NotFair — 24/7 Goal-Driven Marketing Agents
State a business ambition in plain language; NotFair verifies the metric, sets a baseline, and runs a disciplined improvement loop 24/7 once you confirm.
Evidence shows: README describes user confirmation flow (plan confirmed in chat before loop starts), spend envelopes, observation windows, pause button, and code changes via PRs. However, no actual code or config is provided to verify these mechanisms. Deductions: lack of code-level evidence, least privilege, data flow transparency, and sensitive data handling (e.g., OAuth tokens unencrypted) only partially addressed.
Evidence shows: README and CI workflows (typecheck, test) suggest some self-consistency, but no test results or dependency lock files are provided. Deductions: dependency availability unverified, failure messages only partially mentioned (e.g., doctor's Fix line).
Evidence shows: README describes multiple use cases (SEO, paid ads, goal loops) and platform connections, but lacks detailed boundary descriptions or environment adaptation specifics. Deductions: trigger precision (e.g., goal confirmation) and specific environment (e.g., Apple Silicon) adaptation only partially addressed.
Evidence shows: README provides installation instructions, CLI reference, examples, and license, but lacks version history, changelog, and detailed known limitations. Deductions: naming stability (e.g., toprank rename) and versioning information incomplete.
Evidence shows: README displays output examples (e.g., audit reports, goal screen), but no actual run results or cost-benefit analysis. Deductions: marginal value and cost-benefit based on description only, lacking empirical evidence.
Evidence shows: Claims in README (e.g., server-verified metrics) lack traceable tests or data sources. Deductions: lack of cross-source corroboration and clear separation of facts and inferences.
- OAuth tokens are stored unencrypted, posing local exposure risk.
- Dependencies are not pinned, posing supply chain risk.
- Publisher identity is unverified; evaluate with caution.
What does this agent do, and when should you use it?
NotFair is an open-source, locally-run application that combines goal-driven agents with a marketing skills plugin. It turns vague ambitions like "grow organic clicks 30% in 90 days" into verifiable metrics, establishes a baseline via server-side verification, then runs a disciplined improvement loop—measure, evaluate past moves, make one new move, log the prediction, and repeat. The UI offers one screen per goal with chat on the left and a live progress chart, target line, and check diary on the right. NotFair also ships a Claude Code plugin with SEO, Google Ads, and Meta Ads skills for hands-on audits and campaign management. All data stays in a local SQLite file, and it runs on your machine using your existing Codex or Claude Code login—no new subscription or hosted runtime.
Runs npx notfair@latest to start a local server and opens the UI at http://127.0.0.1:3327. Users connect platforms (Google Ads, Meta Ads, Search Console, etc.) via OAuth, then state a goal in chat. The agent explores data sources, authors a metric query, verifies it by re-running server-side, backfills ~30 days of history, and proposes a target, cadence, and budget. Upon confirmation, the loop starts: each check, the platform mechanically measures the metric, wakes the agent to review past actions, and logs at most one new move with a falsifiable expected effect. Agents can make code changes via pull requests when a GitHub repo is attached. CLI commands include doctor, start, status, stop, logs, update, autostart.
- Marketing managers who want to set a quarterly business goal (e.g., “increase organic clicks by 30%”) and have an agent continuously monitor and improve SEO efforts.
- PPC specialists aiming to minimize wasted ad spend; the agent suggests reversible changes within a budget envelope and reviews them periodically.
- Growth leads connecting Stripe and Google Analytics to track revenue or CPA targets, receiving ongoing progress from the agent.
- SEO professionals using the plugin for one-off audits, such as finding traffic drops or quick wins using Search Console data.
- Developers wanting to extend the agent to custom metrics via the
localsource, which runs a shell command to obtain the number.
What are this agent's strengths and limitations?
- Behavioral discipline: agents are judged only on server-verified metrics, log falsifiable predictions, and observe observation windows, preventing fake activity.
- Fully local data: all state in ~/.notfair/db.sqlite, no new subscription, no hosted runtime.
- Wide integration: first-party MCP servers for Google Ads, Meta Ads, Search Console, plus PostHog, Mixpanel, Stripe, Supabase, and custom MCP endpoints.
- Comes with a comprehensive skill library for hands-on marketing tasks (SEO, ads audits), complementing the autonomous agent loops.
- Platform-specific: requires Apple Silicon Mac and Node 20+; not available on Windows/Linux.
- Runtime dependency: relies on your existing Codex CLI or Claude Code subscription; no LLM keys of its own.
- OAuth tokens stored unencrypted in SQLite, requiring local account security.
- Validation of long-running agent loops is limited beyond documentation; effectiveness may vary.
How do you install or deploy this agent?
Prerequisites: Apple Silicon Mac, Node 20+, and Codex CLI or Claude Code installed and authenticated. Run npx notfair@latest doctor for preflight, then npx notfair@latest to start the UI. Alternatively, install globally: npm install -g notfair. For the plugin, in Claude Code: /plugin marketplace add nowork-studio/notfair, then /plugin install notfair@nowork-studio.
How do you use this agent?
- Create a project in the browser UI and pick your harness (Codex or Claude Code). 2. Connect one or more data sources (e.g., Google Search Console) from the Connections page—one-click OAuth. 3. State a goal in the chat, e.g., “grow organic clicks 30% in 90 days.” The agent will verify the metric and propose a plan; confirm to start the loop. 4. Use CLI commands to manage the app:
notfair status,notfair stop,notfair logs. 5. For manual tasks, use plugin skills like/notfair:seo-analysisor/notfair:google-ads.
FAQ
What does it cost?
What permissions does it need?
What if the agent makes a mistake?
Can I add a custom data source?
local source that runs a shell command, or by providing a custom MCP URL—any MCP server speaking HTTP with OAuth works.