kickoff

See it live →

A company that builds itself, steered from a pocket. Runs on Claude Code or OpenCode. Your org outlives your vendor.

Not a better prompt and not a single agent — a small, persistent system: one coordinator, a crew of specialists, memory that compounds across sessions, and a phone as the control plane. Give it one brief and it plans, builds, tests, reviews and reports while you steer in a sentence at a time.

         YOUR PHONE                            ANY ENGINE
    Telegram · one sentence              Claude Code · OpenCode
             │   ▲                              │
             ▼   │                              │  the SEAM:
     ┌───────────┴───────────────────┐           │  one supervision contract,
     │          COORDINATOR          │◀──────────┤  two interchangeable batteries
     │ decides · dispatches · relays │           │
     └───────────────┬───────────────┘           │
                     │ in parallel, own context  │
                     ▼                           │
        planner → builder → reviewer → deployer │

The leverage model

Level	Shape	Where the human sits
L0	prompt → output	you are the CPU
L1	agent → one task	hand off, then review
L2	system → whole domain	build the builder once, then steer

Parallel. Persistent. Async. Context-partitioned. Pocket-accessible. And the core is engine-agnostic — your org's memory, charters and history live in YOUR repos. Engines are replaceable components — we flipped the entire fleet mid-flight to prove it. Models are still someone's service; switching them costs one line, not a migration.

It drives agents — you approve the irreducible

The org runs everything reversible autonomously: scaffolding, building, testing, committing, pushing behind local gates, previewing over Tailscale. It stops for exactly three things:

• SPEND — anything that bills: "ready to ship X to Y, ~Z/mo — go?"
• DESTRUCTION — data loss you can't undo
• CREDENTIALS & PERMISSIONS — rotating secrets, granting access

Everything else: decide, run, report. One-line beats while working — silence reads as death.

The standard org (engine-neutral by design)

Adopt wires the union — everything both engines need, always on disk:

Concern	Lives in	Claude reads	OpenCode reads
Charter	CLAUDE.md → AGENTS.md	✓	✓
Crew charters	.claude/agents/ + .opencode/agent/	✓	✓
Memory recall	memory/ + MEMORY_DB	recall hook	memory_search tool
Commit attribution	native	native	plugin + git hook
Telegram bridge	per-org token + state dir	official plugin	opencode-telegram

Switching engines costs one line:

export WORKER_ENGINE=opencode   # or claude — in .kickoff/instance.env
kickoff up                      # supervisor cycles; nothing else moves

No porting. No re-adoption. No lock-in. Your memory, tracker, charters and history are the asset; engines are batteries.

Pick your battery

	Claude Code	OpenCode
Needs	Claude subscription/API	Node ≥ 22 (free-tier models available)
Bridge	official telegram plugin	opencode-telegram
Ships in core since	v0.1	v0.38

Either alone suffices; running both is belt-and-suspenders — when one vendor has a bad week, your fleet doesn't notice.

Quickstart

# 1 · install (pins a reviewed release; prints tag @ commit so you can audit)
curl -fsSL https://raw.githubusercontent.com/vinceferro/kickoff/core-v1.0/install.sh | sh

# 2 · adopt a project — installs the standard org, engine-neutral
cd ~/my-app && kickoff adopt

# 3 · wire your pocket — one Telegram bot per org, isolated by state-dir + token
kickoff setup

# 4 · pick your battery
echo 'export WORKER_ENGINE="${WORKER_ENGINE:-opencode}"' >> .kickoff/instance.env
kickoff up

# 5 · steer from anywhere
#   brief:      "plan and build X"        → beats arrive while it works
#   preview:    https://<box>.ts.net/<project>/<app>
#   boards:     http://<tailnet-box>:9001/mc/<project>/   (tailnet-only, token-gated)

Deep dive: QUICKSTART · RUNNING · ADOPT · TOOLING · the live site

What it realistically does

• Take an idea to a running, tested first slice — previewed from your phone within the hour
• Carry it toward production through human-approved deploys
• Run a multi-function org around it, each specialist under a least-privilege charter, all reporting into Mission Control
• Compound: session 40 is smarter than session 1 because memory compounded — not because the model changed

It does not write production software unattended, and it does not replace you. It replaces the part of you that context-switches.

---

Proof-of-work: this repository is developed BY its own system — releases v0.38→v0.39 were cut, gated, and fleet-deployed by coordinators running on both engines.