Sylph

The open-source company brain.
Run your entire company with AI agents, skills, and a self-improving context.

Created by nao Labs (YC X25) to run our company.

What is this?

Sylph is a repo to help you build your company brain: a repo of all your context by domains, skills to perform actions with or for you, agents that will act as real AI employees (the Sylphs), and a core principle: a self-improving loop.

It is agent agnostic, can be used in Claude Code, Codex, Cursor, or any AI coding agent that can work on a git repo.

I wrote the full story of how it works and how I built it on my substack here: I built a company brain to run my YC startup.

What you can do with sylph

Sylph works with all domains of your company: product, marketing, sales, legal, recruiting, etc.

You can:

• Automate processes: creating content, events, brand assets, product issues, outbound campaigns etc.
• Gather information: run reports on client activity, product activity, or domain progress on your roadmap
• Plan work: structure scattered to-dos and client feedback into weekly and daily todos
• Have agents working for you: run agents in the background to gather information and prepare tasks for you
• Build and auto-improve your knowledge base: every domain knowledge is stored in the repo and is auto-enriched
• Create any new skill you need: with /sylph-create-skill
• Collaborate with your team: use the same repo for all your team to work on common knowledge

Sylph main concepts

CAO: Chief Agent Officer - that's you.

AGENTS.md: The single source of truth for all agents, explaining how to work in the company brain.

CONTEXT.md: Your company's identity - what you do, who you serve, how the product works, etc.

Domain context: Each business domain has its own folder. This is where domain-specific knowledge lives and grows.

Skills: Processes to perform a task according to your company context. They all end with a self-improving loop.

Agents: AI employees with defined roles, routines, and assigned skills. Like superheroes with specific superpowers!

Connectors: MCP integrations that give agents access to external tools.

Self-learning: The core principle. After you approve an output, the skill diffs what it drafted vs what you kept, and rewrites its own rules to match your edits. Every cycle makes the next one better.

Quick start

1. Fork this repo

gh repo fork getnao/sylph --clone
cd sylph

2. Run the setup skill

Open the repo in your AI agent and run:

/sylph-setup

This will personalize sylph to your own personal and company context.

3. Set up skills

Skills are pre-configured but can be personalized to your own setup - just set them up as you need them.

/sylph-setup-skill linkedin

4. Configure your agents

Agents in the repo are already pre-configured, but you can personalize their tasks and scopes with this skill:

/sylph-setup-agent chief-of-staff

5. Try it

/linkedin-writer Write a post about [topic]
/create-campaign Target Series A data teams
/hr-screening Screen applications for senior engineer
/investor-update Draft the May update

Directory structure

AGENTS.md           - canonical instructions for all agents
CONTEXT.md          - global company description and context

.claude/
  skills/           - all skills for your company

agents/             - AI employees with defined routines and skill assignments

domain folders:
brand/              - brand assets, guidelines, voice
content/            - all content by channel
sales/              - outbound sequences, proposals, CRM
customer-success/   - meeting follow-ups, user research
product/            - roadmap, issues
finance/            - investor updates
partnerships/       - consulting, freelance, influencer
events/             - hackathons, meetups, remote
admin/              - internal ops (email, Slack, CRM)
hr/                 - hiring, screening
legal/              - contracts and legal
compliance/         - compliance notes

Skills

Skills are detailed, self-improving instruction files. They live in .claude/skills/ and are invoked via slash commands.

Domain	Skills	What they do
content	linkedin, blog, substack, newsletter, x, reddit, website, slack-community	Content creation per channel
sales	create-campaign, crm, customer-report	Outbound campaigns, CRM, client briefs
brand	brand-guidelines, brand-designer	Brand assets and guidelines
finance	investor-update	Investor updates
hr	hr-screening	Application screening
product	create-issue	GitHub issue creation
events	event-manager	Event creation and management
legal	create-contract, review-contract	Contract drafting and review
ops	zero-inbox	Inbox triage
writing	email-writer	Email drafting
sylph	sylph-setup, sylph-setup-agent, sylph-setup-skill, sylph-create-skill	Initial setup, agent configuration, per-domain setup, skill creation

Agents

Agents are AI employees that run on schedules. Each has a role definition and a recurring routine.

Agent	What it does	Cadence
Chief of Staff	Inbox triage, project pulse, weekly planning, daily briefing	Daily
CMO	Content planning, drafting, docs maintenance, release comms	Daily
Product Manager	Issue creation from feedback, PR monitoring, daily recap	Daily
Customer Success	Follow-up drafting from CSM backlog	Weekly
Head of Data	Analytics reports, metric monitoring	Daily
Head of Sales	Pipeline management, outbound campaigns	Weekly
Executive Assistant	Admin, HR, finance, accounting ops	Daily
Brand Designer	Visual and motion assets	On-demand

Customization guide

You can add new domains by creating folders in the folder root. You can create new skills & agents with the /sylph-setup-skill <name> skill.

Why "Sylph"?

Sylphs are invisible spirits of the air - creatures that shape the world without being seen. They were made famous by the ballet La Sylphide (the author may or may not be a ballet nerd). That's what this repo does: AI agents running your company behind the scenes, while you stay in control.

Security

This repo contains no backend, no tracking, no telemetry, and no network calls. It's just markdown files that AI agents read. No code phones home, no data is sent anywhere, and no credentials are bundled. See SECURITY.md for the full audit and verification commands you can run yourself.

License

MIT - use it however you want.