---
title: "Krowk — the portable coding agent for Claude, GPT and Grok"
canonical_url: "https://krowk.com/"
last_updated: "2026-10-01T19:45:03.721Z"
meta:
  description: "A fast, secure coding agent harness for your terminal. Run Claude, GPT, Grok or any OpenAI-compatible model on your subscription or an API key, switch models mid-conversation, and publish the work as links that unfurl in GitHub and Slack."
  "og:description": "A fast, secure coding agent harness for your terminal. Run Claude, GPT, Grok or any OpenAI-compatible model on your subscription or an API key, switch models mid-conversation, and publish the work as links that unfurl in GitHub and Slack."
  "og:title": "Krowk — one coding agent, every model"
---

**A coding agent harness for people who live in one. **

# **One agent. Every model. Wherever you work. **

Run Claude, GPT, Grok or any OpenAI-compatible model on the subscription you already pay for. Switch models mid-conversation without starting over. It reads your Claude Code setup as it is, starts in under 150 ms, and publishes its work as links your team can open.

Quick start

**macOS, Linux and Windows. MIT licensed. **[~~Other ways to install~~](https://krowk.com/docs#install)**. **

first run

```
# Install krowk
$ curl -fsSL https://krowk.com/install | bash

# Sign in to Claude, ChatGPT, Grok, or add a key
$ krowk connect

# Open the agent in this repository
$ krowk
```

**Claude**** ChatGPT**** Grok**** OpenRouter**** Claude Code**** Codex**** Any OpenAI-compatible**

**Portability**

## **Your work is not locked to one vendor, one machine or one app. **

A vendor's agent runs that vendor's models. krowk keeps one session, one set of permissions and one history, whichever model is answering and wherever you pick it up.

### **Every major model**

Claude, GPT, Grok, OpenRouter or any OpenAI-compatible server, on the Claude, ChatGPT or SuperGrok subscription you already pay for, or an API key. It can drive Claude Code and Codex too.

### **One conversation, any model**

Hit a rate limit or want a second opinion? Switch model, account or engine mid-session and the conversation comes with you. No copy-paste, no starting over.

### **Every operating system**

macOS, Linux and Windows, as one static binary with no runtime.

**Next**iOS and Android.

### **Every client**

The terminal: an inline TUI that keeps your scrollback, and a headless mode for scripts.

**Next**Desktop and mobile, on the same sessions.

**Later: cloud runners.** The same engine on machines we run, so agents keep working with nobody attached, taking a brief to a shipped result while you review the evidence.

**The agent**

## **Built for the tool you keep open all day. **

A harness sits in every keystroke and runs commands with your credentials. Fast and safe are the floor, not features.

**Performance**

### **Under 150 ms to a prompt**

Cold start is a CI budget, not a benchmark we ran once. Nothing before the first frame waits on a vendor.

### **Zero work when idle**

No polling, no timers while it waits for you. Under 30 MB of memory idle, one static binary on disk.

### **Your terminal, not a canvas**

Inline, not full-screen: finished output lands in your normal scrollback, searchable and copyable, over SSH and in tmux.

**Security**

### **Claude Code's permissions**

The same modes, the same allow, ask and deny rules, the same settings files. Deny beats allow in every mode but unhinged, the one you opt into.

### **Commands read as the shell reads them**

A line is split into every command it runs, and an allow rule never covers one krowk cannot read. It covers git only in forms known to run nothing else.

### **Your keys stay out of reach**

krowk's home and credential files are fenced from the agent's own file tools unless you say so, and an untrusted repository cannot widen its permissions.

### **A budget the engine enforces**

Cap a session in dollars or tokens. The limit is checked before every model call, not by a hook after the money is spent.

**Your setup, as it is **

- Reads AGENTS.md, CLAUDE.md and .cursor/rules
- Loads Claude-format skills, hooks and .claude/agents
- Subagents in parallel, each with its own context
- A todo list the model keeps current
- Every session on this machine in one list, from krowk, Claude Code, Cursor or opencode
- Publishes its screenshots, diffs and logs as links

**Publishing**

## **It shows its work, as a link your team already opens. **

Screenshots, recordings, reports and generated pages become permanent links that unfurl as a card in Slack and Basecamp and render inline on GitHub and Linear. The card names the repo, commit, branch and agent, so nobody has to ask what produced it.

