Skip to content
famaoai-creatorPublic

About

Open-source organization work loop engine — governed multi-agent orchestration with goal-driven workers, evidence-gated missions, and an organizational memory that improves every run.

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

1 star

Watchers

0 watching

Forks

Latest commit

 

History

6,282 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Kyberion

Kyberion

License: MIT Node.js >=24 CI Status

Category Category Category Category Category

An organization work loop engine.
You phrase outcomes. Kyberion plans, runs, and remembers with evidence.

Intent → Plan → Result

Kyberion turns a request into a visible plan and a verified result. You say 今週の進捗レポートを作って or この PDF をパワポにして; it picks the tools, asks only when something is genuinely ambiguous, and hands back the result, the artifact, the evidence that future work builds on, and the next action.

It is OSS and self-hosted: your data stays on your machine, every side effect is governed, and every run leaves an audit trail.

Start here

I want to… Go to
Try it (5 minutes) Quick Start → docs/QUICKSTART.md
Understand what it is What is Kyberion? → docs/WHY.md (日本語)
See what it can do What it can do → docs/SCENARIO_CATALOG.md · CAPABILITIES_GUIDE.md
Use it day to day How you work with it → docs/SURFACES.md · docs/user/
Deploy / operate it docs/operator/DEPLOYMENT.md · docs/operator/
Extend it or contribute docs/developer/EXTENSION_POINTS.md · CONTRIBUTING.md
Look up a term docs/GLOSSARY.md — three tiers: first-win, contributor, and FDE
Check what is actually built Project Status → status index

What is Kyberion?

Knowledge work is moving from "I do this manually with LLM help" to "I delegate and verify". The winning system is not the most chat-fluent model but the engine that captures intent reliably, keeps evidence, and accumulates organizational memory. Full thesis: docs/WHY.md.

The Kyberion work loop: intent → plan → execute → evidence → learn, with learning feeding the next mission's team

Most agent frameworks stop at "execute". Kyberion closes the loop:

  • No evidence, no "done". Finishing a work cycle checks every success criterion against actual artifacts and verifications. Unsatisfied gaps automatically dispatch gap-closing work — the wording of a request never substitutes for its purpose.
  • The work loop improves itself. Every finished work cycle runs a retrospective: deterministic execution stats ground improvement proposals (human-ratified, never auto-applied), and measured outcomes improve future staffing.
  • Frontier-model discipline on any model. The working philosophy — read before write, one change one verification, no retry without a new hypothesis, evidence-based completion — is injected into every worker prompt, so small models inherit the habits that make frontier models reliable.
  • Governance by architecture, not by prompt. Three-tier knowledge isolation is enforced at the file-IO boundary. Customer conversations are physically separated from mission state. Outbound sends always pass an approval gate. An append-only audit chain records everything.
  • Workers get briefed, not dumped. Each worker receives a role-scoped context pack — mission goal, acceptance criteria, and the top hints distilled from previous runs — under an explicit size budget.

Core concepts in one minute

Concept What it is
Mission One unit of work with its own git repo, state and evidence; survives 24h+ runs.
ADF / pipeline Declarative, schema-validated plan format (sub-pipelines, on_error recovery). Ready-made ones live in pipelines/.
Actuator A governed capability module (browser, file, voice, code, …). 33 today — catalog.
Surface A human-facing entrance: Chronos, Concierge, Presence Studio, terminal HUD, chat bridges, capture pads.
Tier & tenant personal/ → confidential/ → public/ knowledge, scoped per tenant; nothing leaks downward.
Stance customer/{slug}/ overlay that swaps identity, connections and policy for the entity you act as — without forks. Not a tenant (how they differ).

New here? Read docs/CORE_CONCEPTS.md — the 5 concepts you need first. Concept map: kyberion-concept-map · Parent architecture: organization-work-loop.


Quick Start

Start here — canonical cold-start source: docs/QUICKSTART.md (it also explains which onboarding command to use when). Map of all docs: docs/README.md. This page is the short version. Day-2 tenant / organization / activation work: docs/INITIALIZATION.md. Documentation authority map: docs/documentation-source-map.json.

