DevLoop MCP: all 28 tools

One prompt. The agent picks the tool, passes the build, and gets an app id back it can feed straight into the next call.

The agent launches the build, then confirms it actually rendered — because the tool tells it to. observe_screen hands back a real screenshot, which is what you see on the device.

One prompt. The agent reads back every build in the account — id, name, version and platform — and picks the one to test next.

One prompt. The agent resolves the build to a short-lived signed URL, then pulls the binary down — ready to install on a device.

One prompt. The agent retires the stale build and reports what it removed — a soft delete, so nothing is lost for good.

One prompt. The agent turns a described flow into a structured test case — steps, expected result and metadata — and gets an id back it can queue straight away.

One prompt. The agent patches just the step that changed and bumps the version — the rest of the case stays exactly as it was.

One prompt. The agent reads back every test case in the account — id, name, category and platform — and picks what to run next.

One prompt. The agent pulls the full case — every step and the expected result — so it knows exactly what to execute before queueing a run.

One prompt. The agent uploads a reference file and links it to the case, so the expected state rides along with every run.

One prompt. The agent queues the case on a real device and gets a run id back it can poll — the point where authoring turns into a live run.

One prompt. The agent reads the run board — what's running, what passed, what failed and on which device — and decides what to look at next.

One prompt. The agent reads the full result — status, duration, step tally and a link to the recorded Loop replay.

One prompt. The agent walks the per-step trace — each action, its status and how long it took — to see exactly what happened on device.

One prompt. The agent reads back the compiled plan — setup, act, verify — so the run is deterministic and reusable on the next build.

One prompt. The agent turns a failed run into a filed bug — summary, clean repro steps, the build link — with the failing run attached as origin evidence.

One prompt. The agent sends only the field that changed — here the sharpened repro steps — and everything else on the bug stays exactly as filed.

One prompt. The agent checks what's already open against the build — by human-readable app name, no id lookup — before it files a duplicate.

One prompt. The agent links a new failing run to the bug that already covers it, so the recurrence is on record without a second ticket.

One prompt. The agent uploads the crash recording and finalizes it on the bug — evidence that travels with the ticket, not a link that rots.

One prompt. The agent reads the whole case file — repro, linked runs and a signed download URL for the evidence — then pulls the recording down.

One prompt. The agent saves the flow it just tapped through as a named step sequence — credentials as runtime params, replayable in one call with no model round-trips.

One prompt. At session start the agent reads back every recorded routine — versions, trigger hints and params — and picks the shortcut instead of re-tapping the flow.

One prompt. The UI changed, so the agent appends a new version — old versions stay in history, and concurrent edits conflict cleanly instead of overwriting each other.

One prompt. The agent turns the raw device recording into a shareable Loop — converted, timelapsed past 60 seconds, uploaded, and linked for the PR.

One prompt. The agent reads the device pool — phones and tablets, platform, OS version and SMS capability — and picks where the suite should run.

One prompt. The agent reads the org's shared taxonomy and files new test cases into an existing bucket instead of inventing its own.

One prompt. The agent does the arithmetic — cases × devices — and validates the balance before queueing, so a suite never dies half-way through the night.

One prompt. The agent lists stored accounts by name and id only — the user picks one, the secret is inserted on-device and never enters the transcript.

One prompt. The skill hands the agent a routing table and the request lands with the subagent that owns it — the main agent orchestrates, the subagent drives the device.

One prompt. Before anyone reads app source, the skill forces a classification — and this failure turns out to be a held device session, not a bug.

One prompt. The skill walks the agent through closing a bug the right way — verify first, sparse status edit, human handle in the report.