gates that stop embarrassing sends.
- **Prospecting Partner** (prospecting-partner, skill, production, client-gated) — Research, qualification, segmentation, and prioritization method for prospect, partner, investor, and candidate lists; the upstream half of the outreach pipeline. Author: Braydon. Recommended uses: Building ranked target lists with evidence instead of vibes; Pre-outreach research sweeps an agent can run repeatedly.
- **Golden Voice Packets** (golden-voice-packets, skill, beta, client-gated) — Craft system for writing and speaking in a named voice: voice packets distilled from edited drafts, with a learning loop that turns user corrections into reusable tone guides. Author: Braydon. Recommended uses: Executives who need an agent that writes in their actual voice; Building a house voice a whole team can execute.
- **Model-Mediated Development** (model-mediated-development, skill, beta, client-gated) — Design discipline for systems where an AI model owns the judgment calls (routing, ranking, extraction, verification) instead of hiding decisions in regexes, thresholds, and fallback code. Author: Braydon. Recommended uses: Teams shipping AI features that keep accumulating hard-coded heuristics; Architecture reviews of model-facing APIs and agent workflows.
- **Folio** (folio, applet, beta, client-gated) — A personal knowledge store with notes, daily notes, collections, and search behind a local app, CLI, and MCP server, installed and operated in your environment. Author: Braydon. Recommended uses: Replacing scattered notes with one queryable knowledge base an agent can read and write; Engagements that need a shared client knowledge store.
- **msgvault** (msgvault, applet, beta, open) — A local email-and-message archive service: synced mail on your own storage, a supervised daemon, and CLI and MCP search so agents can query message history without cloud APIs. Author: kenn-io. Recommended uses: Executives who want their message history queryable but self-hosted; Engagements needing email-derived intelligence with data staying on-prem.
- **PAIA Workboard** (paia-workboard, applet, beta, client-gated) — A Workgraph-backed task board for agent work: claiming, status, comments, planning, and handoff between humans and agent workers on one surface. Author: Braydon. Recommended uses: Coordinating a mixed human and agent workforce on real task queues; Giving an executive one pane over everything the agents are doing.
- **Meeting Tooling (paia-meetings)** (paia-meetings, applet, beta, client-gated) — Meeting capture and tooling service exposed to agents over MCP: calendar-aware meeting workflows, prep, and follow-through that run where the calendar lives. Author: Braydon. Recommended uses: Executives whose meeting prep and follow-up should run itself against their calendar; Engagements building calendar-aware agent workflows.
- **Standing Analyst Panels** (earthworks-panels, applet, beta, client-gated) — Recurring expert-ensemble panels (practitioner, buyer-voice, and adversarial seats) that pressure-test strategy, drafts, and workshop design before anything goes to a real audience. Author: Braydon. Recommended uses: Pressure-testing positioning, pricing, and drafts against buyer reaction; Recurring design review without hiring a standing committee.
- **Google Workspace CLI skill set** (gws-cli-skill-set, skill, production, open) — Thirty-five skills that drive the Google Workspace CLI across every Workspace surface: Drive, Docs, Sheets, Slides, Calendar, Chat, People, Tasks, Meet, Forms, Keep, Admin reports, Model Armor, cross-service workflow recipes, and the profile router that picks the right account before any call. The set builds on the open-source Google Workspace CLI, whose repo ships the tool and its agent skills. Author: Google. Recommended uses: Agent workflows over a company's Google Workspace: documents, sheets, calendar, and chat from one CLI; Reference architecture when an engagement needs Workspace-side automation design.
- **msgvault skill wrappers** (msgvault-skill-wrappers, skill, production, open) — Four wrapper skills that query an archived mail and message store from agent sessions: archive operations, message search, SQL analytics over volume, and attachment export. Three install from msgvault, the open-source archive service whose repo ships the service and its query skills; the fourth is a local operations reference for the archive it searches. Author: kenn-io. Recommended uses: Executive message history an agent can search without cloud APIs; Email-derived intelligence with the archive staying on-prem.
