A software factory
in your pocket.
rdn-swarm runs a full 9-role software-development pipeline on your phone — planning, coding, reviewing, testing, and scoring real apps against your own provider keys. No backend. No account. No telemetry.
Launching on Google Play & the App Store. Bring your own Anthropic, OpenAI, Gemini, or DeepSeek key.
Not a remote control for a server somewhere. The whole orchestrator lives here, running between your thumb and your battery.
Bring your own keys
Mix providers freely — a different model for every role.
Catalogs refresh live from each provider's own /models endpoint — pricing is bundled and fully editable.
Provider names and logos are trademarks of their respective owners. rdn-swarm is not affiliated with or endorsed by Anthropic, OpenAI, Google, or DeepSeek.
The five tabs
Home · Hub · History · Data · Settings
A bottom nav routes between five surfaces. The Hub becomes the live run dashboard mid-pipeline, and the New Run setup when idle.
Your fleet at a glance
Aggregate stats across every run on this device — total runs, success rate, average score, total spend, tokens in and out. A persistent Launch a new run button sits above the recent-run list; completed runs get a green rail, errors a red one.
- Fleet-wide metrics
- Recent runs with score chips
- One-tap new run
- Color-coded status rails
The live SDLC dashboard
A phase track across the top, a big score readout beside the run-status pill, then live cost and token cards with a per-provider breakdown and cache-hit rate. Watch agents work in real time as the pipeline streams.
When it finishes, the Summary tab renders the generated README natively — no webview bridge, just text.
A replayable archive
Every run is recorded — events, generated files, screenshots, costs. Filter chips split the list by lifecycle state. Tap any row to replay it through the same dashboard, rebuilt from the persisted event archive.
A head-to-head benchmark rig
Leaderboard ranks providers, models, roles, or projects by score and efficiency. Token Usage gives the fleet-wide cost and token breakdown with CSV export.
Same brief, same rubric, across providers — so the comparison is fair, not vibes.
Your keys. Your device.
Paste a provider key, validate it against the live API, and it is stored encrypted with an AES-256-GCM master key from the hardware Keystore. Keys never appear in logs.
48 models indexed, refreshed live from each provider's catalog.
The pipeline
Nine roles. One device.
A real software team, modeled as specialized agents. Each role can run on a different provider and model — and several run at once.
- 1AnalystOpenAI
Reads your brief and decomposes it into components and testable acceptance criteria.
- 2Project ManagerDeepSeek
Splits the work into non-overlapping assignments — one owner per file, no collisions.
- 3Developers ×NAnthropic
Write the code in parallel, each developer owning its own files.
- 4Integration ArchitectDeepSeek
Verifies cross-file consistency — selectors, element IDs, imports, references.
- 5QA Reviewers ×NAnthropic
Review the code in parallel and file issues by severity: critical, major, minor.
- 6Visual QAAnthropicMobile-first
Renders the running app and reads the screenshots and DOM as a vision model.
- 7Test AuthorDeepSeekMobile-first
Writes deterministic acceptance tests the orchestrator runs against the app.
- 8Feedback CoordinatorOpenAI
Consolidates every finding and scores the run across the five dimensions.
- 9Summary GeneratorGoogle
Writes the final README — what was built, how to run it, what is left.
Defaults shown. Every role is reassignable to any provider and model from the run's Configuration tab.
Developers and reviewers run in parallel — this is the swarm, mid-review, on a phone.
Quality scorecard
Five dimensions. One honest score.
Each dimension decays multiplicatively with open issues using per-severity half-lives. The overall score is a geometric mean — so a single dead dimension fails the whole run, it doesn't just nudge a number.
Correctness
Does the code actually do what the brief describes? Drops fast on critical bugs.
Completeness
How many acceptance criteria pass — produced by the Test Author's deterministic runner.
Integrity
Cross-file consistency: selectors, element IDs, imports and references all resolve.
Quality
Code style, structure, idiomatic patterns, and the absence of review findings.
Accessibility
Visual-QA findings against the rendered DOM — focus order, contrast, semantics.
