workers-personal-agent

The workers-personal-agent is a personal AI agent that runs entirely on Cloudflare's free tier. It uses one Durable Object per conversation, stores long-term memory as markdown in R2, provides semantic recall from an embedding index in Durable Object SQLite, and includes a deep-research mode that fans out across child Durable Objects. The agent supports web chat over a WebSocket, where the UI holds no state and setState in the Durable Object persists and broadcasts updates. It also offers deep research, where the agent proposes a plan first, then fans out one child Durable Object per angle, each with its own fifty subrequests. Long-term memory is stored as markdown in R2, with semantic recall over an embedding index and a keyword fallback. History compacts instead of truncating: at 80% of the selected model's context window the oldest turns fold into a rolling summary, and go to an append-only archive first. Tool results outlive their round, as a fetched page is kept whole in the thread's SQLite. The agent can write skills for itself, saving workflows as markdown in the vault and listing them in every system prompt. Scheduled work includes reminders, recurring tasks, and a nightly reflection pass, all via Durable Object alarms. Several conversations are listed in a sidebar and registered in the vault. MCP servers can be added in a settings dialog with OAuth or a bearer token. Guards at the boundaries include a subrequest counter and zod schemas for tool calls. The stack includes Cloudflare Workers, Agents SDK, Durable Objects, R2, Workers AI or any OpenAI-compatible API, zod, Vue 3, Vite, Tailwind, shadcn-vue, and vitest.