- **roborev skill family** (roborev-skill-family, skill, production, client-gated) — Eight review-workflow skills for the roborev review daemon: requesting commit and branch reviews, working the iterative fix loop, responding without code changes, addressing specific jobs, and closing findings. Wired to per-repo review configuration; the daemon is local tooling with no verified public distribution, so the set installs with an engagement. Author: Braydon. Recommended uses: Code review loops an agent runs on its own branches; Design review on material diffs before they merge.
- **tier routing skill** (tier-routing-skill, skill, production, client-gated) — Routes each turn of work to the right model tier before it starts: judgment and final calls stay on the strong model, volume generation drops to the fast tier, and anything that touches private data runs on local models only. Encoded as a disposition an agent applies to every task. Author: Braydon. Recommended uses: Keeping strong-model spend on judgment instead of bulk generation; Private data work routed to local-only models by default.
- **host and fleet runbook skills** (host-fleet-runbooks, skill, production, client-gated) — Four runbook skills that recover and orchestrate an agent fleet: post-reboot workspace recovery, discovering and steering coding agents across terminal panes, and two coordinator playbooks for multi-agent work with mandatory verification. Authored against real hosts, so they install with an engagement and get rewritten for yours. Author: Braydon. Recommended uses: Recovering an agent workspace after a reboot or OS update; Orchestrating multiple coding agents in terminal panes with verification gates.
- **folio skill wrapper** (folio-skill-wrapper, skill, production, client-gated) — Skill wrapper that drives the Folio CLI for notes, daily notes, collections, and search from agent sessions. The Folio applet entry describes the knowledge store behind it; the wrapper installs alongside a Folio setup. Author: Braydon. Recommended uses: Note operations and knowledge lookup from any agent session; Daily-note and collection workflows against a Folio store.
- **supernote skill** (supernote-skill, skill, production, open) — Drives Supernote Cloud notebooks from agent sessions: list and read pages, append, move, and plan notes, with auth handled by the bundled CLI. Open source; the repo builds both the CLI and the skill that runs it. Author: Braydon. Recommended uses: Filing and reading handwritten notes from an agent session; Turning a Supernote into a queryable part of the workflow.
- **obsidian-vault-intelligence** (obsidian-vault-intelligence, skill, production, client-gated) — Legacy vault access, migration, and debugging for Obsidian-based stores. Useful when an engagement starts on an Obsidian vault and consolidates onto a queryable knowledge store. Author: Braydon. Recommended uses: Migrating an existing Obsidian vault without losing structure; Debugging vault plugins and link rot during a consolidation.
- **go-minitrace-transcript-analysis** (go-minitrace-transcript-analysis, skill, production, client-gated) — Analyzes archived coding-agent transcripts with go-minitrace: find sessions by repository and date, attribute implementation work to a session, query normalized transcript tables, and report evidence with explicit caveats. Reads local transcript stores, so it deploys per engagement. Author: Braydon. Recommended uses: Attributing implementation work to past agent sessions; Comparing agent behavior across tools with cited evidence.
- **paia skill wrappers** (paia-skill-wrappers, skill, production, client-gated) — Two wrapper skills for the workboard and program surfaces: one covers task-board operations (claiming, status, comments, handoff), the other covers research-backed program management with approval gates and execution ledgers. The Workboard and program applet entries describe the services; the wrappers install with them. Author: Braydon. Recommended uses: Working a task graph from any agent session; Program management with model-authored records and review gates.
- **typesafe-ai skill** (typesafe-ai, skill, production, open) — Third-party skill for building with TypeSafe System One models: typed judgments with calibrated probabilities, used in product code like programming primitives for routing, ranking, extraction, and verification. MIT-licensed skill file; the platform and docs are the upstream. Author: TypeSafe. Recommended uses: Designing AI features where judgment calls become typed decisions; Cookbook patterns for routing, ranking, and verification with Jev.