Kyberion's first visible result comes in three short steps:

  • 30 seconds: run pnpm kyberion doctor and see Kyberion's readiness/value boundary
  • 5 minutes: run the clean browser smoke and get active/shared/tmp/first-win-session.png
  • 15 minutes: read the Quickstart structure map, then inspect the pipeline and actuator entrypoints

Requires Node.js 24+ (.nvmrc / package.json engines) and pnpm.

git clone https://github.com/famaoai-creator/kyberion.git
cd kyberion
pnpm install
pnpm build
pnpm env:bootstrap --manifest kyberion-toolchain
pnpm kyberion doctor
pnpm pipeline --input pipelines/verify-session.json

env:bootstrap verifies the Node 24+ floor and warns if Playwright browsers are missing. The last command opens a local first-win page and writes active/shared/tmp/first-win-session.png. pnpm exec playwright install chromium is optional: without Chromium the pipeline writes its governed text fallback instead of hiding the readiness result.

Path Prerequisites Time Command Notes
First-win Node 24+, pnpm ~5min the five commands above Writes the screenshot (or the governed fallback)
Voice first-win macOS only (native TTS; not available in Docker) ~5min pnpm pipeline --input pipelines/voice-hello.json Run after the browser smoke
Docker Docker Desktop ~10min build docker compose --profile deploy up Headless services only — voice/GUI actuators need the native macOS path

Not sure where to go next? pnpm kyberion setup report --persona first-time-user is the entry guide: it tells you whether to start with Chronos, the concierge, the voice path, or a messaging surface, and whether auth/setup is still blocking that route. If a browser, voice, or media actuator is missing a local dependency, check it with pnpm deps:check --actuator browser (or voice, media-generation).

Already have onboarding JSON? Skip the wizard: pnpm onboarding apply --identity knowledge/public/templates/onboarding/identity.example.json --dry-run (copy and edit the template, then rerun without --dry-run).

To understand the structure in 15 minutes, read docs/QUICKSTART.md sections 4-10, then inspect pipelines/verify-session.json, CAPABILITIES_GUIDE.md, and docs/developer/EXTENSION_POINTS.md. For a server / customer deployment: docs/operator/DEPLOYMENT.md.


What it can do

Every capability is a governed actuator or a ready-made pipeline — nothing here is an unbounded shell. Op-level catalog: CAPABILITIES_GUIDE.md · pipelines: pipelines/README.md.

Area What you get
Browser & desktop Record a web flow once and replay it reliably; drive the desktop with screenshot grounding (Set-of-Marks detectors, OS accessibility on macOS / Windows); terminal (PTY) control.
Documents & media Read PDF / PPTX / DOCX / XLSX / HTML; generate documents and slide decks from semantic briefs; image, audio and video perception and generation; narrated video.
Voice Browser speech in, OS-native or self-hosted speech out; talking avatar; meeting join, minutes and follow-up.
Code Refactor, scaffold, analyze, review; SDLC-cycle pipelines; delegated subagent work.
Services & network Governed fetch; Slack / Google / Notion / Microsoft 365 / email / calendar integration; deployment and cloud-operation actuators.
Knowledge & memory Search, distill and reuse organizational hints, including zero-LLM history search (SQLite FTS5 + CJK trigram, tier-isolated); working memory; volatile-knowledge GC.
Organization operations Operating-model control plane (purpose, services, routine operations, incidents, cadences, decisions — six work_shape kinds beyond solution projects), governed project management, and the canonical context chain tenant_slug → organization_id → project_id → mission_id → task_id.
Multi-tenant & multi-agent Tenant registry with isolated knowledge roots, deny-unless-brokered cross-tenant access, an HMAC-signed peer mesh (pnpm peer:register), and Co-Session coordination for several provider CLIs in one checkout (model).
Governance & trust Approval gate on every outbound effect, append-only audit chain, OTel-style traces, provenance-gated plugins (pnpm plugin:install; third-party code needs human approval), goal-driven workers with budgets and restart recovery.

