Krowk 发布编码智能体 harness：会话可跨模型与机器同步

A coding agent harness: one session, on any machine, model or agent.

Krowk is an agent harness built for power users who run coding agents all day. It runs Claude, GPT, Grok or any OpenAI-compatible model, using your existing subscription or an API key, or drives Claude Code and Codex. A session can move to another model or agent mid-conversation, and syncs to your other machines end-to-end encrypted. It follows Claude Code's permissions, skills and hooks, keeps track of every agent session on your machine, and publishes screenshots, diffs and logs as links that unfurl in GitHub, Slack, Linear and Basecamp. It assumes you're comfortable in a terminal, with git and with configuration files.

krowk connect            # sign in with a Claude, ChatGPT or SuperGrok subscription, or an API key
krowk                    # open the agent
krowk -p "fix the build" # run one prompt headless (--output-format json)
krowk --resume           # continue an earlier session

Install

curl -fsSL https://krowk.com/install | bash
npx @krowk/cli
cargo install --locked --git https://github.com/krowkcom/krowk --features harness krowk

Builds are published for Linux and macOS (amd64/arm64) and Windows (amd64). On a workstation the installer gives you the full build, which includes the agent. In CI and containers it gives you the lean build, which only publishes. Pass --full or --lean to choose (curl … | bash -s -- --full).

The agent

• A prompt that stays put. The agent takes the whole terminal, and the prompt and status line stay on the bottom rows while you scroll the conversation with the mouse wheel or PgUp/PgDn, and dragging over it copies. When you leave, the conversation is printed into your normal scrollback. Choose Screen: inline in /settings (or set "tui": {"screen": "inline"} in ~/.krowk/config.json) to draw at the bottom of the terminal instead, with everything in your normal scrollback as it streams, which suits SSH, tmux and phone terminals.
• Tools. It can read, write, edit, run bash, grep, glob and publish. It also keeps a todo list, runs subagents in parallel, each with its own context and a cheaper model, and asks you with options to pick from when a choice is yours.
• Compatible with Claude Code. It uses the same permission modes, allow/ask/deny rules and settings files. It reads AGENTS.md, CLAUDE.md and .cursor/rules, and loads Claude-format skills, hooks and .claude/agents without changes. It adds an unhinged mode that never asks for permission.
• Multiple accounts. You can connect several accounts per provider and rename them in /connect. A second account is added with --name work.
• Cost limits. --max-usd and --max-tokens stop a session before it goes over a limit, subagents included.

In the prompt, / lists the commands and ? shows the keys.

Worktrees

Agents running at once can each work in a git worktree of their own, on a branch krowk/<hex>, so they don't overwrite each other's files.

• Subagents. isolation: "worktree" on the subagent call, or isolation: worktree in an agent definition's frontmatter, runs the child in a new worktree started from your files as they are, uncommitted changes included. When it finishes, an unchanged worktree is removed; a changed one is applied back to your working tree (never your index or HEAD) and removed. A conflict, a submodule change or a change to .git, .claude, .codex or .krowk keeps it, and the summary names it.
• A session of its own. krowk --worktree (or krowk -p --worktree) starts the session in a new worktree. On exit an unchanged one is removed and a kept one is named.
• Ready to build. Submodules are initialised from your checkout, target and node_modules are cloned in where the file system makes that nearly free (worktrees.seed), the ignored files your .worktreeinclude lists (.env, local config) are copied, and worktrees.setup runs a command such as npm ci in the sandbox. A repository's own setup command runs only once you trust it. On btrfs a worktree is a snapshot, made in tens of milliseconds whatever the repository's size.
• The machine's limits. Build and test commands (cargo, make, npm, pytest, …) from every krowk on the machine queue for builds.slots slots, a quarter of the cores by default. Subagents and --worktree sessions share subagents.maxHost agent slots. A worktree is refused when its disk is short of 4 GiB beyond twice its expected size.

