Watch your coding agent work — and take the helm. A live, steerable board your agent keeps while it works. Glance at it; grab the wheel any time. 实时可操舵的单任务看板:agent 干活时你看着,随时改目标,它下一步就跟上。
English | 中文
The agent works → it hits a fork and needs you (the board turns red, you get a notification) → you type one line of steering on the board → it reads it and carries on (staging only). That loop — you change direction mid-task, the agent follows — is what helm is.
Most "watch your agent" tools are one-way mirrors: you can look, but if it drifts you wait until it's done. helm makes the board a steering wheel. You edit the goal or a steering note right on the page; the agent re-reads them at its next checkpoint (helm goal) and adjusts course. No reload, no flicker, no lingering process — and each update costs the agent only a few tokens, so the board stays honest minute to minute.
- A long task you want to watch without hovering. A migration, a refactor, a test-fix loop. You work on something else; a glance at the board tells you where it is and whether it's stuck.
- You want to correct course without interrupting. Heading the wrong way? Change the goal in one line ("staging only", "don't touch the schema"). No interrupt, no new prompt — it follows on the next step.
- It should call you, not the other way around. When it genuinely needs a decision, the board lights up "needs you", with a desktop notification and a 🔴 count in the tab title.
A live board on 127.0.0.1 (warm light, dark on toggle):
- Left, always in view — Now (current action, status pulse, progress) · Plan (steps as progress) · Your steering (editable goal + note, saved back to the agent) · Needs-you.
- Right, the stream — Activity timeline (color-coded, pitfalls included) · Decisions & assumptions (assumptions flagged, so you catch a wrong turn early) · Artifacts (links, images, data).
A real board's state lives in examples/sample-board.json.
helm is drop-in: zero runtime dependencies, runs on Node's stdlib. No npm install, no build step.
npx skills add chasey-myagi/helm # via skills.sh
# …or just drop the folder in:
cp -r helm ~/.claude/skills/ # Claude Code · Codex · any SKILL.md runtime
⚠️ Don't alias the CLI to barehelm— it collides with Kubernetes'helm. Call it by path (the SKILL.md does).
Just ask, in any of these shapes (English or 中文):
- "do X, and give me a live board to watch progress"
- "I'm stepping away — put up a board I can come back to, and let me change the goal"
- "这个迁移分几步,我想盯着、卡住了叫我"
- "keep me posted on a live board while you refactor this"
Tip: the surest trigger is to name the board — "give me a helm board", "做个 helm 看板". On a long multi-step task a capable agent will often offer one on its own; if it doesn't, just ask.
HELM=~/.claude/skills/helm/dist/helm.mjs
$HELM init --title "Migrate auth → JWT" --goal "Move sessions to stateless JWT, lose nobody"
$HELM plan "Back up users" "Add RS256" "Migrate sessions" "Cut over" "Regression"
$HELM step 2 # 1 done, 2 active, progress auto-set
$HELM event ok "users table backed up"
$HELM decide "RS256 over HS256" ; $HELM decide "staging first" --assumption
$HELM goal # checkpoint: read the human's edits, follow them
$HELM ask "Staging or prod first?" # a real blocker → notifies the human
$HELM done "All sessions migrated, tests green"Full command reference in SKILL.md.
| helm | work-canvas | vibe-kanban · agent-kanban | |
|---|---|---|---|
| Form | live board (SSE) | static HTML snapshot, finalized once | live board |
| Direction | two-way — you edit the goal, it follows | one-way, read-only | mostly task assignment |
| Scope | single task, one focus | single artifact | multi-agent fleet |
| Install | zero-dep drop-in | paste-to-install ritual | npx, runs a server |
Live boards exist. Static self-contained reports exist. Multi-agent fleet boards are a crowded field. But single task + you change the goal mid-run + the agent follows — that lane is helm's.
This repo was built while helm watched. The board in the demo above is the same one that tracked helm's own development — the plan, the decisions, the moments it needed a human call — from first sketch to this README. Dogfooded end to end.
- Writes only its own state. helm writes
<project>/.helm/<task>/state.jsonand serves a local read-only board. It never touches your source and makes no outbound requests. - No daemon, no resurrection. The board process exits on its own when nobody's watching or the task has gone idle. It never auto-starts and never restarts itself.
- Honest by design. The agent follows your steering at checkpoints (it calls
helm goal); it doesn't pretend to be interruptible mid-step. Assumptions are taggedassumptionso a wrong turn is easy to catch early.
helm/
├── SKILL.md how the agent uses it (triggers · workflow · commands)
├── dist/ runtime, committed, zero-dep: helm.mjs · server.mjs · board.html
├── src/ TypeScript source: types.ts (shared BoardState) · cli · server · board
├── examples/ sample-board.json · demo.gif · record-demo.sh (reproducible demo)
├── *.md PRODUCT · DESIGN · DEVRULES · ARCHITECTURE
└── build.mjs · package.json · tsconfig* (dev only — esbuild + tsc)
Edit src/, run npm run build to regenerate dist/. Design notes in DESIGN.md; the technical design in ARCHITECTURE.md.
- Browser, end to end: init → live SSE updates → edit steering on the page →
helm goalreads it → blocked notification → process self-exits. All verified. - Eval (skill-creator): 3 real multi-step tasks, with vs without the skill. With helm, board-quality assertions passed 100% (22/22) vs 13.7% without — and the agent's process variance dropped sharply (tokens ±0.5k vs ±9.4k).
MIT © 2026 chasey