Agent-Builder

A factory that turns reviewed templates into a runnable Claude Code agent crew. Browse dev-agent classes as 3D characters, tune each one, forge a portable .claude/ bundle, and drop it into any repo to run a coordinated, concurrency-safe multi-agent workflow — optionally wired to the enforced Valkyrie pipeline.

Open the Agent Crew Builder →

1 · Build it here

Browse the roster as real 3D characters, pick dev-agent classes (CTO, Reviewer, Implementer, Tester…), tune each one's stats, and give the crew a mission. Every class is a reviewed sc-compose template — you fill variables, you never hand-write agent prompts. Two skills do this: /crew-pick (assemble) and /crew-author (invent a new class). The builder UI is the friendly front-end.

2 · Run it in your repo

Forge produces out/<crew>/.claude/. Copy it into any repo and invoke the third skill, run-crew. Claude then orchestrates the team — dispatching agents, routing their results over a file blackboard, and honoring quality gates.

What's in the bundle

Real Claude Code subagents (.claude/agents/*.md), an orchestrator skill (run-crew), the coordination contract (PROTOCOL.md), a crew charter (CLAUDE.md), and a dependency-free crew-task CLI that scopes every run to its own task folder so concurrent or repeated runs never collide. Not a description of a team — a working team. Optionally a valk-config.mdbinds it to Valkyrie's DESIGN→PRD→ISSUES→TDD pipeline — omit it and Valkyrie behaves exactly as before (a proven no-op).

Why templates

An agent fills variable holes but can't restructure a reviewed class. So a crew is reproducible, auditable as one diff, and fails before runtime instead of drifting silently. The discipline is the product.

How it works

1 · Build → Run. You assemble a crew here; Forge writes a real .claude/ bundle; you run it in your repo (or Valkyrie).

Build · this app
pick classes + stats
/crew-pick
/crew-author · UI
templates = the contract
forgeout/<crew>/.claude/
Run · your repo / Valkyrie
cp .claude/ into a repo
invoke run-crew
Claude orchestrates them
a working crew

2 · Comms + workflow = a file blackboard. Subagents share no memory — only files. The orchestrator routes; agents read-before / write-after, each run scoped to its own tasks/<id>/ so concurrent or repeated runs never collide. That is both how they talk and how the workflow advances.

run-crew · the orchestrator
1 · read HANDOFF tail → pick next2 · Task-dispatch5 · route msg · honor “blocked”
.claude/handoff/
tasks/<id>/HANDOFF.mdper-run state
tasks/<id>/msg-a→b…scoped messages
tasks/INDEX.mdledger of runs
PROTOCOL.mdshared rules
3 · scoped brief ↓4 · status line ↑
agent · Task
read → work → write
isolated context

scoped per task — concurrent or repeated runs never collide

3 · In Valkyrie the workflow is enforced, not just suggested: valk-config.mdbinds your crew to the gated DESIGN→PRD→ISSUES→TDD pipeline — and a proven no-op when it's absent, so vanilla Valkyrie is byte-identical. Valkyrie on GitHub ↗

valk-config.md
DESIGN → cto, pm
PRD → pm
ISSUES → planner
TDD → impl…
DESIGN
cto + pm
PRD
pm
ISSUES
planner
TDD
impl + test + rev

🔒 each stage gate enforced by the valk hook

New here? Read SOP.md for the full architecture, or just open the builder and click a class.