Quickstart

June 9, 2026 · View on GitHub

Install

Mac / Linux / WSL — one line

curl -fsSL https://raw.githubusercontent.com/waitdeadai/claudemax/main/install.sh | bash

Windows — PowerShell 5.1+ or 7+

irm https://raw.githubusercontent.com/waitdeadai/claudemax/main/install.ps1 | iex

Manual (advanced)

git clone https://github.com/waitdeadai/claudemax ~/.claudemax
cd ~/.claudemax
pnpm install
pnpm build
sudo ln -sf $PWD/packages/cli/dist/index.js /usr/local/bin/cmax

Remote-from-phone flow (Tailscale + tmux + ntfy + QR onboarding)

curl -fsSL https://raw.githubusercontent.com/waitdeadai/claudemax/main/setup.sh | bash

Authenticate

For Claude Max users (recommended):

# Just be logged into Claude Code interactive (claude --version should work)
# claudemax routes all provider calls through @anthropic-ai/claude-agent-sdk
# which bills against your Agent SDK credit pool (\$100 Max5x / \$200 Max20x).

Or API key fallback:

export ANTHROPIC_API_KEY=sk-ant-...
export CMAX_PLAN=api

First check

cmax doctor

You should see:

  • plan: max5x or max20x (auto-detected)
  • billing: subscription
  • credit: $100/mo or $200/mo
  • parallel cap with hardware + credit-aware bounds
claude plugin marketplace add waitdeadai/claude-plugins
claude plugin install llm-dark-patterns@waitdeadai-plugins

First run — ask and achieve

In any project directory:

cmax ask "add a /health endpoint that returns build sha and uptime; cover with a test"

That is the entire daily-driver UX. No flags, no model picker, no decomposition by hand.

The SOTA-2026 pipeline runs automatically:

  1. /deepresearch the topic. Sonnet collects sources in parallel; Opus synthesizes. Source ledger persisted to memory.research_sources.
  2. multispec decompose — Fable 5 (while included →2026-06-22; Opus after) authors N sub-Specs with a DAG of dependencies, rollup completion conditions, and per-sub-Spec write-sets.
  3. /specqa (parallel) — Haiku checks each sub-Spec has mechanically-checkable verifyHints.
  4. /introspect (parallel) — Opus rates confidence per sub-Spec; blocks at confidence < 6.
  5. Mode auto-selection — Mode A (SDK subagents in one query()) for ≤ 5 sub-Specs / short runs; Mode B (Claude Code Agent Teams with shared task list + worktree isolation) for big swarms or write-set overlap.
  6. Parallel /goal per DAG leaf — Sonnet executes by default; Opus on router escalations (security, novelty, complexity ≥ 7).
  7. Per-sub-Spec /verify (parallel) — blind Opus re-reads repo and runs each verifyHint.
  8. Rollup /verify — blind Opus checks the integration conditions across all sub-Spec outputs.
  9. Memory record + state snapshot. ntfy.sh push to phone if NTFY_TOPIC is set.

Throughout: bundled dark-patterns hooks (35 of them, wired by cmax init) block vibes, fake citations, aggregator hallucination, and credential leaks.

Power-user flags (same engine; cmax ask is the friendly entry point):

$\text{bash} \text{cmax} \text{run} "<\text{goal}>" --\text{variant} \text{opusolo} # \text{all}-\text{Opus} \text{exec} \text{for} \text{novel}/\text{security}/\text{auth} \text{work} (~3 \times \text{cost}, \text{max} \text{effectiveness}) \text{cmax} \text{run} "<\text{goal}>" --\text{mode} \text{teams} # \text{force} \text{Claude} \text{Code} \text{Agent} \text{Teams} (\text{Mode} \text{B}) \text{parallelism} \text{cmax} \text{run} "<\text{goal}>" --\text{no}-\text{research} # \text{skip} /\text{deepresearch} \text{for} \text{simpler} / \text{well}-\text{known} \text{goals} \text{cmax} \text{run} "<\text{goal}>" --\text{no}-\text{verify} # \text{skip} \text{independent} \text{verification} (\text{not} \text{recommended}) \text{cmax} \text{run} "<\text{goal}>" --\text{max}-\text{turns} 80 # \text{cap} \text{each} \text{sub}-\text{Spec}'\text{s} /\text{goal} \text{loop} $

This will:

  1. /deepresearch (web-current sourced research, persisted to memory.research_sources)
  2. multispec decompose (Fable 5 →2026-06-22, then Opus → N sub-Specs + DAG + rollup conditions)
  3. /specqa each sub-Spec (Haiku quality gate)
  4. /introspect each sub-Spec (Opus confidence/assumption gate)
  5. Auto-select Mode A or Mode B parallelism
  6. parallel /goal per DAG leaf (Sonnet executes by default; Opus on router escalations)
  7. per-sub-Spec /verify (parallel, blind Opus)
  8. rollup /verify (blind Opus against rollup conditions)
  9. Memory record + state snapshot
cd my-project
cmax taste init

Auto-derives taste.md + taste.vision from repo signals + /deepresearch on SOTA at current time. NO 10 questions. ~30–90 seconds.

Install skills into a project

cd my-project
cmax init                              # writes .claude/skills/* + .claude/hooks/*

Now inside Claude Code:

/cmax migrate the user model to drizzle, preserve all reads, ship tests

Inspect routing decisions

cmax route "refactor the auth middleware" --complexity 6 --domain auth
# → tier:opus, escalated:true (security domain)

cmax route "summarize 500 commit messages"
# → tier:haiku

cmax route "design the cache layer" --tier opus --cost-ceiling 2

Inspect memory

cmax memory search "auth migration"
cmax memory runs --limit 20

Common patterns

Cost-capped run

cmax run "<task>" --max-turns 80               # cap goal loop per sub-Spec
cmax doctor                                    # confirm budget headroom first

Force /opusolo (all-Opus, max effectiveness)

cmax run "<task>" --variant opusolo

Force Mode B (Claude Code Agent Teams) on a small task for testing

cmax run "<task>" --mode teams

Goal only (skip spec writing if SPEC.md exists)

cmax goal SPEC.md

Verify a prior run

cmax verify SPEC.md

Long-running overnight

cmax overnight SPEC.md --budget-credits 50
# SIGTERM-safe: re-run picks up from last checkpoint

Research a topic without running anything

cmax research "AI SDK v6 + Vercel AI Gateway + Claude Agent SDK integration patterns"

What can go wrong

  • cmax doctor shows plan: max5x source: default → auto-detect failed. Set CMAX_PLAN=max20x (or whatever you actually have) in your shell.
  • First cmax run fails with "no API key" → you're on the API path. Either set ANTHROPIC_API_KEY or claude login first.
  • Mode B sub-Specs fail with "command not found" → Claude Code CLI not on PATH; which claude to confirm. Fallback to --mode solo (Mode A).
  • /goal hits max-turns → either the SPEC is unbounded or the task genuinely needs more turns. Re-spec or raise --max-turns.
  • /verify returns partial or failed → iterate /goal against the failing conditions; re-verify. If two iterations don't move it, the SPEC is wrong — re-spec, don't re-grind.

What this is not

  • Not a chatbot wrapper.
  • Not a multi-provider abstraction.
  • Not autonomy without a SPEC.
  • Not "Opus for everything" or "cheap for everything." Effectiveness max, cost-aware.