autonomous company engine

Your AI company
that runs itself

Send one directive. The AI CEO classifies it, routes it to agents, executes the work, and asks you only when it matters. You approve. It ships.

uvx kompany onboard --yes
Open source · Python 3.11+ · bring your own LLM key · live in 90 seconds
Read the source → Get updates
you › launch the product homepage, research design first

[ceo] directive classified: new project
[ceo] created KPNY-001 · phase 1: design research
[agent:research] 40 reference pages analyzed 
[agent:build] site built, newsletter wired 
[ceo] gate: deploy to production?

you › go
[ceo] live at runkompany.com  · cost: $0.03
How it works

One directive in. Shipped work out.

Kompany is not a chatbot. It is a company structure made of agents — with a CEO that routes, gates that protect you, and a ledger that remembers everything.

01 · directive

You speak

One sentence in natural language. No tickets, no specs, no meetings.

02 · route

CEO classifies

The AI CEO turns it into a project, picks the agents, sets the phases.

03 · execute

Agents work

Research, build, write, publish — autonomously, in parallel, with budgets.

04 · approve

You decide

Irreversible actions stop at a gate. One word from you releases them.

Features

Built like a company, not a chat window.

An AI CEO that actually manages

Every directive is classified, scoped, and delegated. Projects get IDs, phases, and gates. Unclear requests come back as one sharp question — not a wall of guesses.

[ceo] KPNY-001 · phase 1/3 active
├── research: done
├── design: gated → awaiting GO
└── build: locked

[ledger] ai cost today: $0.14

Approval gates, not blind autonomy

Agents run free on reversible work. Anything irreversible — deploys, payments, public posts — stops and waits for your one-word approval. Full audit trail in the ledger.

[gate] publish to production?
  diff: +412 −0 · cost est: $0.02
  risk: public, reversible via rollback

approve · revise · reject · snooze

Memory that compounds

Every project, decision, and lesson lands in a four-layer memory system. Your company gets smarter every week — raw episodes become distilled patterns, patterns become reflexes. See how it works →

[memory] 4 layers · 1 SQLite store
├── episodes (raw project records)
├── reflections (per-task lessons)
├── distilled patterns (cross-project)
└── recall (utility-ranked injection)

recall: before every task
Memory system

A company that remembers what it learned.

Kompany doesn't just log history — it distills it. Every finished project becomes a structured episode, every episode feeds a retrospective, every retrospective extracts patterns, and every pattern gets injected back into the agents that need it. The company gets sharper with each directive.

Project ships Episode Reflect Distill Recall Next task is smarter
Layer 1 · Raw

Episodes

When a project is delivered, six source tables — tasks, ledger, decisions, debates, audit log, agent memories — flatten into one structured EpisodePayloadV1 JSON document. One row, one project, the full story.

project_episodes table
retention: full for recent N, summary for older
rebuildable from source tables anytime
Layer 2 · Per-task

Reflections

Agents write reflections during execution — what worked, what broke, what they'd do differently. These land in agent_memories tagged category='reflection' and flow into the episode payload at delivery time.

agent_memories table
knowledge_type: experiential
permanent, merged during distillation
Layer 3 · Cross-project

Distilled patterns

The Chief of Staff agent runs a retrospective across recent episodes, extracts durable cross-project patterns, and UPSERTs each one into memory keyed by (agent_role, pattern_key). Inferred-only claims with no evidence trail are rejected before they pollute memory.

runs as CoS directive
evidence-trace guard: no episode IDs = rejected
confidence score stored in metadata
Layer 4 · Injection

Recall

Before every task, relevant memories are pulled and injected into the agent's prompt. Retrieval is utility-ranked: keyword hits + recency bonus + access frequency + pattern confidence. Memories that keep getting used rank higher over time.

score = keyword + recency + log(access) + 2×confidence
access tracking: only on injection
stale facts auto-expired via valid_until
# project delivered → episode materialized [engine] run_retrospective(KPNY-001) [episodes] materialize → EpisodePayloadV1 [episodes] record_or_update → project_episodes [audit] learning.episode_recorded # CoS runs distillation across recent episodes [cos] distill(since=30d) [cos] 12 episodes in → 5 patterns extracted [guard] 2 inferred-only claims rejected (no evidence_episode_ids) [memory] upsert 3 patterns → agent_memories [audit] learning.distillation_run · cost: $0.04 # next directive — recall before task [agent:build] recall(query="deploy static site", limit=5) [memory] 3 patterns matched · score-ranked · access_count++ [agent:build] prior learnings injected into prompt
6 → 1
Source tables flattened into one episode row
utility-ranked
Recall scoring: keyword + recency + access + confidence
evidence-gated
Distilled patterns without episode evidence are rejected
Showcase

Running in production: Nova.

Nova is an AI voice run entirely by Kompany — operating a real business with a public ledger. Every decision, expense, and dollar earned is published daily. Current mission: turn $50 into $1,080.

case 01 · live

The Swedexpress Ledger →

The public ledger Nova writes every day: decisions, costs, revenue, mistakes included. Radical transparency as proof of work.

voice · @prompt_nova

Nova on X →

The same agent narrating its own build in public — posts, launches, and daily updates, all generated and gated through Kompany.

FAQ

Questions, answered directly.

What is Kompany?

Kompany is an open-source autonomous AI company engine. You send a one-line directive in natural language; an AI CEO classifies it, creates a project, and routes tasks to specialized agents that execute real work — research, building, publishing — under per-task budget caps. Irreversible actions wait at approval gates for the human founder.

How is Kompany different from a chatbot or a single AI agent?

A chatbot answers; Kompany operates. It has a company structure: a CEO that routes work, multiple executing agents, an approval inbox, a persistent ledger of every decision and cost, and a 24/7 daemon. It runs multi-turn execution sessions via Claude Code CLI, Codex CLI, or opencode — on your own subscription or API key.

How do I install Kompany?

One line: uvx kompany onboard --yes. It checks your environment, stores your LLM API key in an encrypted vault, applies a starter company template, and prints next steps — in under 90 seconds. Requires Python 3.11+ and an API key for one supported LLM provider.

Is Kompany safe to run autonomously?

Autonomy is bounded by design: every task has a budget cap, and side effects — deploys, payments, public posts — stop at an approval gate until the founder approves. A negative balance never cancels goals silently; everything is recorded in the ledger.

Is Kompany actually used in production?

Yes. Nova (@prompt_nova) is an AI voice run entirely by Kompany, operating a real business with a public daily ledger — decisions, costs, and revenue published at the Swedexpress Ledger.

Updates

Watch a one-person AI company get built.

Build logs, real costs, real revenue numbers. No hype — just what the agents actually shipped this week.

One email a week. Unsubscribe anytime.