Get started

For Claude Code, Copilot CLI and Codex CLI

Ask Natterjack.

Where the business meets the codebase.

One dashboard on top of your team’s repos. Engineers start apps and agent runs from the browser. Product, ops and support ask the code directly, read-only, and keep what the business knows right next to it.

Get started View on GitHub

Free to install · Runs on your machine · v0.17.2

Two audiences, one room

The questions and the code, in the same place.

Engineers get a home base for running things. Everyone else gets a way to ask how things work, and an answer from the code itself.

For engineers

Stop juggling terminals.

  • Start the stack for any worktree with one click, on the right ports, with logs.
  • Run Claude Code, Copilot or Codex from the browser and answer its questions with buttons.
  • Explore the code on a free open model, on your laptop or the team’s GPU.
  • Let a run watch your PRs instead of babysitting them.
  • New laptop? The Machine page lists what’s missing and installs it, and Connections signs you in to the team’s tools.
For product, ops and support

Ask the code, not an engineer.

  • Ask how something works and get an answer from the real code, read-only.
  • Write down what the business knows in Knowledge, right from the page. No GitHub seat needed.
  • Bring in Notion, Confluence or Drive through connectors. Answers cite the notes and pages they used.
  • A simpler dashboard with the build-and-run pages hidden, and answers in the style your role needs.
  • No git access? Download a read-only snapshot and ask away.
  • Ask for a report or an export and download the file from the run.

Fewer “quick question” interruptions for engineers. Faster, first-hand answers for everyone else.

What’s inside

A home base for the whole team.

Natterjack installs a dashboard into your workspace. Here’s what’s on it.

Runs

Start an agent from the browser. Answer it with a click.

Run Claude Code, GitHub Copilot CLI or OpenAI Codex CLI, picked next to Model and Effort. When the agent needs a decision, it asks on the page and you answer with buttons. Every run keeps its transcript, its subagents and a Changes tab with a diff of what it touched.

  • Type while it works and your message is queued for the next turn. Turn on Auto-send and you can walk away.
  • Type / in any chat box to pick a skill or command.
  • “Continue in terminal” hands the session to the agent’s own resume command when you want the keyboard back.
  • Files a run makes, like a PDF, a spreadsheet or an export, download from the run page.
  • Switch models mid-run, and see which model each subagent ran on.
  • Plan mode (read-only) is a switch on each run.

Models

Open models for the legwork. Claude for the hard part.

Run code search and exploring on an open model, at no Claude cost. The run is still Claude Code with the same tools, skills, CLAUDE.md, plan mode and run page. Only the model behind it changes.

  • On this computer: install Ollama and download a recommended model from the page. It checks your GPU, memory and disk and tells you what fits.
  • For the team: list a model you host on Ollama, vLLM or LiteLLM, and everyone gets it, hosted dashboards included. Those turns don’t leave your network.
  • Smart routing (experimental): Claude plans, edits and answers, and hands its broad code searching to an open model.
  • A model that can’t call tools is listed but can’t be picked, and a missing model never quietly falls back to Claude.

Knowledge

What the business knows, where Claude can read it.

Company, Customers, Finance: each area has an owner, notes and a review date. Notes live in your own S3, Google Cloud Storage or Azure bucket, and anyone can edit them from the page. No GitHub seat needed.

  • Reads like an Obsidian vault: [[wikilinks]], backlinks, tags and folders. An existing vault works as it is.
  • Notes past their review date are flagged until someone marks them reviewed.
  • Docs in the repos sit alongside links to Notion, Confluence or Drive, and each card says whether Claude can reach it.
  • Separate buckets or keys per area, so only Finance’s key holders read Finance.

Issues

Your tracker, next to the code it’s about.

The team’s board from Linear, Jira or GitHub Issues, with two buttons on every ticket. Other trackers can be added through adapters.

  • Explain answers read-only, against the real code.
  • Implement takes a ticket end to end with the ready-made multirepo plugin, or your team’s own /implement skill.
  • Tracker keys stay on the server and never reach the browser.

Apps & Workspaces

Start the whole stack. In order. On the right ports.

Start and stop each app, or a whole stack, for each git worktree, with logs. No more remembering commands and ports, or keeping six terminals open.

  • Stacks start in the order you set: database, then API, then web.
  • Apps keep running when the dashboard restarts.
  • Runs and terminals can start apps too, so a dev server outlives the run that started it.
  • On Windows, no console windows pop up.