- **use-railway** (use-railway, skill, production, open) — Third-party skill for operating Railway infrastructure from agent sessions: projects, services, databases, buckets, deploys, infrastructure as code, environments, domains, metrics, feature flags, recovery, and the account signup and MCP paths. Author: Railway. Recommended uses: Provisioning and deploying client infrastructure through an agent; Infrastructure work that should stay conversational instead of dashboard-driven.
- **laya skill** (laya-skill, skill, beta, open) — Local runtime skill for Laya, the open-source non-autoregressive System 1 decision engine: choice, score, and noul primitives that turn natural language and application state into typed judgments with calibrated probabilities, running locally. The Laya repo is the upstream; the model entry lists the weights. Author: Braydon. Recommended uses: Fast local classification, gating, and triage steps without an API call; The open build path when a judgment step must run offline.
- **herdr fleet skill** (herdr-skill, skill, production, open) — Operating skill for herdr, the open-source pane-based agent multiplexer: dispatch and briefing templates, push-based completion protocol, review gates, and a verification ladder for the worker pipeline. The upstream ships the multiplexer and its own agent skill; this one is the fleet doctrine I run on top of it. Author: Braydon. Recommended uses: Coordinating worker agents in herdr panes from inside a session; The dispatch, review, and verification loop for a pane fleet.
- **r (Subspace review)** (subspace-review-skill, skill, production, open) — Routes one readable Markdown file into a focused review surface in a terminal split, collects anchored comments and suggested edits from a human reviewer, and returns the validated result to the agent. Built on the Subspace TUI through herdr or tmux. Author: spacedock-dev. Recommended uses: Human review of agent-written documents without leaving the terminal; Collecting anchored feedback a model can act on.
- **keynote-slides** (keynote-slides, skill, production, client-gated) — Builds single-file HTML slide decks: brand-ready templates, minimal navigation, a narrative engine with proven storytelling structures, and generated media for slide visuals. Used for client decks and workshop material. Author: Braydon. Recommended uses: Client-ready decks an agent drafts and revises; Workshop and talk material in one portable file.
- **white-paper** (white-paper, skill, production, client-gated) — Builds single-file HTML white papers with print-optimized layout, narrative arcs, communication frameworks, and SVG diagrams for research, strategy, and thought-leadership content. Author: Braydon. Recommended uses: Research reports and strategic deliverables that print clean; Long-form positioning an agent structures and drafts.
- **perplexity-research** (perplexity-research, skill, production, client-gated) — Web research through the Perplexity API: authentication, search strategy, and structured output patterns an agent can interpret, so research sweeps return evidence instead of prose. Author: Braydon. Recommended uses: Research sweeps that need citations and structure; Background briefs assembled inside a working session.
- **github skill** (github-skill, skill, production, client-gated) — Bash-script wrappers over the GitHub API: issues, pull requests, repository search, and state updates an agent runs without hand-writing API calls. Author: Braydon. Recommended uses: Issue and PR workflows driven from agent sessions; Repository research across an organization.
- **linear skill** (linear-skill, skill, production, client-gated) — Bash-script wrappers over Linear's GraphQL API: issue creation from conversations, work-item tracking, and project status queries. Author: Braydon. Recommended uses: Turning conversations into tracked work items; Project status queries without leaving the session.
- **root-cause-investigation** (root-cause-investigation, skill, production, client-gated) — Investigation method for incidents, regressions, recurring problems, and surprising metrics: evidence collection, causal chains, and postmortem discipline an agent runs before anyone proposes a fix. Author: Braydon. Recommended uses: Incident reviews that end in root cause, not blame; Recurring problems investigated systematically instead of patched again.
- **shell-scenario-panel** (shell-scenario-panel, skill, production, client-gated) — Scenario planning with worldview elicitation, external scenario construction, and actor-relative impact analysis, documented so decisions under uncertainty get worked through against several futures instead of one forecast. Author: Braydon. Recommended uses: Strategy decisions that deserve more than one future; Preparedness and positioning work with explicit assumptions.
