# Nimrobo

> Nimrobo is a local-first Mac app, an outcome harness that runs an AI agent on a loop against one real number: try an action, measure it, keep what worked and drop what didn't.

Point Nimrobo at one number. Every run, it reads back what its earlier runs proved and disproved about that number, decides what's worth trying next, and states what it expects before it acts. The real result settles that prediction and joins the record the next run reads from. It runs on your Mac, on your existing ChatGPT Plus/Pro sign-in or your own API key. Download: https://www.nimroboai.com/download.

## What does Nimrobo do?

Nimrobo turns "an agent did some work" into "we know whether it helped." Every run is one bounded action tied to a hypothesis, and the reward that comes back decides whether that hypothesis was supported, refuted, or inconclusive. Verdicts stay on the record, so later runs start from what already worked instead of starting cold.

The app is organised around five surfaces:

- **Outcome dashboard**: the north star, the guardrails that can't regress, open experiments, and recent runs on one screen.
- **Levers**: durable classes of action believed to move the number, each moving from candidate to proven or retired.
- **Hypotheses**: the prediction, written before the run and frozen the moment the first result lands.
- **Experiments**: one hypothesis per experiment; the reward snapshot settles the verdict.
- **Runs and control**: one bounded action per run, explore or exploit, in Default, Manual, or Plan mode.

## What is the outcome loop?

An outcome project is one outcome loop: a git-tracked folder holding your north star and guardrails, a lever map, gate criteria, and every run on the record. The cycle is:

- **Set up**: write `goal.md` (north star + guardrails), `run.md` (the lever portfolio), and `gate.md` (the completion contract).
- **Run**: pick one lever, ship one action, record a falsifiable hypothesis and the reward it earns.
- **Update**: promote levers that worked, retire levers that failed.

Each run is either `explore` (test an uncertain hypothesis) or `exploit` (rep a proven one).

## Is Nimrobo just another coding agent?

No. Nimrobo wraps a coding agent rather than competing with it. The coding agent writes and runs the work; Nimrobo is the harness around it: the north star you're chasing, the levers you're betting on, and a record of which hypotheses held.

## How is this different from evals?

Evals score a model on a fixed benchmark before anything ships. The outcome loop optimizes an agent against a moving, real-world outcome (signups, reach, recovered revenue, organic traffic) after the work ships, and compounds across runs by remembering what moved the metric and what did not.

## Platform, control, and privacy

Nimrobo is macOS-only: the shell sandbox that confines a run is built on macOS Seatbelt. Every run has a mode (Default, Manual, or Plan), and any action reaching past the sandbox stops at an approval. Your outcome folders, runs, sessions, artifacts, reward snapshots, and provider credentials stay on your Mac by default.

## Pricing

Nothing in the harness is gated; only capacity scales.

- **Free**: $0/mo, one active outcome project.
- **Starter**: $5/mo, unlimited outcome projects and 250 hosted web searches a month.
- **Plus**: $20/mo, unlimited outcome projects, 500 searches, and $15 of Gemini credits.

## Core concepts

- [North star](https://www.nimroboai.com/docs/outcome): The one real-world outcome that counts as progress: stable, measurable, and owned by goal.md.
- [Guardrail](https://www.nimroboai.com/docs/outcome): A constraint that stops a cheap win from quietly degrading quality (e.g. "no engagement bait", "bounce rate must not rise").
- [Lever](https://www.nimroboai.com/docs/outcome): A mechanism you believe moves the north star: something you can pull, like hook quality or signup friction.
- [Lever portfolio](https://www.nimroboai.com/docs/outcome): The archive of all levers in run.md, each tagged candidate, promising, proven, exhausted, or retired.
- [Action](https://www.nimroboai.com/docs/outcome): One concrete step on a lever this run: a rep of a proven recipe, or a fix to something in the path.
- [Diagnostic metric](https://www.nimroboai.com/docs/outcome): A measurement that explains why a lever did or did not move the north star.
- [Hypothesis](https://www.nimroboai.com/docs/outcome): A falsifiable prediction attached to one lever, recorded in Superdense before the outcome is known.
- [Experiment](https://www.nimroboai.com/docs/outcome): The Superdense record binding a hypothesis to one or more runs/artifacts and a reward window.
- [Mode](https://www.nimroboai.com/docs/outcome): Each run is either explore (test an uncertain hypothesis) or exploit (rep a supported one on a proven lever).
- [Gate](https://www.nimroboai.com/docs/outcome): The reusable completion contract that says what must be true before a run is called complete.

