All products
Personal context layer for AIJul 2026 – present

Dayze

A personal context layer for AI, built and operated by an agent fleet

dayze.com

3,568

Merged PRs (Jul 4 – Oct 24, 2026)

2,151 (60%)

From agent-named branches

3,447

Merged in last 30 days

426 PRs

Busiest day (Oct 21)

The Challenge

People want ChatGPT, Claude and Gemini to know their life (who they met, what they planned, what happened) without handing over raw data or trusting a black box. The brief was open-ended: make personal context usable by any assistant while keeping it private, source-linked and correctable.

Tech Stack

MCPOAuthNext.jsSupabase / PostgresOpenAIAnthropic ClaudeGoogle GeminiVitest

Client

Phi Digital (own product)

The problem

Assistants are only as useful as the context they have. Dayze turns a person’s people, plans, places and moments into a private, source-linked life graph that connected assistants can read from, and write corrections back to, with the owner in control.

The production system, owned end to end

  • Live at dayze.com: web app, life graph, Dayze Agent (voice and chat) and a public MCP server that ChatGPT, Claude, Cursor and Gemini connect to.
  • Phi owns product, code, infrastructure, deploys and self-hosted CI runners. Merging to main ships production.

Technical depth

  • MCP server with OAuth grants: per-connection access/refresh tokens and scopes such as context.read and context.write, with owners able to narrow to category-level scopes. Missing scopes return a 403 INSUFFICIENT_SCOPE naming what is needed.
  • Context graph: people, events, places, moments and relationships with evidence links, conflicts and temporal state.
  • Retrieval: semantic search merged with phrase-first keyword search, with keyword fallback when embeddings are unavailable.
  • Multi-model routing: 50 task types (chat, reranking, voice, transcription, embeddings) routed across OpenAI, Anthropic, Google Gemini and Kimi models from one routing table.
  • Agent QA: 2,200+ test files plus MCP smoke tests, full MCP QA runs and “does it know me” harness evals run against the live tools.

Judgment: privacy, defaults and guardrails

  • New connections get read and write only; delete, mail access and sensitive reads (PII, travel secrets, financial details) must be granted explicitly by the owner.
  • Destructive tools follow preview, then apply, with a restore window; agent-initiated previews queue for the owner to approve.
  • Responses pass a redaction contract; risky product changes ship behind flags that default off, and sensitive auth changes wait for founder approval.

Real failures caught and fixed

  • PR #488 (Oct 1): the MCP get_events tool ignored query and date filters, so assistants saw the same week for every question. Filters now apply.
  • PR #3325 (Oct 22): fast duplicate tool calls could skip the audit row in mcp_tool_calls. Caught in release verification; every call is now recorded.
  • Security: agent security review found and closed cross-user data-access paths and over-detailed error responses before exploitation, and kept merged contact records scoped to the right person.

What the agent fleet did

A fleet of coding agents (mostly OpenAI Codex, plus Claude, Cursor and in-house agents) runs in parallel lanes across three Macs. A planner turns Linear issues into tasks; agents write the code, open the PRs, run the tests and QA; a merge coordinator enforces ownership, holds and required CI; a human approves anything gated. Of 3,568 merged PRs, at least 2,151 came from agent-named branches; the real share is higher because agent work on goh/ branches is not counted.

Sources · measured 2026-10-24

  • Merged PR counts: GitHub gohluke/dayze-webapp (private), measured 2026-10-24 19:26 UTC with gh pr list --state merged.
  • Non-security PR numbers above are in the same repository; security fixes are described at category level only.
  • Fleet operations (three Macs, self-hosted CI runners, merge coordination): phi.llc docs/dayze-ops-runbook.md and speed-board/README.md.
  • Model routing (50 task IDs in lib/ai/model-routing.ts) and test files (2,207 *.test.ts/tsx) counted in the Dayze repo at commit da7992e6a (2026-10-20).

Ready to build something like this?

Start with Discovery Sprint (US$2,500, credited to the Build Pilot), email luke@phi.llc, or ask our AI agent how we would apply this approach to your workflow.

Buy Discovery SprintContact Us