A capsule is a seed. It carries everything an autonomous agent needs — intent, identity, methodology, and memory structure — and grows on any compatible machine.
A capsule is a directory that defines an autonomous agent. You declare what you want the agent to do in a manifest. The intent computer gives it knowledge, an engine, and a machine to run on. One command deploys it:
Four phases, modeled on seed germination:
A capsule = intent + identity + methodology + context. The methodology IS the product.
The thinnest layer in the stack. Just a declaration of what you want to happen:
Skills activate on conditions, not just schedules:
0 6 * * 1 is Monday 6am
Process invoices, categorize transactions, generate financial reports
4 skills — email → ingest → categorize, weekly reports, daily anomaly detection
Crawl client sites, analyze SEO health, deliver actionable audit reports
4 skills — weekly crawl → analyze → report, daily issue monitoring
Curate content ideas, draft platform-specific posts, optimize engagement
4 skills — daily curate → draft, weekly calendar, Friday engagement review
Identify, qualify, and nurture leads through research-driven outreach
5 skills — prospect → qualify → enrich → outreach, weekly pipeline report
The same capsule runs on an exe.dev VM, your MacBook, a Linux server, or a Docker container. The runtime detects the substrate and adapts: systemd on Linux, launchd on macOS, cron as fallback. Gateway API on exe.dev, Claude CLI locally. The capsule author writes methodology once.
This scaffolds a capsule directory with an empty manifest, identity template, and vault. Fill in the intent. Write the identity. Add skills as markdown files. Deploy.
A skill can be as simple as "when email arrives, extract the invoice data" or as involved as a multi-step research workflow with web scraping, analysis, and reporting. The instructions are natural language. The LLM is the runtime.