One verb per sense

Day to day you rarely write a pipeline — you use one command per sense (Markdown on stdout, --json for structure):

Direction Command Direction Command
Document → text pnpm kyberion read Brief → document pnpm kyberion write
Image → text pnpm kyberion see Prompt → image pnpm kyberion draw
Audio → text pnpm kyberion listen Text → audio pnpm kyberion speak
Video → timeline pnpm kyberion watch Document ↔ document pnpm kyberion diff
Ask in plain words pnpm kyberion ask "…" Approve / reject pnpm kyberion approvals

pnpm kyberion with no arguments is the terminal home: a status digest plus your next move. Every command is listed in the generated CLI Reference; task-oriented notes are in the Commands Guide. Verb inventory: capability-verb-inventory.

Design-system-governed output

PPTX and video are authored as semantic briefs; a single style cascade and text-measured layout fitting keep output on-brand without per-slide hand-tuning.


How you work with it

Pick the entrance that fits the moment. All of them share the same missions, approvals and audit chain.

Entrance Use it when Start
Terminal home / HUD You live in a shell and want the next action. pnpm kyberion · pnpm tui
Concierge / Presence Studio You want a front desk: ask, decide, follow progress. pnpm surfaces reconcile
Chronos You supervise: intervene, review artifacts, audit. pnpm chronos:dev
Chat bridges & voice You are in Slack / Telegram / Discord / iMessage, or hands-free. docs/SURFACES.md
Capture pads You have something on your desk: a sketch, notes, a file. pnpm pads

Surfaces — one role per screen

Each surface answers one question and shows its role in the header. The two human-facing surfaces (Concierge and Presence Studio) share one five-item rail — ホーム / 頼む / 決める / 進み具合 / 設定 — so they read as a single front desk. Full role map, ports and access rules: docs/SURFACES.md.

Chronos Mirror — control tower home with tenant scope, view switcher, next action and per-mission agent status
Chronos Mirror · :3000
Control tower: what is the system doing, where should I intervene?
Concierge — the ホーム (Home) page on the shared kyberion-base design system: next action, response status and the five-item rail
Concierge · :3050
CEO secretary: what do I need to decide right now? (決める · 設定)
Presence Studio — the ホーム (Home) page: today's briefing, the ask box with request chips, what needs your decision and progress, behind the shared five-item rail
Presence Studio · :3031
Companion: what are we working on together, by voice or text? (ホーム · 頼む · 進み具合)
Operator Surface — read-only audit monitor listing missions and capability bundles, on the shared kyberion-base design system
Operator Surface · :3331
Audit monitor, read-only: what happened, with evidence?
Computer Surface — live mirror of the browser or terminal Kyberion is operating, with session, executor and status tiles
Computer Surface · :3040
Mirror: what is Kyberion doing in the browser or terminal right now?
Terminal HUD — Ink TUI with operator cockpit, intent preview and the mission panel
Terminal HUD · pnpm tui
Terminal cockpit: missions, work items, runtimes and intent preview without leaving the shell

Beyond the screens, Slack / Telegram / Discord / iMessage bridges share one approval contract and a durable outbox (mechanisms hermetically tested; external-service E2E is still being proven), and voice runs through voice-hub and Presence Studio.

Shared UI (A2UI kyberion-base)

All 5 UI surfaces above render from one design system: the kyberion-base A2UI catalog (ui:* component types with JSON Schema props), one token-driven stylesheet (kyberion-ui.css), and two renderers — React (@agent/shared-ui) for the three Next.js surfaces and a dependency-free vanilla DOM renderer for the two static-HTML surfaces. Every component ships in both light/dark themes and en/ja locales. See every component at once in Presence Studio's /ui-gallery:

UI gallery — every kyberion-base component (page header, nav rail, tabs, next action, table, charts, forms, status pills) in light and dark, English and Japanese

Details, schema and CSS sources: docs/developer/design/DESIGN_SYSTEM.md.

pnpm surfaces reconcile              # start the surfaces declared in active-surfaces.json
pnpm chronos:dev                     # or run the control tower alone
pnpm tui                             # terminal HUD (pnpm tui --once for a non-interactive snapshot)