- **tech-debt** (tech-debt, skill, production, client-gated) — Finds and removes dead code, duplicates, and quality issues with archive-based safety: multi-signal detection, then a two-stage review before anything is deleted. Author: Braydon. Recommended uses: Codebase cleanups an agent runs with a safety net; Paying down cruft before a handoff or audit.
- **webwright** (webwright, skill, production, client-gated) — Scripted browser automation for real web tasks: form filling, multi-step flows, data extraction, and reusable automation scripts when a task needs a real browser. Author: Braydon. Recommended uses: Web workflows too structured for manual clicking and too real for an API; Reusable automation an agent writes and reruns.
- **wg-cli** (wg-cli, skill, production, open) — Drives the Workgraph task system from the shell: the verified verbs for listing, claiming, logging, completing, and failing tasks, plus inbox coordination and daemon checks. wg is open source; the upstream points at its repo, and the Workgraph bridge entry covers the in-harness tools. Author: Braydon. Recommended uses: Task-graph operations from any terminal or agent session; Agent coordination through task inboxes.
- **workspace-cleanup** (workspace-cleanup, skill, production, client-gated) — Intelligent workspace cleanup using multi-signal detection (similarity, timestamps, references) to identify clutter, with a two-stage safety review before archiving anything. Author: Braydon. Recommended uses: Reclaiming project directories without deleting the wrong thing; Pre-handoff tidy-ups with an audit trail.
- **gemini-cli-extension-porting** (gemini-cli-extension-porting, skill, production, client-gated) — Ports agent workflows into Gemini CLI extensions: extension layout, context files, command patterns, and guardrails enforced with tests, CI, and pre-commit hooks. Author: Braydon. Recommended uses: Moving a proven agent workflow onto a second harness; Multi-harness delivery when a team runs mixed tooling.
- **reminders-management** (reminders-management, skill, production, client-gated) — Manages Apple Reminders through MCP: list organization, reminder creation and updates, search, and tie-ins to productivity workflows. Author: Braydon. Recommended uses: Personal task capture an agent handles end to end; Keeping reminder lists organized without manual sorting.
- **grok-aurora-cli skill** (grok-aurora-cli-skill, skill, beta, client-gated) — Multi-provider image and video generation workflows through the grok CLI: make commands, provider-specific generation, async task tracking, clip stitching, and output evaluation. Wired to generation APIs through local tooling with no public repo, so it installs with an engagement. Author: Braydon. Recommended uses: Campaign and deck visuals generated in batch; Video clips assembled from generated segments.
- **The Collective** (collective, applet, beta, client-gated) — Intent-accountable collectives of model-based intelligence: one shared intent per collective, live headless agent teams running as persistent sessions, and a light operator plane with explicit grants and a human-only kill switch. The fleet builds and improves itself against qualities the operator has written down. Author: Braydon. Recommended uses: Standing up a client's own intent-accountable agent collective; Engagements that need persistent, self-directed agent teams with explicit authority boundaries.
- **Z.ai** (provider-zai, provider, production, open) — API home of the GLM model family. My standard plane runs here: GLM-5.2 at high reasoning as the task-graph default, with GLM-5-turbo as the configured fast tier. Author: Z.ai. Recommended uses: The default model plane for dispatched agent work; High-reasoning coding and writing models at subscription rates.
- **LunaRoute** (provider-lunaroute, provider, production, open) — My default model gateway: one subscription account serving the whole hosted roster, from GLM base, vision, and flash variants to DeepSeek, kev-4b, and the image models, with background serving modes for long jobs and an account-wide catalog that syncs into the agent harness. Author: LunaRoute. Recommended uses: One subscription surface for text, vision, and image models; Background serving variants for unattended generation jobs.
