Zene

Local coding agent CLI. Read, write, and edit files, run bash, and manage sessions — privately, on your machine.

curl -fsSL https://zene.sh/install.sh | sh

Or download pre-built binaries from GitHub Releases (Linux & macOS).

Refactor error handling in turn.rs to use anyhow

I'll inspect the module and apply a focused edit.

tool · Edit(crates/core/src/turn.rs)

Refactoring complete. Tests passed.

Message Zene…

What is Zene?

Zene is a local-first AI coding agent written in Rust. It runs in your project directory, reads and edits files with high precision, executes shell commands, searches the web, and keeps conversation sessions on disk.

The default local UI is the Web Agent console (via zene-gateway). Zene also supports headless prompts, Keel sandbox profiles, permission modes, stdio/HTTP MCP servers, lifecycle hooks, skills, and an ACP bridge for editors.

Keel Sandbox

File and shell access are constrained by configurable Keel sandbox profiles: off, workspace, read-only, strict, or a custom profile from ~/.zene/sandbox.toml. Credential paths such as ~/.ssh and **/.env* are denied by default. Network tools can be gated with an allow_hosts egress allowlist.

# Launch Web Agent / REPL in the current directory
zene

# Auto-approve Write / Edit / Bash
zene --yolo

# Stricter sandbox profile
zene --sandbox strict

Context Compaction

Long sessions stay within the model window. Zene tracks usage-driven context water level and applies full-replace compaction with an input ladder: verbatim → fitted → lossy. Oversized tool outputs are truncated first; older turns are summarized when needed. Recover with /compact, /rewind, and /fork.

Steerable Control

Keep the agent on track without waiting for a turn to finish. Queue follow-up instructions mid-step with /steer <msg>, or cancel the in-flight turn with Esc / /cancel.

# Inside a session, while Zene is running:
/steer check if compiling before running tests

Planning Mode

For complex work, switch to /plan. Actuators lock to read-only tools while Zene explores the workspace and writes a proposed plan under .zene/plan.md. Returning to coding mode requires explicit confirmation.

Extensible with MCP & Skills

Register MCP servers over stdio or HTTP in ~/.zene/mcp.json (or project .zene/mcp.json). Probe connectivity with zene mcp doctor. Add skill instruction files under .agents/skills/*/SKILL.md.

# ~/.zene/mcp.json
{
  "servers": {
    "my_database": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-postgres",
        "postgresql://localhost"
      ]
    }
  }
}

Commands

Command	Description
zene	Launch the local agent in the current directory
--yolo	Auto-approve Write / Edit / Bash
--sandbox <profile>	Keel profile: off, workspace, read-only, strict, or custom
-p / --prompt	Run a single headless prompt; optional --output-format json
--worktree	Run the session in a dedicated git worktree under .zene/worktrees/
sessions	List saved conversation sessions for the current directory
--session <id>	Resume a previous conversation session by ID
mcp doctor	Probe configured MCP servers
acp	Speak Agent Client Protocol over stdio JSON-RPC

Configure

Configure Zene globally in ~/.zene/config.toml or override per project in .zene/config.toml. Set provider keys in the file or via environment variables. Custom Keel profiles live in ~/.zene/sandbox.toml.

# ~/.zene/config.toml
provider = "openai"     # "openai" or "anthropic"
model = "gpt-4o"
include_workspace_context = true
permission_mode = "default"   # default | accept_edits | dont_ask | bypass

[sandbox]
profile = "workspace"         # off | workspace | read-only | strict | custom

[compaction]
trigger_ratio = 0.85
min_keep_messages = 20

[web_search]
provider = "tavily"           # or "duckduckgo"