ZELL is a minimal coding-agent harness in Zig. print / JSON / RPC / TUI share one core. Less is More keeps the surface small; Security First is a separately specified Guarded mode that is not available yet.
$ zell -p "What is 2+2?" 4 $ zell --mode json -p "hi" | jq -r .type zell.protocol agent_start # illustrative output · check the actual release $
Zell has no Node or npm runtime requirement. Build from source today; artifact size, linkage, and supported targets are measured per release rather than assumed from the language.
# 1. 克隆并构建（需要 zig 0.16+） git clone https://github.com/Jackallink/zell && cd zell zig build -Doptimize=ReleaseSmall # 2. 放入 PATH cp zig-out/bin/zell ~/.local/bin/
zell command right after install. All four modes live in one binary: print · json · rpc · tui, plus -c session resume. Deliberately not: no daemon, no built-in MCP, no permission-prompt stack.
Visible behavior · Zig-native runtime
Less is More keeps workflow policy out of the core. Security First is an explicit, small future boundary—not a claim that the current process is sandboxed.
Follow-up, tool use, and visible events form one documented agent loop. print / JSON / RPC / TUI expose the current implementation; behavior is verified incrementally, not claimed by slogan.
$ zell -p "重构 fetch 重试"
I added exponential backoff to the fetch retry logic.
No Node or npm runtime is required. The current baseline runs with the invoking user's authority: it has no sandbox, filesystem confinement, network policy, or key isolation. Use external OS/container isolation when required today; Guarded is planned and will fail closed until its platform backend is proven.
$ zig build -Doptimize=ReleaseSmall $ ls -lh zig-out/bin/zell # record the target, release, size and startup measurement
--mode json streams events to scripts, --mode rpc speaks stdin commands / stdout NDJSON — IDEs and CI mount directly.
$ zell --mode json -p "hi" | jq -r .type zell.protocol agent_start $ printf '%s\n' '{"id":"1","command":"prompt","text":"hi"}' | zell --mode rpc {"type":"zell.protocol","version":1,"capabilities":[…]} {"type":"response","command":"prompt","success":true,"id":"1"} {"type":"agent_start",…}
One letter, two layers: order in mathematics, life in its root.
In mathematics, the Riemann ζ function is a compact metaphor for order in complex systems. Here it is a naming reference, not an implemented multi-agent bus.
In Ancient Greek, ζάω (zao) means "to live, to be alive". It is a naming motif for a compact coding-agent runtime, not a claim of autonomous hardware, self-pruning, or self-bootstrapping features.
loop, footprint, steer, session — every one an inspectable primitive, not a buzzword. Terminal output below is illustrative, close to real command behavior.
Two visible loops: the outer drives follow-ups and convergence, the inner executes tool calls, emitting events line by line.
$ zell -p "跑通 checklist"
The checklist has 3 completed items and 1 pending item.
One Zig binary and a small dependency surface. Artifact size, linkage, memory use, and target support are release-specific measurements, not universal product guarantees.
$ zig build -Doptimize=ReleaseSmall $ ls -lh zig-out/bin/zell # publish a target-specific measurement with each release
In the TUI, Enter submits and Alt+Enter follows up; in RPC, steer and follow_up share the transport but have different queue semantics—see Protocol v1.
$ zell --mode rpc
# client sends command JSON after the startup handshake
{"command":"prompt","text":"写 is_even"}
{"command":"steer","text":"不用 % 运算符"}
{"command":"follow_up","text":"压成一行"}
Sessions persist as structured JSONL messages; -c resumes history. They are local conversation records, not protocol-event logs or immutable audit records.
$ zell --name oncall -c -p "继续上次的修复" … $ cat .zell/sessions/oncall.jsonl {"role":"user","content":"继续上次的修复"} {"role":"assistant","content":"…"}
A human in the terminal, scripts pulling data, CI assertions, IDE embedding — one core, four exits.
Streaming TUI; Enter / Alt+Enter set the pace, built for a human in the loop.
$ zell
◇ 加载 session…
▸ 提问：解释 zig 的错误联合
· (streaming…) 错误联合同时表达值与错误…
One-shot prompt, answer straight to stdout — the plainest and most composable.
$ zell -p "2+2"
4
Structured event-stream output — scripts and assertions consume it directly.
$ zell --mode json -p "hi" | jq -r .type
zell.protocol
agent_start
stdin commands / stdout NDJSON, designed for IDE and embedding scenarios.
$ zell --mode rpc
{"type":"zell.protocol","version":1,"capabilities":[…]}
{"command":"prompt","text":"hi"}
{"type":"response","command":"prompt","success":true}
{"type":"agent_start",…}
The do / don't boundary is respect for the product — Less is More, no claiming unshipped capabilities.
One Zig binary with four modes. Docs state the protocol, limits, and current security boundary.