- **Moonshot AI** (provider-moonshot, provider, production, open) — Direct API route for the Kimi model family; Kimi K3 is in my interactive rotation for coding-heavy sessions. Author: Moonshot AI. Recommended uses: Coding-strong alternate models when the primary plane is busy; Long-context reading and editing in the Kimi family.
- **Ollama** (provider-ollama, provider, production, open) — Local open-weight model runtime. I keep a local lane on it for private, offline inference alongside the hosted planes. Author: Ollama. Recommended uses: Private inference where data stays on the machine; Offline and air-gapped agent work.
- **GLM-5.2** (model-glm-5-2, model, production, open) — Z.ai's GLM-5.2, my standard work plane. Task-graph workers and dispatched agent runs execute on it at high reasoning by default; the gateway adds a vision serving of the same model, catalogued on its own row. Author: Z.ai. Recommended uses: Default execution model for agent task graphs; Long coding sessions that need steady high-reasoning output.
- **GLM-5.2 Vision** (model-glm-5-2-vision, model, production, open) — The gateway's vision serving of Z.ai's GLM-5.2: the same model with image input, on the LunaRoute plane rather than the Z.ai direct route. Flex and background variants cover bursty work and unattended work. Author: Z.ai. Recommended uses: Image input while staying on the GLM-5.2 line; Screenshot and diagram reading inside coding sessions.
- **GLM-5-turbo** (model-glm-5-turbo, model, production, open) — Z.ai's GLM-5-turbo, configured as the task-graph fast tier for high-volume generation where throughput beats depth. Author: Z.ai. Recommended uses: Bulk drafts and research sweeps at low cost; Fast turnaround steps inside multi-model workflows.
- **GLM-5.3** (model-glm-5-3, model, production, open) — Z.ai's GLM-5.3 family, the strong tier of my interactive plane. The default interactive model is the vision variant (glm-5.3-vision), with the base model and flash variant one selection away. The rotation serves it through the gateway, background variant included, with the Z.ai direct route enabled as the alternate. Author: Z.ai. Recommended uses: Daily-driver interactive sessions, including image input; Strong-tier default when a session needs more than the standard plane.
- **GLM-5.3-flash** (model-glm-5-3-flash, model, production, open) — The flash variant of GLM-5.3: the weak/fast tier of the interactive plane for quick questions and short edits between strong-tier calls. Served through the gateway with a background variant, and enabled direct on Z.ai as the alternate. Author: Z.ai. Recommended uses: Quick interactive round-trips while working; Short edits and lookups inside a larger session.
- **Kimi K3** (model-kimi-k3, model, production, open) — Moonshot AI's Kimi K3, routed direct from Moonshot's API; the coding-strong alternate in my model rotation. Author: Moonshot AI. Recommended uses: Coding-heavy sessions as a second engine; Long-context reading and editing.
- **DeepSeek-4.1-flash** (model-deepseek-4-1-flash, model, production, open) — DeepSeek's flash model as served through my gateway, including a background variant for unattended jobs. Author: DeepSeek. Recommended uses: High-volume extraction and classification sweeps; Background jobs that finish on their own.
- **HiDream-O1** (model-hidream-o1, model, production, open) — Open image-generation model, served through the LunaRoute image-generation MCP as hidream-o1; defaults to 2048×2048 output for large-format work. Author: HiDream-ai. Recommended uses: High-resolution generation from text prompts; Cover art and hero visuals.
- **FLUX.2 [klein]** (model-flux2-klein, model, production, open) — Black Forest Labs' klein-sized FLUX.2, served through the LunaRoute image-generation MCP as flux2-klein; tuned for few-step generation (my default is four steps). Author: Black Forest Labs. Recommended uses: Fast iteration on generated visuals; Batch generation where volume is the point.
- **Qwen-Image** (model-qwen-image, model, production, open) — Alibaba's Qwen-Image line, served through the LunaRoute image-generation MCP as qwen-image-2512, with an edit-capable variant (qwen-image-edit-2511) for modifying existing images by instruction. Author: Alibaba. Recommended uses: Prompt-faithful image generation; Editing uploaded images without a redraw.
