35% cheaper, same quality

Meethstack

Small models do the work. Frontier models direct it. Agent skills that turn a cheap model into your engineering team and a frontier model into its senior engineer.

View on GitHub
six engine modules
82/82 hidden tests
  1. Wonder · frontier · 0:34
    Intake. Implement the six engine modules exactly as BENCH-SPEC.md says. Done when tsc and vitest pass. I'll review exact strings, diff hunks and cron weekday rules.
  2. Spark lead · 1:14
    Plan: 6 sub-agents in parallel, one module each. Tests first, one named test per spec example. Then I integrate and run an independent reviewer. Approve?
  3. Wonder · frontier · 1:37
    APPROVE
  4. 6 builders working · 141 tests written · reviewer found 2 bugs
  5. Spark lead · 11:12
    Done. The reviewer caught a template lookup bug and a cron weekday mismatch. Both fixed, with 4 regression tests. tsc clean · 141/141 passing.
  6. Wonder · frontier · 11:35
    Ran the checks myself. APPROVE

How it works

The frontier decides why and what. The small model works out how.

  1. Frontierstep 1

    Writes the intake

    About 20 lines: why, acceptance criteria, constraints, and the risks it will check. No spec, no code.

  2. Small modelstep 2

    Does the engineering

    Writes the spec, plan and tests, builds alone or with 1 to N sub-agents, and runs an independent reviewer that tries to break it.

  3. Frontierstep 3

    Approves at checkpoints

    Wakes only when the plan or the result is ready, and answers APPROVE or a few lines of REVISE. Waiting costs nothing.

hstack workflow: the frontier writes the intake, the small-model lead plans and builds with sub-agents and an independent reviewer, and the frontier approves at the plan and done checkpoints.

Benchmark 1 · six engine modules

Same task, same hidden tests. A third less spend.

The small model wrote 95% of the output for 10 cents. Exact token counts, including every sub-agent, and a hidden 82-test suite neither side saw.

Cost to build the same task
Frontier alone
Muse Wonder 1.0
$0.00
hstack
Wonder directs · Spark builds
$0.00
−35%
cost · $1.59 vs $2.46
100%
quality · 82/82 hidden tests, both

The skills

Three skills today. A skill bank next.

hstackfrontier

The senior engineer: intake, plan review, terse directives, final review, token discipline.

hstack-leadsmall model

Owns the task end to end: spec, plan, fan-out, tests-first build, independent review, checkpoints.

hstack-practicesmall model

Engineering habits small models lack: exact specs, state tables, real-environment facts, security, honest reports.

Skill bank coming next

Curated skills that make the small model better on its own: testing, security, frameworks, data and UI quality. The more it knows, the less the frontier has to correct.

Install in one line

The installer asks for your agent, your frontier model and your small model, then sets everything up. Muse Code today, more agents later.

Installcurl -fsSL https://hstack.shellhaki.dev/install | sh
From sourcego install github.com/shellhaki/hstack/cmd/hstack@latest && hstack
  1. 1. Run the installer and pick your models.
  2. 2. Start your agent on the frontier model and run /hstack-activate.
  3. 3. Give it a real task. The small model does the work; the frontier directs.