{ "worktrees": { "setup": "npm ci" }, "builds": { "slots": 4 }, "subagents": { "maxHost": 16 } }

scripts/bench-worktrees measures all of this on this repository: 20 worktrees made and checked at once, and three subagents applied back.

Sessions

The full build keeps a local store of your agent sessions, whether they came from krowk, Claude Code, Cursor or opencode.

Sync

Sessions sync between your machines end-to-end encrypted: neither krowk's servers nor its relay can read them. Run a session on one machine, then follow it and prompt it from another.

Publishing

krowk push screenshot.png --pull-request="https://github.com/acme/storefront/pull/412"

✓ Uploaded screenshot.png → https://krowk.com/a/art_2e1d
  412 KB · run run_8Kd2wq

Each link shows a card with the repo, commit, branch, PR, session and agent, all detected from git and CI. You can push without a key: anonymous uploads last 24 hours and can be claimed into a workspace with krowk claim.

Useful push flags include --private, --caption, --link, --metadata key=value, and --destination github|slack|linear|…, which prints the paste format that tool expects. Metadata is public, so never put a secret in it.

In CI, the repository doubles as a GitHub Action:

- uses: krowkcom/krowk@v0
  id: krowk
  with:
    files: screenshots/**/*.png
    token: ${{ secrets.KROWK_TOKEN }}

It outputs urls, markdown (ready to post as a PR comment), run-slug and json, and adds the links to the job summary.

Using krowk from other agents

• skills/krowk/SKILL.md teaches another agent how to use krowk. The installer adds it to ~/.claude/skills.
• In Claude Code, /plugin marketplace add krowkcom/krowk and then /plugin install krowk@krowk add the skill and the MCP server together. When krowk drives Claude Code itself, the plugin's server stays off and krowk's own tools are used.
• Output is JSON when piped. krowk help --json describes every command, and --jq filters the output without needing jq installed: URL=$(krowk push shot.png --jq '.data.artifacts[0].url').
• krowk-mcp serves the same features over MCP stdio, for any other MCP client, or for Claude Code without the plugin: claude mcp add krowk -- krowk-mcp.
• Exit codes: 0 ok · 1 bad command · 2 not found · 3 needs credentials · 4 refused · 5 rate limited · 6 transfer failed · 7 server error · 8 gone.

Configuration

Everything krowk keeps is in ~/.krowk/ (move it with KROWK_HOME): settings in config.json, keys in credentials.json, and sessions in sessions/. A repository can commit .krowk/config.json to pin a workspace or add permission rules.

MCP servers

krowk's own agent uses the MCP servers you already set up for Claude Code: mcpServers in ~/.claude.json (user and per-project) and in Claude's settings.json, plus mcpServers in ~/.krowk/config.json. A repository's .mcp.json is used only after you trust the repository, and your own servers of the same name win over it. Each server takes Claude Code's shape, with ${VAR} and ${VAR:-default} expanded:

{
  "mcpServers": {
    "github": { "command": "github-mcp-server", "args": ["stdio"], "env": { "GITHUB_TOKEN": "${GITHUB_TOKEN}" } },
    "docs":   { "type": "http", "url": "https://mcp.example.com/mcp", "headers": { "Authorization": "Bearer ${DOCS_TOKEN}" } }
  }
}

The model sees two tools, mcp_search and mcp_call, however many MCP tools there are, so they cost no context until one is used. No server starts until the first search. Permission rules name a tool as Mcp(server:tool) or mcp__server__tool, and Mcp(server) names all of a server's tools; a server denied whole is never started.

Development

make check   # clippy, unit tests and golden tests
make dev     # fast build, linked into ~/.cargo/bin
make mock    # local registry; run any command with --dev

See CHANGELOG.md for what changed in each release, and crates/krowk-bench for the performance and size budgets.

Who uses Krowk?

• Primevise
• Rinkta

License

MIT