- **Laya** (model-laya, model, beta, open) — Open-source (Apache 2.0) System-1 judgment model that turns natural language and application state into typed judgments with calibrated probabilities. I run it locally on my GPU as the open counterpart to the commercial judgment APIs. Author: NandhaKishorM. Recommended uses: Fast local classification, gating, and triage decisions; Judgment steps that must run without an API call.
- **Jev** (model-jev, model, production, open) — TypeSafe's flagship System One model: typed answers with calibrated probabilities over a commercial API, instead of generated text. The commercial counterpart to open-weights Laya, and the model kev-4b is benchmarked against. Author: TypeSafe. Recommended uses: Programmable common sense: routing, ranking, extraction, and verification steps in product code; Judgment calls that need an API and a service level, not a local GPU.
- **kev-4b** (model-kev-4b, model, production, open) — Kev-4B, Jared Palmer's open (Apache-2.0) System-1 decision model: a LoRA plus pointer head on Qwen3.5-4B-Base that serves the TypeSafe /v1/systemone contract and is publicly benchmarked against Jev (it is a Jev-class model, not Jev itself). My LunaRoute account serves it under the same id for arbiter assignment in planning runs. Author: Jared Palmer. Recommended uses: System-1 arbiter assignment in planning pipelines.
- **qwopus36** (model-qwopus36, model, experimental, open) — Qwopus3.6-27B-v2-MTP, Jack Rong's multi-token-prediction fine-tune of Qwen3.6-27B, pulled as the Q4_K_M GGUF from Hugging Face via Ollama and served locally in my llama-lanes stack. Origin verified from the on-disk Ollama manifest; it is a downloaded community build, not one of mine. Author: Jack Rong. Recommended uses: Local experimentation in the interactive plane.
- **pi-subagents** (pi-subagents, extension, production, open) — Extension for the pi coding agent that adds bounded subagent delegation: parallel reviews, scripted multi-child workflows, async work, and worktree isolation. Author: Nico Bailon. Recommended uses: Teams on the pi harness that need safe parallel agent fan-out; Code review and research sweeps without context contamination.
- **context-mode** (context-mode, extension, production, open) — pi extension that keeps raw command output out of the model's context window. Code runs over big logs and files in a sandbox; only the derived answers enter the conversation. Author: Mert Köseoğlu. Recommended uses: Agents drowning in build logs, API responses, or large file reads; Long sessions where context rot degrades judgment.
- **pi-web-access** (pi-web-access, extension, production, open) — Web access for pi agents: multi-provider web search, URL fetching, GitHub repo cloning, PDF extraction, and YouTube and local video understanding, all behind one tool surface. Author: Nico Bailon. Recommended uses: Agent research sessions that need search and page reads without leaving the harness; Video, PDF, and repository intake inside a coding session.
- **pi-chrome** (pi-chrome, extension, production, open) — Lets pi drive your existing signed-in Chrome profile after explicit authorization, so agent browsing works with the sessions you already have. Author: tianrendong. Recommended uses: Agent browsing against sites you are signed into, with the consent gate explicit; Testing flows that depend on an authenticated browser profile.
- **pi-package-manager** (pi-package-manager, extension, production, open) — A web dashboard for browsing, installing, and removing pi packages, plus a /packages command inside the TUI. Author: JVKE. Recommended uses: Keeping the pi extension set visible and manageable from one surface.
- **@narumitw/pi-retry** (pi-retry, extension, production, open) — Retries the provider errors that lose a turn for nothing: empty-detail responses, Codex websocket limits, and stalled requests. Author: narumiruna. Recommended uses: Long unattended runs where a transient provider error should not end the job.
