case · drive DaVinci Resolve from any AI agent

§01 — Pitch

operates davinci resolve / open-source sdk + hosted api / macos · resolve 19 + 20

$pip install case-sdk copied

Claim 50 free runs · PH github →

works with Claude Code · Codex · OpenCode · OpenClaw · Hermes · any MCP client

FIG. 01 · DaVinci Resolve · live replay

click for sound

↓ scroll to drive the agent

§ · Structured failure

structured_failure()/ what your agent gets back when the click misses

Your agent crashed at 'element not found'.
case returns the typed code → your planner replays.

without case raw computer use · pyautogui · pixel-driven

$ python agent.py import_clip▌

Exception: Could not click

element at (847, 312)

at PyAutoGUI._click(x=847, y=312)

at runProcedure('import_clip', …)

at orchestrate(['select_dir','open','poll'])

… 18 more frames …

your agent: crashes

you: grep logs at 2:47am

with case typed enum · retry flag · runbook link

› await case.call("import_clip", { path: "/Users/me/clip.mov" })▌

→ resolved 142ms

{

ok: false,

code: "AX_TIMEOUT",

retry_safe: true,

repair_hint: "modal not painted",

recovery: "/runbook/ax_timeout"

}

your agent: retries → escalates

you: sleep

click → watch each agent resolve in real time

§02 — Surface area

catalog()/ click anything · receipts inline

Every procedure is one call.
Open one to see why it ships.

Each procedure carries a typed schema, an idempotency label, a structured failure shape, per-version reliability, and a recovery doc. 66 procedures live · 12 shown.

§03 — Path in

quickstart()/ one terminal

From install to working agent
in 10 minutes.

Three commands. One terminal. Working agent in 10 minutes.

brew install: tap and pull the harness binary.
case init: probe the target app, register typed procedures.
case demo: run a verified composite end-to-end.

Drop the MCP endpoint into Claude SDK, Anthropic CU, Hermes — or any tool-use framework.

macOS · Resolve 20.0+

$ brew install daemonlabs/tap/case

$ case init --target resolve

› detecting cua-driver............ ok

› probing Resolve 20.2 AX tree... ok

› registering 66 procedures...... ok

› mcp endpoint: stdio://case-resolve

$ case demo

› agent: "import 3 clips and cut to beat"

› step 1/8 open_media_pool........ verified

› step 2/8 import_clip × 3........ verified

› step 3/8 add_to_timeline × 3.... verified

› step 4/8 detect_beats............ verified

› …

› done · 8/8 · 1m 47s · 0 model retries

§04 — Compare

compare()/ measured vs stochastic

Same 50,000 calls.
Two different products.

without it

~70%

Stochastic. Silent. Slow.

Zero-shot CU figures it out every call. Failures are silent. Retries compound bugs. Every call hits the frontier model. No number you can put in a contract.

with case

99.6%

Measured. Versioned. Structured.

The same procedure runs across 51,204 customer calls on Resolve 20.2. Failures return {ok, repair_hint, retry_safe}. Idempotent procedures retry. Non-idempotent ones don't. You ship the number to your own customers.

§05 — Audience

built_for()/ scope

Built for the developers
shipping vertical AI agents.

01Your agent works in dev, fails at demo time.
02Customers ask why the export silently dropped audio.
03Pixel-click automation breaks every Resolve update.
04You ship Resolve now — Premiere is next quarter.
05You're hitting the desktop reliability ceiling.

§06 — Investment

your_status_page()/ populated by our numbers

Your customer-facing status,
powered by our measurements.

The numbers you ship to your customers come from us.

01 / operations, not marketing
99.6% across 51,204 calls on Resolve 20.2. 30-day uptime, p95 latency, per-version success — measured on real customer machines.
02 / your status page, our measurements
Drop per-procedure stats into your customer-facing reliability page. Numbers become yours to defend, backed by trajectory data and replay-deterministic audit trails.
03 / the compounding moat
Models will improve. "We know how to drive Resolve" erodes. Verification contracts, idempotency, audit trails, versioning — these compound.

your-product.com / reliability

Desktop procedure statuspowered by case · last 30 days

All systems operational

import_clip99.8%

add_to_timeline99.6%

apply_color_node97.1%

render_export99.9%

set_inspector_value99.6%

detect_beats99.5%

We don't sell reliability.
We sell the absence of dread.

Every developer shipping a desktop-touching agent runs the same monologue — "this will fail at demo time and I'll debug it under deadline." Structured failures, idempotency labels, versioned reliability, recovery docs — engineered for one thing: turning that dread into a typed function call with a published number.

§08 — Access

request_endpoint()/ one line

Reliable procedures for the
software your agent can't reach.

If your agent product lives or dies on desktop reliability, we should talk.

Read the spec →

Your agent crashed at 'element not found'.case returns the typed code → your planner replays.

Every procedure is one call.Open one to see why it ships.

Each procedure carries a typed schema, an idempotency label, a structured failure shape, per-version reliability, and a recovery doc. 66 procedures live · 12 shown.

From install to working agentin 10 minutes.

Same 50,000 calls.Two different products.