01 — Visibility
Review what needs it
Risk and confidence stay separate. High-risk nodes rise even when the agent is “sure.”
Open for early testing
A coding agent that knows what it doesn’t know
Z attaches structured, reviewable uncertainty to the code it touches — and learns reusable skills from your work so it doesn’t ask the same questions twice.
Why Z
Wrong changes look like correct ones until something breaks
Agents get more autonomy every month. The failure mode isn’t occasional mistakes — it’s that a bad edit is indistinguishable from a good one until production. Z makes the difference visible before you ship.
02 — Signals
Checkable, not vibes
Detectors read tests, paths, APIs, and blast radius — not fake confidence percentages.
03 — Memory
Teach once, reuse
Skills capture how your repo actually works so the same conventions aren’t re-explained every session.
How it works
Detectors build a tree you can act on
As Z edits your repo, it runs checkable detectors — tests, path keywords, live API verification, blast radius — and builds a tree you can act on.
01
You describe the task
Z decomposes the request into a checklist so requirement gaps surface before they ship.
02
Detectors fire on real signals
Missing tests, unverified APIs, migrations, TODOs, shared logic — each becomes a typed node.
03
You act from the tree — or Z does
Browse risk-first, or let Z take a bounded self-correction pass on the findings it's confident enough to fix on its own.
Skills
Reusable skills
Coding agents forget your project’s rules the moment the chat ends. Teach Z once how your repo works — it remembers, and applies the right playbook on the next matching task.
01 — Create
Paste or capture
Drop in a playbook with /skills add, generate one with z skill create, or let Z offer to save after a solid turn.
02 — Index
Metadata inferred for you
Z writes when to use it, what triggers it, and where the body lives — so you don’t fill out a form.
03 — Apply
Auto-applied later
On a matching task, Z pulls the skill in automatically. No manual attach. No re-explaining Stripe webhooks every session.
Detectors
What gets flagged
Risk and confidence stay separate. High-risk nodes rise to the top even when confidence is high.
Payment / auth / securityPath and symbol keyword match — auto Medium risk at minimum.
Missing or failing testsRelevant suite by module/symbol — fail escalates; none flags Missing Test.
API assumptionNo live-verified call this session, or unverifiable MCP result.
Pattern inconsistencyNo match or conflicting conventions for a new file.
Shared logic / blast radiusReference count above a configurable threshold.
Unverifiable configEnv vars and secrets the agent cannot inspect.
Migration riskSchema changes plus explicit existing-data impact.
High confidencePositive signal when a change matches a tested pattern and passes.
Reinvented solutionsHand-rolled IP/email/date parsers — flagged when the surrounding code is actually about that domain.
Missing sibling registrationA new plugin/handler that matches a family but never got added to the shared registry.
Permissive shortcutsgetattr defaults and broad excepts that quietly absorb a real failure.
Scope driftEditing the wrong file for the right reason, or continuing after the task is done.
Model access
Bring your own model, or let Z route for you.
Bring your own key
Free, forever
Connect Anthropic, OpenAI, DeepSeek, Groq, Gemini, Kimi/Moonshot, or a local Ollama server. Your key goes straight from your machine to the provider — Z never sees it.
Z's router
Automatic routing
Z picks the right model for each task automatically and manages billing across providers for you.
See pricing →Install
Try early testing
Install now and run Z in a project. Bring your own model keys. Rough edges are expected — that’s the point of early testing.
curl
$ curl -fsSL https://raw.githubusercontent.com/Nate-git05/z/main/install.sh | shpip
$ pip install -U "git+https://github.com/Nate-git05/z.git"macOS / Linux. Puts z on your PATH. Then set your API key and run z in a project.