How Zest works
A session's path from your coding agent to your standup: what runs on your laptop, what the API does with it, where it is stored and what reads it.
The short version
A Zest plugin reads each session your coding agent writes to disk, redacts secrets on your laptop and uploads it to the Zest API over HTTPS. The API checks it, redacts it again and stores it under your workspace. Every minute a router turns new sessions and pull requests into session cards, timeline entries and daily totals. Standups, levels and Ask Zest are built from those.
On your laptop
- Capture. Each plugin reads the files or events its own agent already keeps, such as Claude Code's session transcripts or Codex's session logs. Its page says exactly which files and fields: Claude Code , Cursor , VS Code , Codex , Hermes , Grok Build , OpenCode , Freebuff .
- Redact. Before anything is queued, every text field (prompts, replies,
titles, folder names, metadata) goes through the same redactor, which replaces API keys,
tokens, passwords, private keys and email addresses with
[REDACTED]. The patterns. - Queue. Sessions wait in a queue file in the plugin's own folder (for example
~/.claude-zest/queue/), so nothing is lost while you are offline or signed out. - Upload. A background process sends the queue to
/api/v1/ingest, gzipped, about once a minute. A retry resends the same batch id, so a session is never stored twice. - Your settings travel with it. Every 5 minutes the plugin reads your workspace's collection settings. If you or an admin turned off prompts or agent replies, that text is left out of the upload.
At the API
One Cloudflare Worker serves the API, this site and the web app. A plugin's upload goes through these steps before anything is written:
- Check. The token must belong to a member of the workspace, and each event must match the schema. Unknown fields are rejected.
- Strip and redact again. Text for a role your workspace does not collect is dropped, and every free-text string goes through the same redactor the plugins run. This covers an old plugin, or one that skipped its own pass.
- Store. Each upload goes to R2 as one compressed file per session, and the session gets an index row in D1. A retry overwrites the same file with the same bytes.
Limits: 4 MB and 2,000 events a batch, 64 KB of text a message, and 512 MB a workspace a day.
Where it is stored
- D1 is an index. Rows for sessions, people, teams, timeline entries and daily totals, with no message text in them. Every row that holds user data has an expiry date.
- R2 holds the content. Raw uploads are kept 30 days. Each session's transcript keeps every message's text, cut to 4 KB, with no images and no model reasoning. It is kept 30 days.
- Every key starts with your workspace. The server builds every R2 key as
w/{workspace_id}/…from the workspace it has checked you belong to. No client ever sends or receives a key, and both buckets are private.
Retention for every kind of data, and how to shorten it: Security → Retention.
The router
Every minute the router picks up sessions that have been quiet for 10 minutes (or have 200 new events) and pull requests as they arrive.
- Resolve. Rules, no AI: which repo the session worked in (from git remotes, folders and edited paths), and which pull request it opened, matched by repo and branch.
- Split. A model turns the session into short pieces: what shipped, what was fixed, a decision, something that needs a call. It is the cheapest model that passes Zest Bench for this task, reached through OpenRouter.
- Check. Every project, person and pull request a piece names must exist in your workspace, and every number must appear in the session. A piece that fails is dropped, never guessed.
- Fan out. The session card, timeline entries, your stack (models, skills, MCP tools), to-dos and the daily and weekly totals. Totals are rebuilt from the sessions, never incremented, so a re-route changes nothing.
Standups
- Yours is written at your standup hour (18:00 in your timezone unless you change it in Settings → Notifications) from the day's sessions and pull requests. During the day it is refreshed as sessions arrive, at most every 2 hours, and it is sent once.
- Your team's is written at 09:00 in the team's timezone.
- The Zest standup covers what your agents did overnight, at 06:00.
- A standup can carry up to three tips drawn from your last 8 days, never one you saw in the last 14.
Levels
Every night Zest measures each person's last 14 days of agent sessions against the rubric it found across its network: agents a day, agents running at once, repos a day, how often agents verify and ship their own work, and how much of it a harness starts. A level needs every test of its own and of the levels below, and it never drops for 14 days after a rise. A team's level is the median of its members'.
- L0 Hand Coder. You write the code; agents aren't connected yet.
- L1 Pair Programmer. One agent at a time, fewer than 4 a day. You watch.
- L2 Delegator. 4+ agents a day, 2+ at once. You review.
- L3 Orchestrator. 5+ at once across 4+ repos. Agents verify and ship their own work.
- L4 Code Factory. 40+ agents a day, 10+ at once, and a harness starts a third of the work.
Ask Zest
Ask a question in the app or in Slack. Ask Zest sends your question and the Zest numbers it needs to OpenRouter, which routes it to a model vendor (Anthropic, OpenAI, Google or xAI) for the answer. Agents can read the same data
through Zest's MCP server at /mcp with a Zest API key: sessions, standups and
timelines, and they can post timeline entries.
GitHub
A workspace admin installs the Zest GitHub App and picks repositories. From then on GitHub sends pull request, review and comment events to Zest, and a history sync backfills the last 30 days of pull requests, reviews and default-branch commits. Zest uses them to tie sessions to the pull requests they produced and to count what merged and what an AI reviewer flagged. It reads with read-only tokens and never writes to a repository. Connect GitHub · what it can read.
Slack
A workspace admin adds Zest to Slack from Settings → Slack and picks a channel. Personal and team standups post there, and anyone can ask Zest a question by mentioning it or sending it a direct message; the answer comes in the thread. Slack users are matched to Zest members by their Slack-confirmed email, or by a one-time link they confirm.
Zest sends email through Cloudflare's email service, with Resend as the fallback. You choose the lists in Settings → Notifications, and each of these emails has a one-click unsubscribe:
- Daily standup — Your day, every evening.
- Team standup — What your team shipped.
- Zest standup — What your agents did overnight, at 6:00.
- Level-ups — When you reach a new level.
- Setup tips — Help connecting your agents.
- Billing — Receipts and plan changes.
Sign-in codes and account emails always send.
Zest 2 today
This is Zest 2, at meetzest.com. Plugins installed before it keep working: an older build still sends to the old system, and Zest 2 gets a copy of its data within minutes. Updating the plugin moves it to Zest 2 directly, and you stay signed in. Each plugin's page says which build is which.