Website-Builder Agent Base, Claude Code Skills & Vendor CLIs
An AI-cofounder / autonomous-startup-builder SaaS
Overview
The base repository cloned into each project’s workspace to define the website-builder (“CEO”) agent for an AI-cofounder / autonomous-startup-builder SaaS. The entire agent behavior lives as Claude Code configuration under ceo/, with build-time vendor CLIs for company functions kept at the repo root.
Why It Exists
The platform runs many isolated company projects, each driven by an autonomous agent. Rather than hard-code that agent in application code, its behavior is expressed as a versioned repo, CLAUDE.md, skills, subagents and settings, so the agent is reproducible, reviewable, and updatable like any other codebase, and so the sandbox can simply git-clone it.
What We Built
A repo-as-agent-config package. The agent’s behavior is contained in ceo/ (ceo/CLAUDE.md, ceo/.claude/, a site template, a workbench and scripts), explicitly structured so the Agent SDK’s cwd resolves to that subdir and picks up the project settings, skills and subagents (researcher, ui-review). A vendor/ tree bundles build-time CLIs covering company surfaces, ads, tasks, web, leads, crm, scheduler, social, site, files, sprint, assets, thread, email. An INTEGRATION.md specifies the contract with the e2b sandbox and bridge: the repo is cloned to /home/user/company, run from ceo/ via the bridge using @anthropic-ai/claude-agent-sdk (not the bare CLI), with settingSources: ['project'] and the claude_code preset pinned for deterministic loading across SDK versions. GitHub Actions live under .github/.
Technologies & Approach
Built on the Claude Agent SDK and Claude Code’s project-settings model. The key design choice is putting agent config and skills in a git repo so each company’s agent is a clonable, auditable artifact; the integration contract pins SDK defaults explicitly because implied settingSources behavior has varied across SDK releases, ensuring identical behavior in every sandbox.
Outcome / Impact
Gives every autonomous company project a consistent, version-controlled agent definition with a documented sandbox contract, decoupling agent behavior from application code and making it independently testable and upgradeable.
Capabilities Demonstrated
- Repo-as-agent-config (“agent in a repo”) pattern
- Claude Code skills, subagents and hooks loaded from project settings
- Documented e2b sandbox + bridge integration contract
- Bundled vendor CLIs spanning ads, CRM, social, web and more
- Deterministic Agent SDK session bootstrap across versions