Skip to content

unfairly-ai/unfairly-intelligence

v0.8.4

See what works with AI on your team. Content-free session metadata, joined to shipped work, with shared learnings everyone can reuse.

GitHubOpen Website

Unfairly plugin for Unfairly Intelligence

The Unfairly plugin is the portable workflow component for Unfairly Intelligence, backed by the existing unfairly CLI. Its root Agent Plugins manifest carries the shared skills and MCP definition. Codex/ChatGPT Work use the OpenAI extension, Claude Code uses the Claude compatibility manifest, Cursor uses the native Cursor overlay, and Gemini CLI loads the nested gemini/ extension. All lifecycle hooks call the compatibility command unfairly pace capture and emit the same content-free event contract. Employee-facing commands use unfairly intelligence.

The plugin carries its own bundled runtime (runtime/unfairly-intelligence.mjs, generated from cli/src/pace-runtime.ts by npm run bundle:plugin-runtime in cli/), so installing the plugin is the whole install. Hooks call scripts/capture.cjs, which prefers the bundled runtime and falls back to a runtime placed by npx unfairly or a global unfairly. Identity, redaction, Git enrichment, durable local spooling, upload, and policy live in the shared CLI core.

Employees install it from their AI client's plugin screen, using the public marketplace Unfairly-AI/unfairly-intelligence (built from this directory by scripts/build-intelligence-marketplace.mjs). An administrator sends them https://intelligence.unfairly.ai/install?org=<org-id>, which walks through each client and shows the Claude managed-settings snippet for pushing it to everyone. On the first session start the plugin prints a link; the employee approves the machine in the browser (choosing the organization if they belong to several), and the machine receives its own revocable collector credential. A background poller picks up the approval within a few seconds, so the same session connects, and the first sync uploads the last 90 days of sessions after fetching the approved-repository list. After that, session summaries upload at every session start and end.

The terminal path still works: npx unfairly@latest setup --org <org-id> installs and enrolls, and unfairly intelligence backfill --days 90 uploads past sessions from the local transcripts.

Before activation, inspect the exact contract with unfairly intelligence policy. Use unfairly intelligence status to see enabled clients, queue depth, policy/runtime/plugin versions, attestation expiry, and last upload. Collection can be stopped without losing installation identity with unfairly intelligence pause and restarted with unfairly intelligence resume. unfairly intelligence uninstall removes only tracked Unfairly-managed files and local Intelligence state.

It also closes the learning loop. Employees can search the organization's approved marketplace with unfairly intelligence playbooks --query "<goal>", start one with unfairly intelligence playbook-use <playbook-id> --version <version>, and propose a content-safe learning after successful work. Playbook identity and version travel with workflow evidence so the dashboard can distinguish discovery, installation, reuse, and verified outcomes.

The same runtime now includes a local evidence adapter for ChatGPT Desktop, Claude Desktop, and other closed applications. A centrally managed allowlist defines which output or app-state directories may be observed. Output roots produce content-free created/modified artifact facts; app-state roots produce one aggregated activity fact per scan. The adapter never reads file contents, silently expands beyond approved roots, or emits filenames.