By radishbuild
Personal local-first agent: openspec inspired four-phase workflow loop (Explore/Design/Apply/Verify), organic memory, resource provenance, all in a git-versioned vault. Ships skills, slash commands, and a path-scoping hook.
Check a radcad vault for structural inconsistencies — AGENTS.md presence/age, broken MEMORY.md wikilinks, orphan memory files, orphan job folders (missing plan.md), TODO.md entries that don't resolve, and resources missing required frontmatter. Report findings; suggest fixes but do not apply them without explicit user approval.
Scaffold a new radcad vault at the given path — creates the directory tree, writes AGENTS.md/MEMORY.md/TODO.md from the bundled templates, runs git init, and prints next-step instructions.
Use during the Apply phase of a Full-tier radcad job — execute the numbered work tasks in plan.md's Tasks section (not the V-prefixed verify tasks), check boxes immediately after each completion, write generated artifacts under resources/generated/, and pause on ambiguity or plan-wrong discoveries. Invoke after Design has written plan.md.
Use during the Design phase of a Full-tier radcad job — create the job folder and write plan.md, the single job artifact containing Acceptance criteria + Tasks (work tasks + V-prefixed verify tasks) + optional Decisions. Invoke after Explore produces enough clarity to name at least one acceptance criterion. Apply never starts without plan.md already written by this phase.
Use during the Explore phase of a Full-tier radcad job — read memory lazily, pull sources into resources/fetched/, confirm intent, and STOP before writing any job artifact. Invoke when starting a new Full job or when the user's request requires clarification before planning.
Use to write, read, or reason about entries in the vault's memory system — flat MEMORY.md index with wikilinks to organically-named memory/*.md files. Invoke when the user mentions a durable fact worth keeping, when you're considering saving a lesson at the end of Verify, or when you need to check whether something is already remembered.
Use when saving files under resources/fetched/ or resources/generated/ — writes the required YAML frontmatter (source/fetched_at/fetched_by for fetched, generated_by/generated_at/purpose for generated), uses .meta.yaml siblings for binaries, and never creates a central resources/index.md. Invoke when pulling an external source or producing a generated artifact.
Modifies files
Hook triggers on file write and edit operations
Own this plugin?
Verify ownership to unlock analytics, metadata editing, and a verified badge. GitHub access is read-only (username + org membership).
Sign in to claimOwn this plugin?
Verify ownership to unlock analytics, metadata editing, and a verified badge. GitHub access is read-only (username + org membership).
Sign in to claimBased on adoption, maintenance, documentation, and repository signals. Not a security audit or endorsement.
Personal knowledge management — with spec-driven discipline.
radcad is a local-first AI agent for your own files, your own knowledge, your own work. It plans every non-trivial task through an Explore → Design → Apply → Verify loop borrowed from OpenSpec, records the trail as plain Markdown in a git-versioned vault, and stays out of the way on trivial asks. Your notes, your memory, your decisions — auditable tomorrow, diffable next year, browsable in Obsidian or any editor.
The pitch in one line: run your life with structure — propose, verify, remember — and get an agent that actually does it with you.
Prerequisites: Claude Code ≥ 2.1.89, Bun ≥ 1.3, a Pro/Max subscription (claude auth login) or ANTHROPIC_API_KEY.
The radcad GitHub repo ships both the CLI (TypeScript on Bun) and a Claude Code plugin marketplace. You can use either or both.
The CLI launches Claude Code with your vault's AGENTS.md as the full system prompt. The plugin adds skills, slash commands, and a path-scoping safety hook.
# 1. Clone radcad and link the CLI globally
git clone https://github.com/radishbuild/radcad.git
cd radcad
bun install
bun link # makes `radcad` globally callable
# 2. (Optional) Add the plugin marketplace (from your local clone, or once pushed,
# from GitHub directly — both work)
claude plugin marketplace add "$(pwd)" # local clone
# OR, once the repo is on GitHub:
claude plugin marketplace add radishbuild/radcad
# 4. Install the plugin
claude plugin install radcad@radishbuild
If you live in the Claude Code VS Code extension and don't want to use a terminal CLI, you can install just the plugin:
/plugins.radishbuild/radcad (or the local path to your clone).radcad and click Install.Then scaffold a vault manually (mkdir -p ~/radcad-vault/{memory,jobs,resources/{fetched,generated}} and copy the templates from plugin/templates/), or invoke /radcad:init ~/radcad-vault inside any Claude Code session once the plugin is loaded. Note the caveat in Use from VS Code below about system-prompt replacement.
npm / maintained marketplace — not live yet. Once published,
bun add -g radcad(ornpm i -g radcad) will replace the git-clone + bun-link step. The plugin marketplace flow above is already functional today.
# First time: scaffold a vault anywhere you like
radcad init ~/radcad-vault
# Start a session
radcad # uses $RADCAD_VAULT or ~/radcad-vault
radcad --vault ~/my-vault # or point explicitly
That's the whole thing. radcad reads your vault's AGENTS.md as the full system prompt and hands the session to Claude Code. You talk to it; it classifies each request, either acts directly (Light) or walks the four-phase loop (Full), and leaves a reviewable trail in your vault.
Try this first:
You: Plan a Saturday hike.
radcad: Classification: Full — ambiguous (where, how long, who?). Starting Explore.
<asks one clarifying question>
→ writes jobs/2026-04-21-plan-saturday-hike/plan.md
Open ~/radcad-vault/jobs/2026-04-21-plan-saturday-hike/plan.md in VSCode/Obsidian, watch it grow through the phases, git diff it whenever you want to see what the agent did.
plan.md that grows — scope, criteria, tasks (work + verify), decisions — through each phase.memory/travel/, memory/food/ preset folders. Memory filenames emerge from content; a flat MEMORY.md index gives you wikilink-and-descriptor glance-ability.AGENTS.md (the agent's operating rules) lives in your vault, not in the plugin. Edit it to tune the agent's behavior. Plugin updates never stomp your edits.npx claudepluginhub radishbuild/radcad --plugin radcadHarness-native ECC operator layer - 67 agents, 271 skills, 92 legacy command shims, reusable hooks, rules, selective install profiles, and production-ready workflows for Claude Code, Codex, OpenCode, Cursor, and related agent harnesses
Superpowers Plus core skills library for Claude Code: planning, execution routing, TDD, debugging, and collaboration workflows
Persistent file-based planning for AI coding agents. Crash-proof markdown plans (task_plan.md, findings.md, progress.md) that survive context loss and /clear, with an opt-in completion gate and multi-agent shared state. Manus-style. Works with Claude Code, Codex CLI, Cursor, Kiro, OpenCode and 60+ agents via the SKILL.md standard. Includes Arabic, German, Spanish, and Chinese (Simplified and Traditional).
Claude harness - A harness for solo developers (Vibecoders) to handle full-cycle contract development.
Core skills library for Claude Code: TDD, debugging, collaboration patterns, and proven techniques
Tools to maintain and improve CLAUDE.md files - audit quality, capture session learnings, and keep project memory current.