Machine

Onboarding in hours, not days.

Natterjack knows what this computer needs for your stack and checks it live: Node, Go, PHP, Python, .NET, Java, Docker, databases and more. Missing something? Click Install.

  • Checks come from your repos’ own version files.
  • Apps that need a missing tool say so on their card.
  • Build tools can be shown to developers only.
  • Copilot and Codex show up as agents, with Sign in and Install buttons.

Connections

Every tool your agents can reach, and what’s broken.

Every MCP server the agents can use, from claude.ai connectors and plugins to your own, with its live state and the fix one click away: sign in, approve, add, or let dashboard runs use it.

  • The team’s must-have servers show first, with a count in the sidebar for anyone who isn’t connected.
  • A tab for each installed agent. Copy from Claude gives Copilot or Codex the servers Claude already has.
  • Sign in starts from the page, hosted dashboards included.
  • Header and environment values never reach the browser.

People

Give someone a hosted dashboard in a click.

Your company directory next to who has a hosted dashboard. Set people up, start, stop, update or remove them from your own dashboard, on Kubernetes or Azure App Service.

  • Lists everyone in your Google Workspace or Entra ID domain, or one team. Nobody gets a dashboard until an engineer clicks Set up.
  • A profile per person sets their Knowledge areas, run limit and size. Assign their role and their access, full stack or read-only, and they can’t change them.
  • A Roles tab edits the team’s roles: who someone is, and the style Claude answers them in, like plain language for support.
  • Flags dashboards that need an update and people who have left.
  • Every change shows the platform’s own preview before it’s applied. The roster is a file in your repo, so git history is the audit log.

Ask

A question box that starts a run against the real code, read-only by default. While you type, a free local search shows the files the agent will start from. Tick a Knowledge bucket or Confluence and the answer cites the notes and pages it used.

PR watches

A run can watch its pull requests and wake up for a new review or comment, failing checks, a change in review decision, or a branch that’s behind or conflicting.

Links and Infrastructure

Link tiles anyone can add to, for the team or just themselves. Environments, URLs, accounts and IPs with click-to-copy values.

Roles and access

A role is who someone is and how Claude answers them, so support can get plain language instead of file names. Access is separate: read-only hides the build-and-run pages and Implement. On hosted dashboards, an engineer assigns both.

Snapshots

People without git access download read-only copies of the code, published by CI to Confluence, a web server, S3, Cloudflare R2, Google Cloud Storage or Azure Blob, and can still Ask, Explore and read the docs.

And the rest

Search across everything without spending tokens, skills as cards (your claude.ai organization’s skills and your plugins’ skills included), plan usage limits, memory, a file explorer, and a restart button for the dashboard itself.

How setup works

Discovery, not forms.

Run one skill in the folder that holds your repos. Claude works out your setup, asks about the gaps, and hands you a dashboard that already knows your stack.

  1. 1

    Discover

    Claude first checks the folder: ideally a git repo of its own with every repo your product is made of inside it, and it helps you set one up if not. Then it scans your repos: stacks, ports, run commands, databases, CI, deploy targets, ticket keys in your branches, and your design system.

  2. 2

    Interview

    It asks only what it couldn’t work out, in a round or two of questions: which repos, which tracker, which apps run together, and where the business keeps its knowledge.

  3. 3

    Brand

    It writes the config and brands the dashboard with your team’s own colors and logo, and checks the contrast.

  4. 4

    Start

    The dashboard starts on localhost. Teammates pull, start it, and the Machine and Repos pages walk them through the rest.

Your copy, your changes.

The dashboard is copied into your workspace and you own that copy. Change anything. When a new version comes out, /natterjack:upgrade merges it around your changes.

Run it your way

On a laptop, or on your own infrastructure.

Engineers run Natterjack on their own machines. People without a dev machine can use a hosted copy behind your company’s single sign-on. The same workspace serves both.

