How Your Agent Works
What your agent does with CofounderOS through MCP or the CLI: which company it picks, which guides it reads, what it asks you before acting, and how it reports progress.
Besides tools, Cofounder gives your agent guides to read, company context to check, and rules about what needs your approval. You don't need to memorize any of it. Your agent reads the same guidance through MCP or the CLI, and knowing what it does tells you what to expect and what to ask for.
| Step | MCP | CLI |
|---|---|---|
| Get started | ||
| Check who is signed in | ||
| List your companies | ||
| Read the company overview | ||
| Read the roadmap | ||
| Find and read guides | , | , |
| Read recent events | ||
| Post progress |
The sections below name the MCP tool. A CLI agent runs the matching command.
Getting Started
Ask your agent to get started with Cofounder. It reads the onboarding state, which checks your account and says what comes next: sign in, choose a company, create one, start setup, or begin work. Your agent follows that answer. Ask again whenever you're unsure where things stand.
Choosing Your Company
When you sign in with your own account, the agent checks who you are with
user_get and which companies you belong to with company_memberships_list.
It uses the company you name, or your only company, and asks by name when
several could match. It never asks you for an ID.
A company API key is already bound to one company, so an agent using one skips those lookups.
Reading the Guides
Your agent reads Cofounder's guides before it acts. It starts with
cofounder-company-work, then reads the guide for the task:
| Guide | Read it for |
|---|---|
| Creating or resuming a company and getting started on a task | |
| Working on a company: context, guide choice, approvals, and progress | |
| Tool calls that fail, go missing, or lose their sign-in | |
| Cloning, branching, deploying, and handing off the product repository through the CLI, plus adding sign-in, billing UI, or the brand theme to a Next.js app | |
| The same repository work over Git for agents connected through MCP | |
| Finding and saving company documents and brand assets | |
| Contacts, accounts, pipelines, and communication history | |
| Transactional email, one-to-one mail, and marketing campaigns | |
| Preparing, scheduling, and publishing social posts | |
| Stripe setup, going live, and billing in your app | |
| Attaching a Supabase project you already have | |
| Google, GitHub, or other sign-in for your app |
MCP agents read them as resources (skill://<name>/SKILL.md). CLI agents
use cofounder skills show <name>, and cofounder instructions install-skill
installs cofounder-company-work for a terminal agent. The live catalog
may list guides newer than this page.
Checking Company Context
Before deciding what to do, the agent reads only what the task needs:
company_getfor the company overview, what is ready, and a one-line brief of what matters now.roadmap_getfor the recommended next milestone.resources_listandresources_getfor databases, domains, inboxes, and other resources.events_listfor recent work from you and other agents.
What Needs Your Approval
Your agent does what you ask and checks with you on the rest:
- It buys, provisions, and moves money only when you ask.
- When a call fails for billing or permission reasons, it follows the returned recovery step instead of retrying.
- Steps only you can do, like signing in to a provider or confirming payment, come back as a link. The agent hands it to you and waits.
Following Through
Many responses include next_steps: the call to make next, or the step you
need to take. Long-running calls return right away, and the agent checks the
matching status call before it says the work is done.
The agent can also show you views instead of describing them. company_app_open
opens the company Home, Inbox, Messages, Contacts, Roadmap, Domains, Wallet,
Files, Analytics, or Social inside hosts that display MCP apps.
Reporting Progress
The agent posts short updates with events_post when it finds something,
makes a decision, hits a blocker, or finishes work. Other agents on the same
company read them with events_list, so work carries across sessions and
tools.
When You Ask What to Do Next
Ask "what should I work on next?" and the agent reads the company brief and the roadmap. It suggests two or three concrete tasks, recommends one, and waits for you to pick.