**Screenshot · **checkout-after.png

krowk.com/a/9f3c2e1

**Checkout, mobile 390×844**

The frame after the fix. Discount now applies before tax.

<dl>

<dt>agent</dt>
<dd>claude-code</dd>

<dt>repo</dt>
<dd>acme/storefront</dd>

<dt>task</dt>
<dd>#412 discount codes</dd></dl>

**Screen recording · **add-to-paid.mp4

krowk.com/a/4d7b8a2

**Add to cart → paid, 12s**

The whole flow, recorded end to end with no manual steps in between.

<dl>

<dt>agent</dt>
<dd>claude-code</dd>

<dt>repo</dt>
<dd>acme/storefront</dd>

<dt>task</dt>
<dd>#412 discount codes</dd></dl>

**Markdown · **run-report.md

krowk.com/a/b1e60c4

**Run report, #415 coupon rules**

What changed, what it ran, and the two tests that came back red.

<dl>

<dt>agent</dt>
<dd>codex</dd>

<dt>repo</dt>
<dd>acme/storefront</dd>

<dt>task</dt>
<dd>#415 coupon rules</dd></dl>

**HTML · **coverage.html

krowk.com/a/e7a91d3

**Coverage report, discount module**

The generated page itself, rendered in the card. Click through and it runs.

<dl>

<dt>agent</dt>
<dd>github-actions</dd>

<dt>repo</dt>
<dd>acme/storefront</dd>

<dt>task</dt>
<dd>nightly · main</dd></dl>

### **When to publish **

krowk's agent has publishing built in; this one line in CLAUDE.md or AGENTS.md tells it — and Claude Code, Codex, Cursor or a CI step — when to use it. No account for the first push. CI and containers get a lean build that only publishes.

CLAUDE.md, AGENTS.md, or a CI step

```
After any visual change, capture it (screenshot if static, a short screen recording if it is an interaction) and publish it with the krowk skill. If the skill is missing or fails, run `krowk help --json`, then `krowk push FILE --title "what changed" --json`, and paste `paste.url` where you report the work.
```

**Quick start**

## **Three commands to your first turn. **

Install, sign in to what you already pay for, and open it in a repository.

1 — Install krowk

```
$ curl -fsSL https://krowk.com/install | bash
```

2 — Sign in to Claude, ChatGPT, Grok, or add a key

```
$ krowk connect
```

3 — Open the agent in this repository

```
$ krowk
```

[krowkcom/krowk](https://github.com/krowkcom/krowk)

**Pricing**

## **The agent is free. Publishing bills like an API. **

krowk is MIT and runs on your own provider subscription or API key. The plans below are for publishing: links, uploads and storage. Teammates are free on every plan.

### **Free**

For trying it inside a real run.

**$0**

- Links live for 24 hours
- 100 uploads a day
- 100 MB an artifact
- Expired links recoverable for 30 days
- No signup, no card

### **Pro**

For teams whose agents run daily.

**$15**/mo

- Links that never expire
- 10,000 uploads a month
- 2 GB an artifact
- 100 GB of storage included
- Unlimited teammates, a key per agent

[Start Pro](https://app.krowk.com/registration/new)

### **Business**

For when the output can no longer be public.

**$49**/mo

- Links that never expire
- 30,000 uploads a month
- 2 GB an artifact
- 100 GB of storage included
- Unlimited teammates, a key per agent
- Private and shared artifacts
- Per-fetch access logs

[Start Business](https://app.krowk.com/checkout?products=4503afb3-c67f-4ea1-a9ba-ef42c4b2709c)

**Paid plans, past the included amounts **

Included uploads: 10,000 a month on Pro, 30,000 on Business. Storage included is 100 GB on both.

<dl>

<dt>Uploads past the included monthly amount</dt>
<dd>**$1 per 1,000**</dd>

<dt>Storage past 100 GB</dt>
<dd>**$0.05 per GB-month**</dd></dl>

Every key has a per-minute ceiling, always on. One key per agent is one kill switch per agent. A hard spend cap is coming, but not yet.

**FAQ**

## **Common questions **

<details>

<summary>

What is Krowk?

</summary>

A coding agent harness for your terminal. It runs Claude, GPT, Grok or any OpenAI-compatible model on your subscription or an API key, follows Claude Code's permissions, skills and hooks, lets you switch models mid-conversation, and publishes its work as links that unfurl in GitHub, Slack, Linear and Basecamp.