Local

For engineers

  • Runs on each person’s computer at localhost:3333.
  • Uses the Claude Code, Copilot or Codex you’re already signed in to.
  • Everything in the feature tour above, including Apps, Workspaces and Machine.
  • The People page, where engineers manage the team’s hosted dashboards, is here too.
Install it

Hosted

For product, ops and support

  • Each person gets their own container behind your single sign-on: Google Workspace, Entra ID, Okta and others.
  • One shared address. A small router starts each person’s container when they arrive, and stops it when they’re idle.
  • Engineers set people up from the People page, with a profile, a role and their access for each.
  • Everyone signs in with their own Claude seat, so their claude.ai connectors just work.
  • Sign in to GitHub from the page to push and open pull requests as yourself. Commits are signed and show as Verified, and you can keep the sign-in to people with full-stack access.
  • Containers boot from a published snapshot, with no git and no code-host account, then pull each repo’s latest default branch. Knowledge keys come from the deployment’s secrets.
  • A model on your team’s own GPU works here too, so code search and exploring stay inside your network.
  • Runs on Kubernetes (GKE, EKS, AKS), Azure App Service, or a VM with Docker.
  • /natterjack:host walks you through the setup, step by step, and /natterjack:doctor checks it afterwards.
Read the hosting guide

Trust and safety

It can start processes. So it’s careful about who can ask.

The dashboard runs agents and your apps as you, so it only listens to you.

Local and locked down

The dashboard binds to 127.0.0.1 only, with checks against DNS rebinding and other origins, and a per-install token on every POST. No telemetry, and nothing is ever sent to Natterjack’s author.

Read-only by default

Ask runs in plan mode unless you switch it off, whichever agent runs it. People on the reader profile never get Implement. Runs can only use the MCP servers you’ve allowed.

Keys stay put

Tracker keys and MCP server secrets never reach the browser, and bucket keys never go in config. Install buttons and app commands come only from config files, never from the page.

One container per person

Hosted mode isn’t a shared multi-user server. Each person gets their own container, disk and Claude sign-in.

No cloud credentials of its own

The People page uses the kubectl, az, gcloud and git an engineer already signs in with. Every change is previewed first, and deleting someone’s disk means typing their email.

Commits that are really yours

Hosted GitHub sign-in uses the GitHub CLI’s device flow, with no OAuth app to set up. Each person pushes as themselves, within their own GitHub permissions. Limit the sign-in to full-stack access and read-only people can read the code but not push. Signing out removes the key.

Agents and models

Bring the agent your team already uses. Or your own model.

Runs and Ask work with Claude Code, GitHub Copilot CLI or OpenAI Codex CLI, whichever are installed, and with open models on your machine or your team’s GPU. Pick one per run, next to Model and Effort, and set the team’s default. Much of the dashboard, like Apps, Machine, Knowledge, Links and the issue board, is plain tooling that doesn’t need a model at all.

Available now

Claude Code

Sign in with a Claude subscription, or use an Anthropic API key, Amazon Bedrock or Google Vertex AI. Plan-usage meters, the per-run cost cap and memory are Claude-only.

Available now

GitHub Copilot CLI

Runs and Ask with Copilot and its models. It reads your workspace’s CLAUDE.md and skills, and installs the plugin from the same marketplace.

Available now

OpenAI Codex CLI

Runs and Ask with Codex and its models. Plan mode uses Codex’s read-only sandbox, and your team’s CLAUDE.md and skills come along.

Available now

Open models

Qwen, Gemma, Nemotron and more, on your own computer through Ollama or on a GPU your team hosts. Same tools, skills and plan mode, and they don’t count against your Claude usage.

Using an agent or model we don’t support yet? Tell us what your team would use.

Get started

Three commands to a dashboard.

Install the plugin in Claude Code, Copilot CLI or Codex CLI, then run create in the folder that holds your team’s repos.

You’ll need

The dashboard’s Machine page helps with everything else your stack needs.

View the source on GitHub
1 · In Claude Code
/plugin marketplace add lhoezee/skills
/plugin install natterjack@lhoezee-skills
Using GitHub Copilot CLI or OpenAI Codex CLI?
In a shell, for Copilot CLI
copilot plugin marketplace add lhoezee/skills
copilot plugin install natterjack@lhoezee-skills
In a shell, for Codex CLI
codex plugin marketplace add lhoezee/skills
codex plugin add natterjack@lhoezee-skills

Then ask it to “use the natterjack create skill”. Copilot asks before running the plugin’s scripts: approve them.

2 · In the folder with your repos
/natterjack:create

Afterwards, /natterjack:doctor checks everything, /natterjack:upgrade pulls in new versions, and /natterjack:host sets up hosted dashboards. create also offers the multirepo plugin from the same marketplace: ready-made skills to implement a ticket, commit, open pull requests and manage worktrees. Used it back when it was called agentic-os? Here’s how to switch.