Something's compiling

CrewHaus is an open-source compiler that turns a single spec.yaml into a runtime-thin bundle for any agent target shape — CLI, channel bot, RAG pipeline, multi-agent crew, eval harness, voice/realtime agent, browser/computer-use agent, and more.

license apache-2.0 runtime typescript · bun launch june 2026

§ 02 / WHAT'S COMPILING public release · june 2026

spec.yaml → IR → emit(target) One input. Twelve output shapes today, more on the roadmap.
target shapes CLI · channel bot · RAG pipeline · multi-agent crew · eval harness · voice/realtime · browser-use · batch worker · autonomous research · stateful graph · managed multi-tenant · workflow.
license Apache 2.0.

name: hello-crew
target: crew
model: claude-sonnet-4-6
entry: researcher
permissions:
  mode: default
roles:
  researcher:
    instructions: |
      You are the RESEARCHER role in a multi-agent crew. When asked a topic,
      list 3 concise risks (1 sentence each) of the topic. After your list,
      call the `Handoff` tool with target="writer" and reason describing what
      the writer should produce. End your turn after Handoff returns. Do NOT
      try to write the final post yourself — that's the writer's job.
  writer:
    instructions: |
      You are the WRITER role in a multi-agent crew. The previous role
      (researcher) just handed control to you with a list of 3 risks. Write
      a single 4-sentence summary post about those risks. If you need a
      clarifying source check, use the `SendMessage` tool with target="critic"
      to ask a single short question and incorporate the critic's reply.
      Then end your turn with the final post.
  critic:
    instructions: |
      You are the CRITIC role in a multi-agent crew. When the writer asks you
      a question via SendMessage, give a one-sentence direct answer. Be terse.

name: hello-graph
target: graph
model: claude-sonnet-4-6
entry: plan
nodes:
  plan:
    instructions: |
      You are the PLAN node of a 3-node graph (plan → execute → summarise).
      Read the user's input from the Upstream state and produce a 3-bullet
      plan for how to address it. Return only the plan, no preamble.
  execute:
    instructions: |
      You are the EXECUTE node. Read the plan from the Upstream state's
      `plan` field and execute it: produce concrete findings, talking
      points, or evidence in 4–6 sentences. Stay grounded in the plan;
      do not free-associate.
    hitl:
      prompt: "Approve the plan and continue to execute?"
  summarise:
    instructions: |
      You are the SUMMARISE node. Read the plan and execute results
      from the Upstream state and produce a 2-sentence executive
      summary. Do not include lists or bullets — prose only.
edges:
  - from: plan
    to: execute
  - from: execute
    to: summarise

name: hello-workflow
target: workflow
model: claude-sonnet-4-6
permissions:
  rules:
    - { type: alwaysAllow, pattern: Bash }
steps:
  - name: list-files
    instructions: |
      Use the bash tool to run `ls -la` in the current directory.
      Report what you find in 1-3 short sentences (file types, sizes, anything notable).
      Do not editorialize beyond what the listing shows.
    tools:
      - bash
  - name: summarize
    instructions: |
      You will receive the previous step's directory listing as context.
      Write a single-paragraph summary describing what kind of project this
      directory contains based on the files you see. Do not call any tools.

name: hello-batch
target: batch
agent:
  model: claude-haiku-4-5-20251001
  instructions: |
    You are a batch worker. Each job's input is a short text prompt.
    Respond with a single concise sentence (≤ 20 words). Do NOT call
    any tools. End your turn after the one-sentence reply.
queue:
  adapter: in-memory
  visibilityTimeoutMs: 30000
  maxRetries: 3
  seedJobs:
    - "Say hello in one sentence."
    - "Name three primary colors in one sentence."
    - "Describe water in one sentence."
    - "Define recursion in one sentence."
    - "What is the capital of France? Answer in one sentence."
    - "Define entropy in one sentence."
    - "Describe a circle in one sentence."
    - "Name two oceans in one sentence."
concurrency: 4
idempotencyWindowMs: 60000
permissions:
  mode: default

name: hello-browser
target: browser
agent:
  model: claude-sonnet-4-6
  instructions: |
    You drive a chromium browser. Use the available tools to complete the
    user's task:
      - Screenshot() — capture the current viewport.
      - FindElement(description) — get a bounding box + center coords for
        a UI element by natural-language description.
      - Click(x, y), Type(text), Key(combo), Scroll(dx, dy) — drive the
        page (DESTRUCTIVE; require alwaysAllow).
    Be deliberate: take ONE screenshot, find the target element, click
    its center coordinates, take ONE more screenshot to verify, then
    end your turn with a 1-sentence summary of what you did.
driver:
  backend: chromium
  viewport:
    width: 1024
    height: 720
  startUrl: http://127.0.0.1:7325/
groundingModel: claude-sonnet-4-6
permissions:
  mode: default
  rules:
    - type: alwaysAllow
      pattern: Screenshot
    - type: alwaysAllow
      pattern: FindElement
    - type: alwaysAllow
      pattern: Click
    - type: alwaysAllow
      pattern: Type
    - type: alwaysAllow
      pattern: Key
    - type: alwaysAllow
      pattern: Scroll

§ 03 / GET NOTIFIED one email at launch · no marketing