Dayze
A personal context layer for AI: a source-linked life graph that ChatGPT, Claude, Cursor, and Gemini can answer from over MCP. Built and operated end to end by our US sister company, Phi Digital LLC, with an AI agent fleet.
Visit dayze.com- Merged PRs (4 Jul – 24 Oct 2026)
- 3,568
- From agent-named branches
- 2,151
- Model providers routed
- 3+
- Read-only by default for ChatGPT
- Yes
The Challenge
Assistants like ChatGPT and Claude forget who you are between conversations. The ask was ambiguous: make AI genuinely know a person's life without turning that into a privacy liability. That meant deciding what to store, how to prove where each fact came from, and how to let third-party AI clients read it safely.
Tech Stack
The Solution
A life graph of people, plans, and moments where every record links to its source, exposed through a hosted MCP server with OAuth 2.1, per-category scopes, and read-only defaults. A task-based model router picks between OpenAI, Anthropic, and Google models, and CI gates every change on retrieval and connector evals.
The Results
Live at dayze.com and connectable from ChatGPT, Claude, Cursor, and Gemini. Between 4 Jul and 24 Oct 2026 the repo merged 3,568 pull requests; at least 2,151 (60%) came from agent-named branches.
The problem
AI assistants are only as useful as the context they have. Dayze set out to give them a person's real context (people, plans, places, moments) without becoming a data-leak risk. The hard part was not the model. It was deciding what an AI client may see, proving where every fact came from, and keeping the system safe while shipping daily.
The production system
- Life graph with provenance. People, plans, and moments are stored as linked records, each tied to its source, so answers can cite where a fact came from.
- Hosted MCP server at
dayze.com/api/mcp, with separate OAuth 2.1 audiences: a read-only directory, personal read, write, and an owner-only private surface. Access uses OAuth grants or owner-issued API keys with the same scopes. - Scoped grants.
context.read,context.write,context.delete, per-category scopes, and sensitive reads (PII, travel secrets, financial detail) that no general scope implies. - Context packs that fit a token budget, trim sections deterministically, and report every cut, so the client knows what it did not get.
- Retrieval with embeddings plus a reranking step.
- Multi-model routing. One registry maps each task to a model across OpenAI, Anthropic, and Google, plus open-weight options. The rule is the cheapest model that passes that feature's evals, and a guard test fails the build if an unlisted model ID appears.
Judgment calls
- Least privilege by default. New connections get at most read and write. Delete, mail, and sensitive reads must be ticked per connection by the owner. ChatGPT is only offered reviewed read-only scopes, because it pre-ticks everything it is offered.
- Access log. An append-only, user-visible log records which client read which categories and how many records, never the content.
- Preview before destroy. Destructive MCP tools have preview and restore counterparts.
- Human gate on auth changes. Sensitive changes wait for founder approval; for example PR #3329, titled "GATED: needs Luke's yes", tightened legacy API keys to require explicit
mail.read.
Failures we caught and fixed
- Security: agent security review found and closed cross-user data-access paths before exploitation, and connector writes are now private by default.
What the agent fleet does
AI coding agents (Codex, Claude Code, Cursor, and our own named agents) pick up issues, write the code, open pull requests, and run QA. Branch counts over the period: codex 1,843, cursor 119, claude 94, hermes 53, fizz 12, systems 4. Examples: PR #3016 (a claude/ branch) recorded live QA of ChatGPT write flows; PR #3688 (a fizz/ branch) patched a dependency security advisory. CI runs on every PR: lint, typecheck, a "does it know me" retrieval eval, a golden-question set, connector tests, an MCP smoke test, and OAuth issuance, rotation, replay, and revocation race tests.
How we count
Numbers come from gh pr list --state merged on our private product repositories, measured 24 Oct 2026. A PR counts as agent work only if its branch is named for an AI coding agent (for example codex/, claude/, cursor/, hermes/, fizz/, pollen/, honey/, systems/). Agent work on other branches is not counted, so these are floors. For CFA Mock Exam we also count commits on the main branch that carry an AI agent Co-authored-by trailer. We do not publish user, activation, or retention numbers here.
Ready to build something like this?
Start with a free audit, or the S$3,500 assessment (paid by PayNow invoice, credited in full against your build).