</details>

<details>

<summary>

Can I use my Claude or ChatGPT subscription?

</summary>

Yes. krowk connect signs in to the Claude and ChatGPT subscriptions (through Claude Code and Codex, which krowk drives) and to SuperGrok, or takes an API key for Anthropic, OpenAI, xAI, OpenRouter or any OpenAI-compatible server.

</details>

<details>

<summary>

What happens when I switch models mid-conversation?

</summary>

The conversation goes with you: what you asked, what was answered, every tool call and result. The one thing that does not cross is a provider's private reasoning state, which the next model reads as plain text instead.

</details>

<details>

<summary>

Will my Claude Code setup work?

</summary>

It is read as it is: permission modes and allow/ask/deny rules, settings files, AGENTS.md and CLAUDE.md, skills, hooks and .claude/agents. krowk adds one mode of its own, unhinged, which asks nothing and holds no deny rule or fence; only a hook that blocks still blocks.

</details>

<details>

<summary>

Which platforms does it run on?

</summary>

macOS, Linux and Windows today, as one static binary. iOS, Android, desktop and mobile clients are on the way; they are not available yet.

</details>

<details>

<summary>

Is it free?

</summary>

The agent is MIT licensed and costs nothing; you pay your model provider as you already do. Krowk's plans are for publishing: permanent links, upload volume and private artifacts.

</details>

<details>

<summary>

Can my other agents publish too?

</summary>

Yes. Anything that can run a CLI publishes with krowk push: Claude Code, Cursor, Codex, Devin, OpenHands or a bash step in CI. CI and containers get a lean build that only publishes.

</details>

<details>

<summary>

Why not just use S3?

</summary>

Signed S3 URLs expire and unfurl as nothing. A Krowk link is stable and renders as a card in GitHub, Slack, Basecamp, Linear and Notion; the API is retry-safe. Storage is the cheap part.

</details>

<details>

<summary>

What happens to a free link when it expires?

</summary>

Free links expire 24 hours after upload, and the URL returns 410 Gone with the filename and upload time. The artifact sits in cold storage for 30 days; claim a workspace in that window and the link comes back. After that it is deleted.

</details>

<details>

<summary>

What does publishing cost past the included amounts?

</summary>

Two meters. $1 per 1,000 uploads past the included monthly amount (10,000 on Pro, 30,000 on Business), and $0.05 per GB-month of storage past 100 GB. Retries with an Idempotency-Key are never double-counted. No bandwidth or egress charge.

</details>

<details>

<summary>

What does Business add?

</summary>

Privacy and audit. Private artifacts stay workspace-only; shared ones open to whoever holds the link. Access logs record every fetch. Same meters as Pro; 30,000 uploads included instead of 10,000.

</details>

<details>

<summary>

What if an agent goes runaway?

</summary>

In the agent, cap a session with --max-usd or --max-tokens; the engine checks it before every model call. For publishing, every key and keyless address has a per-minute upload ceiling, on by default; runaway loops get 429 with Retry-After inside a minute. A hard spend cap on the publishing bill is coming, but not yet.

</details>

<details>

<summary>

What happens to my links if I cancel?

</summary>

Links stay live for 30 days after cancelling, then sit in cold storage for 30 more; re-subscribe in that window and every link comes back at the same URL. Export every artifact with its metadata as one archive at any time. After 60 days links return a 410 tombstone.

</details>

<details>

<summary>

Is it open source?

</summary>

Yes. krowk is MIT, at github.com/krowkcom/krowk. The upload API is documented at krowk.com/docs, so you can publish with curl if you would rather install nothing.

</details>

<details>

<summary>

Can I self-host?

</summary>

The agent runs entirely on your machine. Publishing cannot be self-hosted yet: the unfurl side depends on our link resolver, and we would rather ship one that works than a self-host guide that half works.

</details>

<details>

<summary>

How do I get a screenshot out of Claude Code into a pull request?

</summary>

Add the rule under Publishing to CLAUDE.md or AGENTS.md. Visual changes are captured, published with the krowk skill and pasted in the PR. Without the skill: krowk help --json, then krowk push. The first push needs no account; with an API key the card carries repo, commit, branch and agent.

</details>

## **Take your agent everywhere you work. **

One install. Your models, your permissions, your sessions.

Publish from another agent

terminal

```
$ curl -fsSL https://krowk.com/install | bash
$ krowk connect
$ krowk
```