Skip to main content
Glama

JK

Give ChatGPT real hands on your computer: it reads your code, edits files, runs tests, and shows you the evidence, while JK remembers the work and keeps it safe.

English | 한국어

JK is a free local MCP app designed and built by Anjingyeong, originally started from ezBuilder/chatgpt2codex. Install it, pick a project folder, and connect it to ChatGPT. From then on the chat you already use can work on your real project instead of only talking about it.

JK is an independent project and is not affiliated with or endorsed by OpenAI. OpenAI, ChatGPT, GPT, and Codex are marks or products of OpenAI.

Why JK exists

ChatGPT is good at reasoning about code, but on its own it can't see your repository, can't run your tests, and forgets where it left off. The usual workaround is copying files into the chat and pasting answers back. That's slow, error-prone, and you never know whether the suggested fix actually works.

JK was built to close that gap without adding another AI subscription:

  • ChatGPT keeps doing the thinking. JK is not a model. It is the local execution layer, so you keep using the ChatGPT plan you already have.

  • Work continues across turns. JK stores the goal, touched files, decisions, and verification results, so "continue what you were doing" means exactly that.

  • Results come with proof. Changes are verified by running your own commands, and the output is shown back instead of a "should work now".

  • Your machine stays yours. Access is limited to the folder you choose, risky actions wait for your approval, and secrets are redacted.

Related MCP server: localterminal-lite

Why use it

Without JK

With JK

Copy files into the chat, paste answers back

ChatGPT reads and edits the files directly

"This should fix it"

The fix is tested and the passing output is shown

A new chat starts from zero

The previous task, files, and decisions are resumed

One long change on your working copy

Work in an isolated task workspace, publish only verified changes

No idea what an agent is doing

Live Control Center with a dependency graph of parallel work

Download

Get the app from the Releases page. Node.js, cloudflared, and ripgrep are bundled, so no terminal is needed.

OS

File

Windows

JK-...-windows-setup.exe

macOS

jk-....pkg

The installers are not code-signed yet. If Windows SmartScreen appears, choose More info → Run anyway only when the file came from the official Releases page.

On macOS, if the package is blocked ("cannot be opened because it is from an unidentified developer"), right-click (Control-click) the .pkg in Finder → Open → Open, or allow it under System Settings → Privacy & Security → Open Anyway.

Prefer the terminal?

With Node.js 22+ installed, one command sets everything up and prints the connector URL, Control Center link, and health check (Ctrl+click to open):

npx -y jk-mcp setup

Run the same command next time; it reuses your folder and connection code. jk start also works after npm install -g jk-mcp.

Quick start

  1. Install and open JK. The tray icon appears near the clock (menu bar on macOS).

  2. Open Settings... and choose the project folder ChatGPT may work in.

  3. Click Auto-generate Token, then Copy Owner Token. Treat it like a password.

  4. Enable ChatGPT web connector and click Start MCP.

  5. Click Copy Connector URL. It ends in /mcp.

  6. In ChatGPT, open Apps & Connectors / Connectors, create a new connector, and paste the URL.

  7. When ChatGPT shows the JK login window, paste the Owner Token.

  8. Try: @jk Read this project's README and summarize its current state.

Connector URL: temporary vs. your own domain

No domain (default)

Your own domain (optional)

URL

https://<random>.trycloudflare.com/mcp

https://mcp.example.com/mcp

Setup

Nothing. JK creates a Cloudflare Quick Tunnel automatically.

Run a Cloudflare Named Tunnel or HTTPS reverse proxy to http://127.0.0.1:7979, then enter the hostname in Owned fixed domain.

After JK restarts

The URL changes. Update the connector URL in ChatGPT and log in again.

The URL stays the same.

Best for

Trying JK, short sessions

Daily use

A domain is nice to have, not required. See the installation guide for the full setup.

What you can ask

@jk Inspect this project and explain how it works. Do not edit anything.
@jk Find the cause of this bug, fix it, and run the relevant tests.
@jk Continue the previous task and finish it.
@jk Implement this in a separate task workspace, verify it, then apply it.
@jk Run E2E and show the passing evidence.
@jk Review the current diff and commit it if verification passes.

You describe the goal in plain language; you don't need to know any tool names.

Features

Work that carries on

  • Persistent work sessions: the current goal, touched files, pending items, decisions, checkpoints, and verification results are saved per project, so follow-ups resume the real task.

  • Checkpoints: restore an earlier state of the files JK changed.

