manzanas

A Mac daemon for multi-agent iOS simulator fleet orchestration: leases, actions, streaming, deterministic state, and an exportable run journal for AI agents (and humans) sharing simulators.

The 90% story: lease a simulator, run a YAML spec, get evidence. Everything else, fleets, physical phones, mirroring, is an opt-in chapter you can ignore until you need it.

https://github.com/user-attachments/assets/983b0548-df1a-41a3-812c-f0f39cfaa01c

1 orchestrator, 7 Codex agents, 7 simulators across 3 Macs, leases mean nobody trips over anybody. Real time, no edits.

manzanasd runs on each Mac host and owns everything stateful, the simulator registry, the lease table, the warm pool, action backends, streamers, golden images, and the run journal. Clients (manzanas CLI, MCP facade, SDKs) are thin and cross-platform, speaking a versioned JSON protocol over HTTP + WebSocket.

Website: manzanas.dapsdev.dev (source in site/).

Five-minute quickstart

No flags to read, no config to write. Four layers, start at whichever matches your machine:

0. One-liner (macOS/Linux): npx

npx manzanasd-client@latest doctor   # downloads the release binary and runs the setup diagnosis

The npm wrapper fetches the right manzanasd/manzanas binary for your platform on install, then proxies every command to it. Use it as the CLI, or point your agent's MCP config at manzanas mcp.

1. Anywhere (no Mac): mock mode

make build                      # writes bin/manzanasd, bin/manzanas, bin/manzanas-broker
./bin/manzanasd --mock          # full daemon + fake fleet + mock action backend

In another terminal, run the whole lease → boot → act → evidence loop as one declarative YAML run:

# hello.yaml, drives the mock login screen (docs/mock.md)
name: hello
target:
  labels: [ios26]
steps:
  - action: type_into_element
    with: {id: username, text: agent}
  - action: type_into_element
    with: {id: password, text: pw}
  - action: tap_element
    with: {label: "Sign In"}
  - action: wait_for_element
    with: {label: "Welcome, agent!", timeout_ms: 5000}

./bin/manzanas run hello.yaml -o evidence.md
cat evidence.md                 # PR-ready markdown evidence of every step

That's the product: one spec, one command, a journaled evidence trail. The same spec runs unchanged against real simulators, and against physical phones, because every target speaks the same actions API.

2. On a Mac with Xcode: real simulators

brew tap baribarigood/tap https://github.com/BariBariGood/homebrew-tap
brew trust baribarigood/tap      # Homebrew >= 6 requires trusting third-party taps
brew install manzanasd            # daemon (+ pulls in the manzanas CLI)
brew services start manzanasd     # launchd service on port 7433
manzanas doctor                   # one-shot setup diagnosis; every failure names its fix

manzanasd needs zero flags for a working single-Mac daemon: sensible defaults for the journal, template library, and element map, simulators enumerated from simctl, physical devices and the mirror off until you opt in. The same manzanas run spec.yaml now drives real simulators. Step-by-step walkthrough (lease, tap, screenshot, release by hand): docs/quickstart.md.

manzanas targets                                        # list simulators
manzanas lease acquire --labels ios26 --agent me --wait # claim one
manzanas tap 200 400 --lease lse_...                    # drive it
manzanas mcp                                            # or hand the tools to an agent

3. Optional chapters (you probably don't need these yet)

• Physical iPhones (WebDriverAgent): docs/devices.md
• XCTest-hostile apps via iPhone Mirroring: docs/mirror-onboarding.md
• Multiple Macs behind one endpoint: docs/broker.md

Why

Agents driving simulators over raw SSH + CLI tools trip over each other and pay huge fixed costs. manzanasd removes both, with measured numbers (M3 Pro, macOS 26.5, Xcode 26.5; reproduce with make bench):

• Leases, not locks: TTL-bounded exclusive claims with FIFO queues, no two agents ever drive the same sim.
• Park/thaw warm pool: idle sims are SIGSTOPped (a parked tree is unschedulable, ~0 idle host CPU no matter what the sim's daemons are doing) and thawed on lease grant: ~0.28 s lease-to-live vs ~7 s for a cold boot (~29 s first boot). The thaw itself is a cached-PID SIGCONT and takes under a millisecond.
• Warm actions: a resident per-sim helper makes an end-to-end tap ~36 ms vs ~950 ms cold (per-action AXe spawn), ~3 s cold on Intel.
• Deterministic state: snapshots, fixtures, per-lease auto-reset, and golden images that stamp out slimmed sims in seconds (~0.75 GB vs ~5 GB stock, via simslim), how one Mac runs a dozen sims.
• Evidence: every mutating op under a lease is journaled, with content-addressed artifacts and a PR-ready markdown export.