## Outcome packs

Outcome packs are ready-made loops: a north star, guardrails, and a lever map your agent can start running today.

- [X Reach](https://www.nimroboai.com/docs/outcome/packs/x-reach): North star: impressions per post, measured via X API public_metrics (impression_count).
- [Landing Page Conversion](https://www.nimroboai.com/docs/outcome/packs/landing-page-conversion): North star: % of unique landing-page visitors who complete the signup action, measured in your analytics tool over a consistent window (7 days is a common starting point, adjust to match your traffic volume).
- [Turn Signups into Active Users](https://www.nimroboai.com/docs/outcome/packs/turn-signups-into-active-users): North star: % of signups who reach the aha moment within a defined window (7 days is a common starting point, adjust based on your product's typical time-to-value).
- [Recover Failed Subscription Payments](https://www.nimroboai.com/docs/outcome/packs/recover-failed-subscription-payments): North star: % of failed subscription payments that recover within a fixed window, measured from your billing provider's failed-payment event to a paid/captured payment event.

## Reward connectors

Connectors let the loop read an artifact's real-world reward through an official CLI or API.

- [GitHub connector](https://www.nimroboai.com/docs/reward/connectors/github): reads reward via `gh`, comments, reactions, merge state, review state, stars, forks, release asset downloads.
- [npm connector](https://www.nimroboai.com/docs/reward/connectors/npm): reads reward via `npm`, downloads, dependents, version adoption.
- [X (Twitter) connector](https://www.nimroboai.com/docs/reward/connectors/x): reads reward via `xurl`, views (impressions), likes, reposts, replies, bookmarks.
- [Stripe connector](https://www.nimroboai.com/docs/reward/connectors/stripe): reads reward via `stripe`, signups, conversions, conversion rate, revenue, MRR delta.
- [Razorpay connector](https://www.nimroboai.com/docs/reward/connectors/razorpay): reads reward via `razorpay`, payments, captured amount, failed amount, recovered amount, invoice status, subscription status.
- [Vercel connector](https://www.nimroboai.com/docs/reward/connectors/vercel): reads reward via `vercel`, visitors, pageviews, top paths, and (with Web Analytics) conversions.
- [Netlify connector](https://www.nimroboai.com/docs/reward/connectors/netlify): reads reward via `netlify`, visitors, pageviews, top sources, bandwidth.
- [Resend connector](https://www.nimroboai.com/docs/reward/connectors/resend): reads reward via `resend-cli`, delivered, opens, clicks, open rate, click rate, bounces.
- [Google Analytics (GA4) connector](https://www.nimroboai.com/docs/reward/connectors/google-analytics): reads reward via `GA4 Data API`, users, sessions, pageviews, events, conversions, retention.
- [YouTube connector](https://www.nimroboai.com/docs/reward/connectors/youtube): reads reward via `YouTube Data API v3`, views, likes, comments, and (via Analytics API) watch time.
- [Instagram connector](https://www.nimroboai.com/docs/reward/connectors/instagram): reads reward via `Meta Graph API`, reach, impressions, likes, comments, saves, video views.
- [Substack connector](https://www.nimroboai.com/docs/reward/connectors/substack): reads reward via `community CLI (sbstck-dl)`, views, opens, likes, comments, subscriber delta (dashboard-only for most).

## Optional

- [Download Nimrobo](https://www.nimroboai.com/download): the macOS build.
- [Outcome loops](https://www.nimroboai.com/outcome-loop): the Mac app itself — one run end to end, the levers, hypotheses, experiments and verdicts it keeps, and the sandbox and approval settings around it.
- [Pricing](https://www.nimroboai.com/pricing): both ways to pay — Outcome Deals (pay up front against a target for 28 days of Google organic clicks, refunded in proportion to the promised lift we miss, with the settlement mechanics and arithmetic in full) and the Free, Starter, and Plus harness plans.
- [Outcome packs index](https://www.nimroboai.com/docs/outcome): all packs and the outcome-loop reference.
- [Reward connectors index](https://www.nimroboai.com/docs/reward): the full connector taxonomy.
- [Superdense on GitHub](https://github.com/Nimrobo/superdense): Nimrobo AI's open-source CLI, which predates the Mac app and runs the same outcome loop from the terminal.