A run is Successful once the score clears your threshold (default 0.80) with zero open critical issues. Otherwise it loops and fixes itself.
You stay in the loop
It iterates until it's good — or until you say stop.
When the score lands below threshold, Swarm enters bug-fix mode: the PM re-assigns only the files with open issues, developers patch in place, and QA re-reviews. You can let it run on autorun, or check in at every iteration and accept, reject, or inspect the output yourself.
Play the result.
These aren't mockups. Each one was built by the swarm from a one-page spec — planned, coded, reviewed, and scored — with zero human edits. The one on the right is running, right now, in your browser.
Connect Four
A real Connect Four — game logic, win detection, gravity, and the board itself, written end-to-end by the swarm.
↑ Playable · Sandboxed iframe · Tap a column. Red goes first.
Behind the bottom nav
The screens that do the work
The requirements library, configuration, the output browser, the in-app preview, the live log, the model catalog, and more.
Hub · New Run
A library of starting points
Thirty bundled briefs — games and utilities — plus your own markdown docs, persisted on-device. Edit raw or render the preview, then assign per-role models.
New Run · Configuration
Assign every role
Nine role cards. Pick a provider and model for each, set max output, or apply one choice to all. Orchestration, dev mode, QA mode, and the scorecard live here too.
Run · Output
File tree + zip export
Generated files in a tree with QA screenshots and tests alongside the code. Export the whole run as a zip through the Storage Access Framework.
Run · App Preview
Run it before you ship it
The generated app boots in a sandboxed WebView inside Swarm, with phone, tablet, and desktop viewport toggles, so you can sanity-check it before iterating.
Run · Live log
Every event, in order
Each phase, every usage tick, each consistency check and generation chunk is persisted. Replay drives the dashboard back through the whole timeline.
Settings · Model catalog
Per-row editable pricing
Providers don't return pricing from /models, so the catalog ships a bundled table refreshed from a live feed. Every row is yours to override.
Data · Leaderboard
Rank the field
Group by model, provider, role, or project and order by score or cost. The same brief and rubric across providers means the comparison is fair.
Settings · Display
Dual theme · 9 accents
Light, Dark, or follow System, with nine preset accents from Deep Blue to Lime. Theme and accent apply across the whole app instantly.
Settings · About
Build · pipeline · license
Version and git SHA, the engine and minimum OS version, the pipeline summary, and deep links to Swarm Online and the bundled license.
Security posture
The keys never leave this device.
Most "AI app" clients hand your keys to a server. Swarm doesn't have a server. Provider keys live in the hardware Keystore, and nothing about your runs is uploaded anywhere — unless you explicitly opt in to sharing anonymous benchmark stats.
Hardware-backed key vault
On Android, keys live in EncryptedSharedPreferences under an AES-256-GCM master key from the device Keystore. On iOS, each key is its own hardware-backed item in the system Keychain. They are redacted from every log.
Excluded from cloud backup
Auto Backup is disabled on Android; on iOS, keys are device-only Keychain items and the run database is excluded from iCloud backup. Nothing is ever copied off-device by the OS.
Direct-to-provider calls
Requests go straight from your phone to Anthropic, OpenAI, Google, or DeepSeek. Rdn Labs is never an intermediary and never sees them.
Sandboxed app preview
The WebView that runs generated apps and powers Visual QA boots with file and content access off, scoped to an isolated per-run origin.
Read the full privacy policy — the short version is that everything stays on your device and the app makes no network calls except the ones you direct.
Bundled benchmark suite
Same input. Same rubric. Real numbers.
Thirty starter briefs ship with the app and double as a benchmark suite — each with deterministic acceptance criteria, so head-to-head provider comparisons are fair.
Connect Four
Tetris
2048
Minesweeper
Wordle
Sudoku
Snake
Chess
Calculator
Pomodoro…and twenty more — Lights Out, Battleship, Reversi, Solitaire, Mastermind, Tower of Hanoi, Calendar, To-Do, and beyond.
Build your first app
on the train home.
Bring your own provider keys and let nine agents plan, build, test, and score a working app — entirely on your phone.
Questions? Read the support & FAQ or email support@reidell.net.