A coding agent that's actually yours. Self-host, BYOK, open source — Claude Code / Cursor capability on real tasks, with no caps and full cost transparency. Your code and your key never leave your machine.
zonecli.dev · GitHub · feedback@zonecli.dev
Subscription agents (Cursor, Codex, Claude Code) hide your usage behind rate limits and a flat monthly price. Zone runs on your own API key — so you pay per token, never get throttled, and see the real cost of every task.
| Zone | Subscription agents |
|---|---|
| Your key, direct to the API | Locked to one vendor's plan |
| No caps — never blocked mid-session | Rate-limited (5-hour rolling windows) |
| Every cent visible, per task | Usage hidden behind a flat price |
| Runs on your machine | Their cloud |
| Open source (AGPL-3.0) | Closed |
# 1 — install
npm install -g zone-ai-agent
# 2 — add your API key
export ANTHROPIC_API_KEY=sk-ant-... # or OPENAI_API_KEY
# or use /keys inside zone
# 3 — run, from any repo
cd ~/your-project && zonePress Shift+Tab for plan mode. You can also add and manage keys from inside the TUI with /keys — no .env needed for personal use.
git clone https://github.com/BedreddinErkan/zone && cd zone
npm install && npm run build && npm linkZone runs via WSL (Windows Subsystem for Linux). Install WSL 2 with a Linux distribution, then run zone from within the WSL terminal.
- Describe a task. In plan mode, Zone drafts a step-by-step plan and waits — you approve, refine, or give feedback before it touches a file.
- It works. It reads your code, makes AST-validated edits, and runs commands behind a safety layer — reading the output and fixing its own mistakes until the task is done.
- You see the cost. Tier-aware model routing and prompt caching keep it cheap; every run ends with exact tokens, time, model, and dollars.
- BYOK — Anthropic + OpenAI, your keys, direct.
- No caps — pay per token, never throttled.
- Cost telemetry — real dollars on every run.
- Plan mode — a read-only investigation (reads files, searches, runs
build/test/typecheck) finds the real root cause and shows a plan before it edits. Approve, refine, or reject. - Atomic safety — patches stage in memory, typecheck on the staged tree, flush all-or-nothing, and auto-rollback on failure. Your working tree never enters a half-broken state.
- Sessions — resume any run where you left off (
--resume). - Model routing — the right model per task; switch any time with
/model. - Self-hosted & open — AGPL-3.0, inspect everything. No telemetry, no vendor dependency beyond your LLM provider.
- @file context injection — type
@pathin the composer to inline a file's contents into the task; tab-complete shows matching repo files. - /image — attach a local image (jpeg/png/gif/webp) for vision-capable models with
/image <path>. - /undo — snapshot-based revert of the last run's file changes; shows a drift warning if files were edited since.
- Web fetch — the agent can fetch a URL via
fetch_url; SSRF-guarded, streaming-capped, re-validates redirects. - User hooks —
.zone/hooks.jsondefinesPreToolUseveto rules andPostToolUseside-effects; home-dir content-hash trust gate. - MCP client —
.zone/mcp.jsonconfigures stdio MCP servers; their tools register asmcp__<server>__<tool>and are available to the agent automatically.
| Command | What it does |
|---|---|
/help |
Show key bindings and commands |
/exit |
Exit zone |
/clear |
Clear transcript |
/cost |
Show session cost |
/permissions |
View and remove trusted command prefixes |
/keys |
Manage API keys (BYOK) |
/sessions |
Browse and resume past sessions |
/init |
Scaffold .zone/memory.md by analyzing repo |
/memory |
Show .zone/memory.md |
/model |
Choose AI model |
/effort |
Set reasoning effort (model-dependent: low → max) |
/summary |
Set summary format (compact/detailed) |
/plan-mode |
Set plan depth (quick/investigate) |
/session |
Toggle session memory (off/on) |
/metrics |
View run telemetry KPIs |
/limits |
Set daily USD cap |
/commit |
Commit last run's changes with scoped git commit |
/autocommit |
Toggle auto-commit after each run (off/on) |
/websearch |
Toggle web search (off/on) |
/image |
Attach a local image file to the next task (/image <path>) |
/undo |
Undo the last run (restore files to pre-run state) |
/feedback |
Send a bug report / feedback to the maintainer |
Stop guessing what your coding agent costs. Every task shows its real cost, and you only pay for what you use — for most developers, far less than a Max plan.
Real Zone runs: a small edit costs as little as $0.02; a full Astro site scaffold costs $0.09; a typical cached feature or refactor runs $0.20–0.50; large multi-file refactors top out around $0.90. At ~5 sessions/day and ~$0.30/session, that's roughly $45/mo in API spend.
Anthropic rates (Jun 2026): Sonnet 4.6 $3 / $15 per MTok · Opus 4.8 $5 / $25 · cache-hit input $0.30 (−90%). Prompt caching is on by default, so multi-iteration runs get cheaper per turn, not more expensive.
Consumer Claude plans don't include API access, and Claude Code under a plan is rate-limited. Zone is BYOK on the provider API —
/metricsshows real history from your own runs.
Run /cost or /metrics in-session for exact per-run figures; full cost logs live in ~/.zone/cost-logs/.
Keys can be set entirely through /keys at runtime — no .env file required for personal use. For headless or scripted setups:
ANTHROPIC_API_KEY=sk-ant-... # at least one provider key
OPENAI_API_KEY=sk-...
ZONE_ENABLE_MESSAGE_CACHE=1 # Anthropic prompt caching (default on)
ZONE_VERBOSE_LOGS=0 # 1 for diagnostic output
ZONE_DAILY_USD_CAP=10 # per-user daily spend cap (USD); 0 = unlimited
ZONE_ORG_POLICY_PATH=... # org-level policy file (JSON)
ZONE_EXPERIMENTAL_SYNTAX_CHECKERS= # CSV of experimental checker ids (e.g. "go,ruby,java")Zone validates patched files inline before writing them to disk. TypeScript and Python are first-class (always active). Go, Ruby, and Java are experimental and opt-in via ZONE_EXPERIMENTAL_SYNTAX_CHECKERS. Each uses a graceful skip — if the binary isn't on PATH, Zone approves the patch rather than blocking. (Rust is omitted; use run_command: cargo check instead.)
- Node.js 22+
- npm 10+
~/.zone/directory (auto-created on first launch)
AGPL-3.0-or-later. Free for personal use, modification, and distribution. Forks must release their source under the same terms; closed-source SaaS forks are not permitted.
If Zone saves you money, you can sponsor its development.
Bedreddin Erkan (@BedreddinErkan) — solo dev. Built with TypeScript, Node.js, Ink, and the OpenAI / Anthropic SDKs. Inspired by Cursor; frustrated by SaaS lock-in. feedback@zonecli.dev