- **@ff-labs/pi-fff** (pi-fff, extension, production, open) — Fuzzy file and content search for pi, powered by the FFF engine: fast find across a repo from inside the session. Author: Dmitriy Kovalenko. Recommended uses: Large-repo navigation without shell gymnastics.
- **pi-compact-before-switch** (pi-compact-before-switch, extension, production, open) — Compacts context with the outgoing model before a switch to a smaller-window model, so the new model starts from a summary it can fit. Author: Eran Sandler. Recommended uses: Moving a session down from a large-context model without losing the thread.
- **@taylorsatula/pi-compaction** (pi-compaction, extension, production, open) — Checkpoint compaction for pi: structured engineering handoff summaries with model-selected cut points and full artifact tracking. Author: taylorsatula. Recommended uses: Long builds where a plain summary loses the engineering state.
- **@taylorsatula/pi-prepare** (pi-prepare, extension, production, open) — Read-only prepare mode for pi: clarifying questions, a structured plan for approval, tracked execution, and a clean-workspace dossier before implementation. Author: taylorsatula. Recommended uses: Front-loading questions and plan review before any file changes.
- **rpiv-todo** (rpiv-todo, extension, production, open) — A todo list for the model, rendered as a live overlay in the TUI that survives /reload and conversation compaction. Author: juicesharp. Recommended uses: Keeping multi-step agent work on rails across a long session.
- **rpiv-advisor** (rpiv-advisor, extension, production, open) — A second opinion the model can request from a stronger reviewer model before it acts. Author: juicesharp. Recommended uses: Cheap insurance on consequential edits: a stronger model reviews before the change goes in.
- **@lunaroute/pi-extension** (lunaroute-pi, extension, production, open) — The LunaRoute extension for pi: login, model sync, request attribution, and first-class web tools wired to my default gateway. Author: LunaRoute. Recommended uses: Keeping the gateway model catalog in sync with the harness; Attributing agent requests to the right workspace.
- **herdr-agent-state** (herdr-agent-state, extension, production, open) — The pi integration herdr installs and manages: it reports live agent-session state to the herdr server over the pane socket, which is what lets panes auto-resume with their sessions. Author: herdrdev. Recommended uses: Agent fleets in herdr panes that survive restarts with session state intact.
- **pi-memory** (pi-memory, applet, production, client-gated) — Memory layer for pi that injects persona and owner facts into agent sessions, so an agent arrives already speaking in the operator's voice and knowing the working context. Built on the paia memory system and delivered privately with engagements. Author: Braydon. Recommended uses: Agent sessions that need persona continuity and owner facts from the first message; Standing up the same memory-backed agent practice inside a client environment.
- **classify-request** (classify-request, applet, production, client-gated) — Builds pi-memory's local classify request to Ollama under a pinned contract: strict JSON, thinking off, and a capped context so the call finishes in seconds. Pure module with its own tests. Author: Braydon. Recommended uses: The judgment step of pi-memory's automatic memory loop.
- **facet-guards** (facet-guards, applet, production, client-gated) — Namespace guards for pi-memory's facet keys: the fixed list of valid namespaces plus a normalizer that strips the subject prefix the local classifier invents. Pure string logic with its own tests. Author: Braydon. Recommended uses: Keeping memory facets queryable instead of silently mistagged.
- **state-stage** (state-stage, applet, production, client-gated) — Triage for pi-memory's staged record writes: separates clean, idempotent, warn, and failed outcomes, keeps the real traceback, and logs to a file instead of the TUI. Pure module with its own tests. Author: Braydon. Recommended uses: Idempotent record staging that never interrupts a session with a false error.
- **bash-guard** (bash-guard, applet, production, client-gated) — Guards every bash tool call in a pi session: rewrites whole-command grep to ripgrep, injects tiered fallback timeouts when the agent sets none, and blocks recursive searches rooted at the workspace container. Author: Braydon. Recommended uses: Agent sessions that stop hanging on pathological shell commands.