Access control

Every HTTP surface resolves a viewer principal and tenant scope server-side (ViewerContext); a client-supplied tenant only narrows what the viewer may already see, never widens it. Enforcement is staged via KYBERION_VIEWER_SCOPE=off|warn|enforce (default warn). KYBERION_API_TOKEN / KYBERION_LOCALADMIN_TOKEN remain compatible all-tenant tokens for the single-operator local workflow; scoped token registrations can restrict a viewer to selected tenants. On Next.js 15+ a same-machine browser is recognised as loopback only through a surface token or KYBERION_TRUST_PROXY=1 behind a proxy that sets x-real-ip. Remote browsers sign in through a shared OIDC login (/login; Google and Microsoft Entra supported, a signed HttpOnly session cookie, only identities bound to an active member) — see docs/developer/SURFACE_OIDC_LOGIN_OPERATIONS.ja.md. Hosted user management is not implied by this boundary. Operations: docs/developer/CHRONOS_VIEWER_SCOPE_OPERATIONS.ja.md.


Local Pads — capture at your desk, hand off to Kyberion

The Capture desk is one 127.0.0.1-only server for the eight capture pads below. Start it once with pnpm pads, choose a pad from the menu, and inspect its authenticated history. It captures something you already have on your desk (a sketch, meeting notes, a screenshot, a file, a clipboard, today's TODO), stores records in the server-derived tenant/tier partition, and never starts a mission or sends anything on its own. The legacy per-pad pages remain available during migration. Every pad renders with the shared UI kit (toolbar, dialog, sketch board with a drawing palette, voice input with a live level meter) in light/dark and English/Japanese.

Sketch Input — draw a diagram, dictate an instruction, hand off PNG + handoff.json
Sketch Input · :8147
Draw a diagram, dictate the instruction, hand off PNG + JSON
Meeting Notepad — notes, dictation, recording and attachments turned into minutes
Meeting Notepad · :8148
Notes + recording + attachments → minutes → handoff
Report Review — an edit/comment/voice layer overlaid on any self-contained HTML report
Report Review · :8137
Edit / comment / dictate on any HTML report, save back in place
Screenshot Annotate — paste or drop an image, draw annotations, hand off to vision
Screenshot Annotate · :8150
Paste / drop an image, mark it up, hand off to vision
Memory Capture — brain-dump notes and tags into a working-memory handoff
Memory Capture · :8149
Brain-dump notes + tags → working-memory handoff
Clipboard Inbox — collect pasted snippets and URLs into an inbox handoff
Clipboard Inbox · :8151
Collect pasted snippets and URLs → inbox handoff
Daily Desk — Journal / TODO / NOW faces edited side by side
Daily Desk · :8152
Journal / TODO / NOW faces, seeded from working memory
Doc Drop — drop PDF, images, text or docx files for ingestion
Doc Drop · :8153
Drop PDF / images / text / docx → ingest handoff
Personal Workbench — link, task, follow-up, decision, expense and daily-review inbox with governed actions
Personal Workbench · :8154
Link / task / follow-up / decision / expense inbox; proposal-only, with governed OCR / email-draft actions

The :81xx ports above are the legacy standalone ports (used only when you start a single pad directly, see below). pnpm pads serves every pad from one server on http://127.0.0.1:8160/.

Start the Capture desk and open the printed URL:

KYBERION_PERSONA=sovereign KYBERION_TENANT=<tenant-slug> pnpm pads
# open http://127.0.0.1:8160/

# Legacy individual entry point (still supported during migration)
KYBERION_PERSONA=sovereign KYBERION_TENANT=<tenant-slug> \
  node_modules/.bin/tsx scripts/meeting-notepad/server.ts        # or sketch-input, daily-desk, …

# report-review wraps an existing report instead of a blank page
KYBERION_PERSONA=sovereign KYBERION_TENANT=<tenant-slug> \
  node_modules/.bin/tsx scripts/report-review/server.ts active/shared/tmp/report.html

What they share:

  • Loopback only. Bind to 127.0.0.1, reject other origins, and require the per-run token printed at startup for every write. The token is not a substitute for human approval.
  • Tier and tenant on the command line. --tier public|confidential|personal --tenant <slug> decide where the session lands; confidential / personal need a server-side KYBERION_TENANT, and Personal Workbench (default tier personal) always does.
  • Proposal, not execution. The unified desk writes durable, scope-partitioned pad records and an authenticated history index. Legacy entry points continue to write their session folder under active/shared/tmp/<pad>/ plus handoff.json. Missions, sends, calendar changes and knowledge promotion stay behind the normal approval gates. Personal Workbench's /action exposes only governed OCR, a knowledge-promotion candidate, and email drafts.
  • One helper, many pads. They are thin twins built on scripts/lib/local-artifact-pad.ts and registered in the protocol-service registry, so adding a pad is a small, reviewable change. Index: scripts/personal-pads/README.md.

