comis

<p align="center">
  <img src="assets/comis-readme-banner.png?v=1.0.42" alt="Comis — self-hosted AI agents you can audit" width="100%" />
</p>

<p align="center">
  <strong>A self-hosted AI agent platform that deploys into your existing messaging channels.</strong>
  <br />
  <sub>Kernel-enforced sandboxes · sigstore-signed releases · multi-tenant primitives · 81% prompt-cache savings · across 9 chat channels.</sub>
</p>

<p align="center">
  <a href="https://github.com/comisai/comis/actions/workflows/ci.yml"><img src="https://github.com/comisai/comis/actions/workflows/ci.yml/badge.svg?branch=main" alt="CI" /></a>
  <a href="https://www.npmjs.com/package/comisai"><img src="https://img.shields.io/npm/v/comisai?color=06B6D4&style=flat" alt="npm" /></a>
  <a href="https://github.com/comisai/comis/blob/main/LICENSE"><img src="https://img.shields.io/badge/license-Apache--2.0-06B6D4?style=flat" alt="License" /></a>
  <a href="https://github.com/comisai/comis/stargazers"><img src="https://img.shields.io/github/stars/comisai/comis?style=flat&color=06B6D4" alt="GitHub Stars" /></a>
  <a href="https://discord.gg/FsqgJkpp"><img src="https://img.shields.io/badge/discord-join-5865F2?style=flat&logo=discord&logoColor=white" alt="Discord" /></a>
</p>

<p align="center">
  <a href="#quick-start"><strong>Quick Start</strong></a> &middot;
  <a href="https://docs.comis.ai">Docs</a> &middot;
  <a href="#how-comis-is-different">How it's different</a> &middot;
  <a href="https://comis.ai/#demo">Demo</a> &middot;
  <a href="https://discord.gg/FsqgJkpp">Discord</a>
</p>

---

## Quick Start

**One-liner** — installs Node.js and everything else (macOS & Linux):

```bash
curl -fsSL https://comis.ai/install.sh | bash
```

**Or with npm** (requires Node.js 22+):

```bash
npm install -g comisai
comis init
comis configure              # choose LLM provider, add API keys, connect a channel
comis daemon start
```

**Or with Docker** (no Node.js required):

```bash
docker run -d --name comis -p 4766:4766 \
  -v comis-data:/home/comis/.comis comisai/comis:latest-slim
docker exec -it comis comis init
```

Verify: `curl http://localhost:4766/health` → `{"status":"ok"}`. Message your agent on Telegram, Discord, or any of the other 7 channels — that's it.

> **Production = Linux host only.** Docker Desktop (macOS/Windows) auto-disables the exec sandbox — fine for dev, never for production. See [Platform Support →](https://docs.comis.ai/operations/docker#platform-support).

---

## See it in action

<p align="center">
  <em>"Have four analysts research NVDA in parallel, then run a bull vs bear debate, and let the head trader make the final call."</em>
  <br />
  <strong>One sentence. A live 7-node DAG pipeline — with isolated memory, budgets, and audit trail per agent.</strong>
</p>

<p align="center">
  <a href="https://comis.ai/#demo">
    <img src="assets/comis-dag-demo-poster.jpg" alt="Click to watch the DAG pipeline demo on comis.ai" width="100%" />
  </a>
  <br />
  <a href="https://comis.ai/#demo">▶ Watch the demo</a>
</p>

---

## Why Comis

**🛡️ Defense in depth — runtime and compile-time.** The LLM is the attack surface. Comis assumes it will be attacked, then makes the abuse expensive: kernel-enforced exec sandboxes (bubblewrap / sandbox-exec), input + output scanning, trust-partitioned memory, per-agent tool restrictions, canary tokens at runtime — paired with **22 ESLint-enforced architectural rules** that block insecure patterns at commit time. Defects don't reach production because they don't reach `main`. [Deep dive →](https://comis.ai/security)

**📦 Supply-chain integrity by default.** Every release is **sigstore-attested via GitHub OIDC**. Workspace packages are bundled with `bundledDependencies` and exact-pinned — no runtime `npm install` of plugins, no transitive surprises, no peer-dependency outages. Comis owns its domain types end-to-end (no external `pi-coding-agent` you can be held hostage by). The supply chain *is* part of the threat model.

**💰 Cost — prompt caching saves 81%.** Dual-prompt architecture, active cache fences, adaptive TTL escalation, sub-agent spawn staggering. **$5.02 vs $26.42** for a 76-call Opus 4.6 session · **94% cache hit rate** on warm turns · **$2.11** for an 8-agent pipeline (788K tokens). [Deep dive →](https://comis.ai/context-management)

**🧊 Context — scales without degradation.** Most assistants silently drop old messages. Comis never deletes a message. A 10-layer context engine evicts stale content, compacts older history into structured summaries while keeping originals retrievable (via `session_search`), then re-hydrates critical references (`AGENTS.md`, recent files, active state) and re-states the original objective after compaction so the agent doesn't drift. A background job extracts user preferences between sessions, so your agent gets better over time without being told to remember.

**🏗️ Hexagonal architecture with `Result<T, E>` discipline.** Core defines port interfaces; ad

... (truncated)