Comb
Open-source · self-hosted · $0 local

Your company's
agentic OS harness.

Comb is Claude Code, but for your company's own agents. Pipe in your data, build agents by answering a few questions, and run them from one CLI — over a governed brain that cites its sources or refuses, respects access scopes, and never plays yes-man.

npm i -g open-company-brain && comb init

How it works

From raw data to working, governed agents — without gluing five tools together.

1

Start the brain

Pipe in any data — paste, upload, an API, or a workflow webhook (n8n). It becomes scoped, embedded, citable knowledge.

2

Create agents

Answer a few prompts and an agent is ready — it inherits citations, cite-or-refuse, and your access scopes.

3

Run & govern

The harness runs agents on any model (local or cloud), with memory, context retention, anti-drift, and a human-approved action layer.

4

Connect anything

Bring any tool, API, or MCP server (even your own) — the harness aggregates them into one governed toolset.

One harness, your model, your tools

The CLI is the harness. It runs the agents; you bring the model and the tools.

┌──────────────────────────── comb (CLI) ────────────────────────────┐ │ │ │ HARNESS ────── AI MODEL (local Ollama · or cloud) │ │ │ │ │ └──────── HARNESS TOOLS (brain · APIs · any MCP) │ │ │ │ governed kernel: embed · access-scoped recall · cite-or-refuse │ └─────────────────────────────────────────────────────────────────────┘

Why Comb

⚖️

Governed by default

Access scopes on every read, a full audit trail, and human-approved actions. The governance lives in the kernel, so every agent inherits it.

🎯

No hallucination, no yes-man

Cite-or-refuse: answers quote their sources or say "I don't know." Agents stay on task (anti-drift) and tell you the truth instead of flattering you.

🔌

Connect any model & tool

Run on local Ollama ($0/query) or a cloud model. Connect any API or MCP server — including your own — into one namespaced toolset.

🧠

Memory & context

Per-agent memory and context retention so agents get better with use, with token-aware context to keep each run efficient.

🏢

Built for organizations

Multi-scope access control is first-class — not a single-user assistant, but a substrate a whole company's agents run on.

🛠️

Yours to self-host

Open-source, runs on your infrastructure. Embed it as a library, run the API for your apps and workflows, or operate it from the CLI.

Use it from the terminal

Install once, then build and run your company's agents.

# set up (guided) — pick local $0 models or a cloud key
npm i -g open-company-brain && comb init

# bring your data in (or point a workflow/API at the ingest webhook)
comb connect knit -- npx -y knit-mcp@latest   # connect any tool / MCP
comb tools                                     # everything an agent can use

# run an agent over your governed brain + connected tools
comb run "what's open across the team this week?" --scopes my-team
comb chat                                      # interactive agent REPL