Project Status

OSS, in active development. Pre-1.0. The roadmap is in docs/PRODUCTIZATION_ROADMAP.md:

  • Phase A — Make first-win 5 minutes. (in progress)
  • Phase B — Make it survive 30 days of continuous use. (foundations landed)
  • Phase C' — Make it contributable in under a week.
  • Phase D' — Make FDE / implementation-support engagements possible without forks.

The strategic positioning is OSS-first, with paid implementation support / FDE as the eventual revenue model. SaaS only after a clear user base exists (see docs/PRODUCTIZATION_ROADMAP.md §0 for the explicit "yes / no" list).

Multi-tenant isolation, the organization operating model, and viewer-scoped surface authorization have landed as engineering foundations; productized SaaS — billing, IdP/SSO, hosted user management — remains explicitly out of scope. The README describes the product; the source of truth for what is actually implemented is the status index: docs/developer/improvement-plans-2026-08/README.ja.md (release history: CHANGELOG.md).


Documentation Map

If you want to Read
Understand why this exists docs/WHY.md / .ja.md
Try it in 5 minutes docs/QUICKSTART.md
Fix a first-run problem docs/user/TROUBLESHOOTING.md
Browse what it can automate docs/SCENARIO_CATALOG.md · docs/user/USE_CASE_QUICKSTARTS.md
Pick a surface / entry point docs/SURFACES.md
Run day-to-day operations docs/OPERATOR_UX_GUIDE.md
Deploy it for a customer docs/operator/DEPLOYMENT.md
Understand the architecture knowledge/product/architecture/organization-work-loop.md
Author a new actuator / pipeline docs/developer/EXTENSION_POINTS.md
Customize for a customer docs/developer/CUSTOMER_AGGREGATION.md / .ja.md
Understand the data flow / privacy docs/PRIVACY.md / .ja.md
Run multi-tenant isolation knowledge/product/architecture/multi-tenant-operations.md
Operate viewer-scoped API access docs/developer/CHRONOS_VIEWER_SCOPE_OPERATIONS.ja.md
Check what is actually implemented docs/developer/improvement-plans-2026-08/README.ja.md
Contribute CONTRIBUTING.md
Report a security issue SECURITY.md

Three audiences, three folders: docs/user/ (using Kyberion) · docs/operator/ (running it as a service) · docs/developer/ (extending it). Questions and showcases: docs/COMMUNITY.md — GitHub Discussions for how-to, Issues for reproducible bugs, SECURITY.md for vulnerabilities.


How It Compares

You've used What Kyberion adds
ChatGPT / Claude.ai Stateful missions, governed execution, a catalog of actuators (browser, file, voice, …), audit chain, reusable memory across runs.
Cursor Code is one actuator among many. The unit of work is a long-running mission with persistent state, not a single chat.
Computer Use / browser agents Mission-scoped state, tier-isolated knowledge, customer aggregation. The browser is one tool, not the substrate.
Zapier / n8n / RPA Replaces brittle rule chains with intent-driven plans. Plans survive site changes via Trace-fed reusable hints.
AI Ops / agent SaaS OSS, self-hostable, customer-data-stays-local. No central server. FDE-ready for implementation engagements.

Project

MIT licensed — LICENSE; third-party licenses are inventoried by pnpm license:audit (generated, not committed). We follow the Contributor Covenant (CODE_OF_CONDUCT.md). Governance: GOVERNANCE.md · MAINTAINERS.md · CODEOWNERS. PRs welcome — see CONTRIBUTING.md.

Kyberion is operator-facing in English, conceptually-authored in Japanese. Both languages are first-class. See docs/DOCUMENTATION_LOCALIZATION_POLICY.md.

About

Open-source organization work loop engine — governed multi-agent orchestration with goal-driven workers, evidence-gated missions, and an organizational memory that improves every run.

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Used by

Contributors

Languages