Changes you can trust

  • Guarded edits: edits to existing files can carry a SHA-256 precondition. If the file changed after it was read, the edit is rejected instead of overwriting newer work.

  • Verification evidence: tests, builds, and E2E runs are executed locally and their output is returned as proof.

  • Task workspaces: implement and verify in a private copy of the project, then publish only verified, reviewed changes back. If the original moved on in the meantime, publishing is refused rather than merged blindly. See Task Workspaces.

Bigger jobs, organized

  • JK orchestration: goal_intake and goal_loop run longer coding loops through Explorer, Oracle, Implementer, Reviewer, Verifier, and Recovery roles. A failed verification re-plans instead of repeating the same attempt.

  • MASS ULW parallel work: split a job into dependency-aware lanes that implement, verify, repair, and review in parallel inside one ChatGPT conversation. See MASS ULW workflow.

  • Optional OMO delegation: hand selected passes to a local OMO / Oh My OpenAgent install while JK keeps state, safety, and verification authoritative.

See what's happening

  • Control Center: a local dashboard for the current task, approvals, execution hosts, and activity.

  • Live dependency graph: parallel work is drawn as a DAG with waves, colour-coded lane states, animated flow into running lanes, and hover highlighting of each lane's dependencies.

Everything a coding session needs

Project selection, code search, narrow file reads, guarded patches, allowlisted commands and shell jobs, Git status/diff/commit/explicit push, dev servers, browser E2E with screenshots, image intake, external MCP routing, and runtime health checks.

Full list of all 70 tools, grouped: docs/FEATURES.md.

Safety model

JK is intended for trusted development environments.

  • Access is scoped to the selected project folder.

  • Sensitive shell, network, Git publication, and destructive operations require your approval.

  • Secret-looking values are redacted from tool output.

  • Project leases and task ownership prevent conflicting concurrent edits.

  • The MCP endpoint rejects requests without Owner Token authorization. Rotating the token signs out existing sessions but keeps the connector registered.

Never share your Owner Token, tunnel credentials, or domain credentials in screenshots, logs, issues, or chats. If one leaks, rotate it in Settings.

For developers: build from source

Requirements: Node.js 22+, npm, and PowerShell on Windows.

npm ci
npm run typecheck
npm test
npm run build

Run from source:

npm run chatgpt:windows   # Windows
npm run chatgpt           # macOS
npm run chatgpt:linux     # Linux

Packaging and E2E:

npm run windows:package
npm run windows:e2e
src/
  auth/           authentication / owner-token support
  code/           search, read, patch operations
  control/        desktop/control safety paths
  control-center/ local operations UI
  e2e/            local browser E2E and evidence
  exec/           command, shell, OMO runners
  executors/      remote/local execution routing
  orchestration/  MASS ULW and longer-running task logic
  policy/         approvals, paths, secrets, shell-job policy
  roles/          JK orchestration roles
  server/         MCP tools and Actions bridge
  state/          persistent project/work-session state
  workspace/      project registry and lease handling

windows/          Windows launcher / tray / installer
macos/            macOS launcher/status-bar path
linux/            Linux launch/install path
scripts/          build, packaging, verification, release helpers
docs/             usage, architecture notes, engineering docs
assets/           JK public assets

Author

JK is designed, built, and maintained by Anjingyeong. Found a problem? Use Report issue in the app sidebar or tray menu, or 문제 신고 · 피드백 in the Control Center; it opens a prefilled issue with your version and OS. Security problems: see SECURITY.md. Pull requests are welcome; see CONTRIBUTING.md.

JK started from ezBuilder/chatgpt2codex, used with the original author's permission. Thanks to ezBuilder for the original work. See ACKNOWLEDGEMENTS.md and Attribution and compliance notes.

License

MIT © 2026 Anjingyeong. If you fork or redistribute JK, please keep the credit to the original repository (NOTICE).

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      C
      maintenance
      Bridges ChatGPT with local computer for controlled file and project management, featuring session-based collaboration and diff tracking.
      4
      Apache 2.0
    • A
      license
      Not graded
      quality
      B
      maintenance
      Remote MCP coding bridge that gives ChatGPT/Codex secure local workspace access, including file retrieval, semantic code intelligence, Git, diagnostics, and guarded shell execution.
      55 npm
      MIT
    • A
      license
      Not graded
      quality
      C
      maintenance
      A bridge that enables the ChatGPT web interface to use local file and shell tools through the official MCP protocol, turning the workspace into an agentic coding environment.
      17
      MIT