Dovsky — Development skill for Claude Code
Local-first control plane for Claude Code and Codex CLI: sandboxed agent jobs, immutable evidence, and human-gated acceptance.
How to install Dovsky
This entry records only its repository, not the path inside it, so there is no
exact command to give. Open quant-dhawan/dovsky and copy the folder into
~/.claude/skills/, or the file into ~/.claude/agents/.
What Dovsky does
Local-first control plane for Claude Code and Codex CLI: sandboxed agent jobs, immutable evidence, and human-gated acceptance.
Alternatives in Development
- Zeron — A native control plane for Claude Code, Codex, Cursor, Devin and other coding agents 1.4k ★
- OpenSquirrel — For people who get distracted by agents 1.3k ★
- Agentcontrolplane — ACP is the Agent Control Plane - a distributed agent scheduler optimized for simplicity, clarity, and control 463 ★
README
Dovsky
**Run Claude Code and Codex CLI like a disciplined team — sandboxed, evidenced, and human-accepted.**
[](https://github.com/quant-dhawan/dovsky/actions/workflows/ci.yml) [](LICENSE)   
[Why Dovsky](#why-dovsky) · [How it works](#how-it-works) · [Quick start](#quick-start) · [Architecture](#architecture) · [Safety](#safety-guarantees) · [CLI](#cli-command-groups)
Dovsky is a local-first control plane for human-led Claude Code and Codex CLI work. One daemon owns provider processes and SQLite state; the `dovsky` CLI submits work and makes jobs, evidence, failures, results, and handoffs inspectable.
It is deliberately single-user and single-host. Runtime state and artifacts stay outside Git under `${DOVSKY_HOME:-~/.dovsky}`. Configuration defaults to `~/.config/dovsky/config.json`; keep configuration and any credential files outside Git with operator-only permissions.
Why Dovsky
Coding agents are fast, but "the agent said it's done" is not a release process. Dovsky puts one accountable loop around every agent run:
| Without Dovsky | With Dovsky |
|---|---|
| Agents edit your working tree directly | Each attempt runs in a private bubblewrap view and its own systemd user scope; only the captured delta is applied |
| "Done" means the model stopped talking | Host gates (--verify, --red-before, --protect, quality commands) must pass on the real tree |
| Crashed or orphaned agent processes are guesswork | Execution leases record PID, boot ID and start ticks; anythi |
Related Skills
Do SOUL Alaya
The local-first memory plane for CLI coding agents. Codex and Claude Code attach via MCP; SQLite-backed, evide
Agent River
Drive Codex & Claude coding agents from Telegram — approval-gated, human-in-the-loop, with zero runtime depend
Lacp
Local-first policy, evidence, memory, and recovery controls for Claude, Codex, Hermes, and other CLI coding ag
Jmunch Console
Local, opt-in control-plane GUI for the jMunch suite (config, index/watcher health, savings/ROI, sessions, age
Omk
Evidence-gated runner for Codex, Claude Code, OpenCode, and local coding agents. Routes tasks into scoped DAG
Agent Knock Knock
Control local Codex and Claude Code from OpenClaw through shared tmux terminals, with seamless human-agent han
Related Agents
Tsk - AI Agent Task Manager And Sandbox
by dtormoen - A Rust CLI tool that lets you delegate development tasks to AI agents running in sandboxed Docke
Paperclip Reviewer
Paperclip 2026.529+ governance-first integration platform code review specialist — control plane + adapters tw
Codex Agent
Runs OpenAI Codex on the user's ChatGPT subscription. Three modes - consult (OS-sandboxed read-only analysis,