- **llama-serve** (llama-serve, applet, production, client-gated) — Probes running llama-server lanes over HTTP and registers the live ones as pi models under the local/ provider namespace, with /llama status, refresh, start, and stop commands. Author: Braydon. Recommended uses: Local llama.cpp lanes as selectable models inside the harness.
- **ollama-tool-schema-fix** (ollama-tool-schema-fix, applet, experimental, client-gated) — Sanitizes the tool schemas in pi's outgoing provider requests so Ollama-served models accept them. The installed build is a diagnostic pass that logs each request to disk. Author: Braydon. Recommended uses: Mixing local Ollama models into pi without schema rejections.
- **runtime view switcher** (zz-runtime-view-switcher, applet, production, client-gated) — Switches the pi TUI between desktop and mobile chrome by terminal width or manual command, owns the footer widget slots, and clears the powerline status line on every view change so chrome never double-renders. Author: Braydon. Recommended uses: One pi console across desktop panes and phone screens.
- **session-sidebar** (session-sidebar, applet, production, client-gated) — A live-updating sidebar for the pi TUI with five tabs: session context and model history, an on-demand summary, the todo list, git state, and pane info. Author: Braydon. Recommended uses: A glance view of the session without leaving the editor pane; Watching context, todos, and git while the agent works.
- **markdown-editor** (markdown-editor, applet, beta, client-gated) — A side panel for the pi TUI that renders markdown as formatted terminal output, so you review what the agent wrote without leaving the terminal. /md opens the most recent markdown the session touched; a gitignore-aware browser picks any other. Author: Braydon. Recommended uses: Reviewing agent-written docs in place; Reading long markdown without a context switch.
- **tmux-panel** (tmux-panel, applet, production, client-gated) — My tmux pane switcher for the pi TUI: a preview-first picker with flip navigation, pane links, visible spawning, and a global tmux manager agent. Author: Braydon. Recommended uses: Driving a tmux agent fleet from inside one pi session.
- **herdr-panel** (herdr-panel, applet, production, client-gated) — My herdr pane switcher for the pi TUI: a preview-first picker with flip navigation, pane links, visible spawning, and a global herdr manager agent. Author: Braydon. Recommended uses: Driving a herdr agent fleet from inside one pi session.
- **pi-bare** (pi-bare, applet, production, client-gated) — Bare completion mode for slow local base models: strips the pi request to just the latest message, with tools and history removed, so a slow prompt pass finishes before the watchdog aborts. Author: Braydon. Recommended uses: Wringing answers out of big local models on modest hardware.
- **pi-privileged-execution** (pi-privileged-execution, applet, production, client-gated) — Runs one approved shell command with administrator privileges after explicit consent: sudo authentication, a deletion-classification gate, and no password in the agent's hands. Author: Braydon. Recommended uses: Global or destructive operations that need root without handing over a shell.
- **pi-worksgood** (pi-worksgood, applet, production, open) — The Workgraph bridge for pi: exposes the wg task verbs (claim, log, done, fail) as agent tools and adds a graph widget so the task board is visible inside the session. Author: graphwork. Recommended uses: Agents that work the task graph under the same gates I use.
- **orca-agent-status** (orca-agent-status, applet, production, client-gated) — Orca-managed integration installed by the Orca app: posts agent status snapshots (session, model, and tool events) to Orca's local hook receiver so the ops view stays current. Author: Orca. Recommended uses: Agent status inside the Orca ops surface.
- **orca-prefill** (orca-prefill, applet, production, client-gated) — Orca-managed integration: when Orca hands the pane a prefilled prompt, this drops that text into the pi editor at session start. Author: Orca. Recommended uses: Composing a prompt in Orca and sending it to the pane clean.
- **orca-titlebar-spinner** (orca-titlebar-spinner, applet, production, client-gated) — Orca-managed integration: animates a braille spinner into the terminal title while the agent works and restores the session title when it stops. Author: Orca. Recommended uses: Telling busy panes from idle ones at a glance.