MCP (Claude Code, Cursor, Codex)

manzanas mcp serves the whole toolset over the Model Context Protocol (stdio), so any MCP-capable agent can lease and drive simulators, with per-session auto-release of leases and self-describing tool errors. Ready-to-paste client configs and troubleshooting: docs/mcp.md.

claude mcp add manzanas -e MANZANASD_ADDR=mac-host:7433 -- /path/to/manzanas mcp

Architecture

┌ Linux / CI / anywhere ────────────┐        ┌ each Mac host ─────────────────────────────┐
│ manzanas (thin client, Go)         │   WS   │ manzanasd (Go daemon)  :7433                │
│  - CLI: lease/tap/observe/...     │◄──────►│  registry ── warm pool (park/thaw, gates)  │
│  - MCP facade (stdio)             │  HTTP  │  leases (TTL, labels, FIFO, auto-reset)    │
│  - eval harness (manzanas-eval)    │        │  actions ── cold AXe / warm simbridge      │
├───────────────────────────────────┤        │  streams (MJPEG fan-out, browser view)     │
│ manzanas-broker  :7440             │        │  state (snapshots, fixtures, golden images)│
│  fleet-wide placement: leases are │───────►│  journal (evidence, artifacts, export.md)  │
│  scheduled across N daemons, then │ probe/ └────────────────────────────────────────────┘
│  clients talk to the owning       │ lease         × one daemon per Mac in the fleet
│  daemon directly (host_addr)      │
└───────────────────────────────────┘

One orchestration core, transports as thin backends: the same YAML spec, leases, journal, and evidence pipeline run against simulators (AXe/simctl), physical iPhones (WebDriverAgent), and XCTest-hostile apps (iPhone Mirroring). What manzanas is not, a generic OCR automation framework, an app-botting tool, is written down in docs/non-goals.md, along with why the repo stays one repo.

Documentation

Start here (the 90% path):

• docs/quickstart.md, zero to a passing run on one Mac: install → doctor → lease → tap → screenshot → YAML run.
• docs/runs.md, the one-call YAML run: schema, verdicts, drift detection, record → replay.
• docs/mock.md, the full daemon anywhere (Linux/CI), no Mac needed.
• docs/journal.md, the evidence trail: run journal format, artifacts, markdown export, replayable spec export.
• docs/mcp.md, handing the tools to an agent (Claude Code, Cursor, Codex).
• docs/troubleshooting.md, manzanas doctor first, then symptoms → causes → fixes.

Going deeper (read when you hit the subsystem):

• proto/PROTOCOL.md, the v0 wire protocol: targets, leases, actions, streams, state, journal, runs, WS surface. The tables in §5 are the authoritative list of action kinds and payloads.
• docs/architecture.md, components, interface contracts, and the product boundary.
• docs/agent-qa.md, a worked end-to-end agent QA session, plus driving animated (React Native) apps.
• docs/resolver-ladder.md, how element resolution works: map/ax/ocr/template rungs, provenance, escalation.
• docs/elementmap.md, the learned element map (fast-path resolution across runs).
• docs/templates.md, the pixel-template library (save_template / tap_template / assert_template).
• docs/warm-pool.md, park/thaw warm pool, footprint watchdog, host safety gates.
• docs/actions-warm.md, warm vs cold action paths; the resident simbridge helper.
• docs/state.md, snapshots, fixtures, per-lease auto-reset, quarantine.
• docs/images.md, golden images: slim once, stamp N sims in seconds.
• docs/streaming.md, MJPEG streaming and the browser view page.
• docs/dashboard.md, the built-in read-only web dashboard at /dash.
• docs/recording.md, per-lease video capture into the journal.
• docs/eval.md, the scenario-driven determinism benchmark harness.
• clients/README.md, manzanas CLI + MCP quickstarts; clients/npm/; GitHub Action in action/.

You probably don't need these yet (opt-in surfaces; each page opens with who it's for):

• docs/non-goals.md, the product boundary: what manzanas is, what it deliberately is not, and the one future split candidate.
• docs/devices.md, physical iPhones as leasable targets (devicectl + WebDriverAgent), and the mirror backend reference.
• docs/mirror-onboarding.md, the guided setup path for driving XCTest-hostile apps through iPhone Mirroring.
• docs/broker.md, multi-Mac federation.
• docs/fleet.md, running a multi-Mac fleet (topology, Tailscale, day-2 ops).
• docs/install.md, launchd install, releases, Homebrew formula.

The physical fleet this runs on (machines, locks, build caching) is site-specific; docs/fleet.md covers the daemon's part.

Status

Everything above is implemented and running on the fleet, leases + queues (incl. pause/resume), warm pool with safety gates, cold (AXe) + warm (simbridge) action backends, composite/batch/assert actions, the resolver ladder with element map and pixel templates, one-call YAML runs with verdicts and drift detection, MJPEG streaming, video capture, snapshots/fixtures/auto-reset, golden images, journal, dashboard, doctor, physical-device support (WDA + mirror), broker, eval harness, CLI + MCP. Verify with go build ./... && go vet ./... && go test ./... (Linux-safe; simctl paths are mocked).

License

Apache-2.0, see LICENSE. Releases up to and including v0.6.0 (and their existing tags) were published under MIT and remain MIT; later releases are Apache-2.0.