Tool	What it does
Teamwork Graph CLI (`twg`)	Gives the Teamwork Graph context arm access to your connected work data.
Benchmark CLI (`benchmark`)	Runs the comparison and stores every result on your own computer.
Benchmark app	The local web app `benchmark run browser` opens. You do the whole benchmark in it: finish setup, choose prompts, start and watch a run, review answers, and read the report.
AI agent	Codex or Claude Code. The benchmark drives your agent to produce both answers.

Before you begin

You need:

An Atlassian Cloud site you can sign in to. Data Center sites are not supported.
At least one supported AI agent installed on your computer: Codex or Claude Code. The benchmark app uses whichever supported agent it finds, and drives it to produce both answers.
Sources of work connected to your Atlassian site. The benchmark requires at least one source in each of Work, Knowledge, and Code, for example Jira, Confluence, and GitHub or Bitbucket. A People source is strongly recommended. Step 5 walks you through connecting them, and more relevant sources usually give Teamwork Graph more useful context, which generally produces a more meaningful comparison.

Nothing else is a prerequisite. You install the Teamwork Graph CLI and the Benchmark CLI in steps 1 and 2, and the benchmark app checks the rest for you, including setting up the Rovo MCP server and signing you in where it is missing.

Step 1: Install or check the Teamwork Graph CLI

If the Teamwork Graph CLI might already be installed, skip ahead and run twg doctor.

On macOS and Linux:

1
2
curl -fsSL --retry 2 https://teamwork-graph.atlassian.com/cli/install | bash

On Windows, run this in PowerShell:

1
2
curl.exe -fsSL https://teamwork-graph.atlassian.com/cli/install.ps1 -o "$env:TEMP\twg-install.ps1"; powershell.exe -NoProfile -ExecutionPolicy Bypass -File "$env:TEMP\twg-install.ps1"

Open a new terminal window and confirm the installation:

1
2
twg doctor

Success: every required check passes. For alternate installation methods and fixes, see Install and set up the Teamwork Graph CLI.

Step 2: Install or update the Benchmark CLI

On macOS and Linux:

1
2
curl -fsSL https://teamwork-graph.atlassian.com/cli/benchmark/install | bash

To install into a different directory:

1
2
curl -fsSL https://teamwork-graph.atlassian.com/cli/benchmark/install | BENCHMARK_INSTALL_PREFIX="$HOME/bin" bash

On Windows, run this in PowerShell:

1
2
curl.exe -fsSL https://teamwork-graph.atlassian.com/cli/benchmark/install.ps1 -o "$env:TEMP\benchmark-install.ps1"; powershell.exe -NoProfile -ExecutionPolicy Bypass -File "$env:TEMP\benchmark-install.ps1"

The Windows installer places benchmark.exe in %LOCALAPPDATA%\Programs\benchmark\bin and adds that directory to your user PATH.

Open a new terminal window and confirm the installation:

1
2
benchmark version --verbose

Success: the command prints the installed version with no setup error.

If the Benchmark CLI was already installed, update it before continuing:

1
2
benchmark update

If benchmark update fails, or still targets an old install channel, rerun the installer for your platform above. Reinstalling is always safe.

Step 3: Open the benchmark app

1
2
benchmark run browser

This command starts a local server and opens the benchmark app in your browser. The app is the benchmark itself, not a dashboard that reports on one: from here you finish setup, choose your prompts, start the run and watch it, review the answers, and read the report, all in the same tab. Leave the terminal open, because closing it ends the session.

Success: your browser shows Set up your environment.

Step 4: Clear the prerequisite checks

The benchmark app checks your environment and lists what it found. Work down the list until nothing is left to fix.

Row	What it means
Your agent, for example Codex Sandbox or Claude Code	The agent is installed and signed in. The benchmark uses a separate sign-in for Codex so runs stay reproducible.
Teamwork Graph CLI	The `twg` binary is installed and signed in.
Rovo MCP	The Atlassian MCP server the Siloed context arm uses. Without it there is no baseline to compare against.
Other MCP servers	Informational. It lists other MCP servers your agent exposes, so you can see what else was in scope.

Each row shows a status badge, such as Checking, Ready, Not installed, or Setup failed. When a row can be fixed automatically it offers a button:

Select Install or Sign in on the row you want to fix.
If the row asks you to authenticate, select Open sign-in and complete the sign-in in the new tab.
If a fix fails, select Try again, or Run checks again to recheck everything.

Success: Continue to sources becomes available. Select it.

The benchmark app changes only the configuration the benchmark itself owns. For Codex it uses its own sandbox home and leaves your desktop Codex configuration untouched. Claude Code keeps a single credential store, so a Claude sign-in here is the same one you would get from /mcp.

Step 5: Connect your data sources

The Choose where context comes from step groups every source the benchmark can read.

Group	Requirement
Work, for example Jira	Connect at least 1
Knowledge, for example Confluence or Google Drive	Connect at least 1
Code, for example Bitbucket or GitHub	Connect at least 1
People	Strongly recommended
Communications, for example Slack or Google Calendar	Optional
Other, for example Assets or Datadog	Optional

Select Connect on a source you want to include, then finish the sign-in it opens.
Select Check connection to confirm a source the benchmark cannot yet read.
Newly enabled connectors can take a few minutes to install and sync, so recheck if a source looks empty.

Success: each of Work, Knowledge, and Code shows at least one connected source, and Continue to prompts becomes available.

Connecting more relevant sources gives Teamwork Graph a denser graph to work with, which usually produces a more meaningful comparison. Your connections stay in Teamwork Graph.

Next step

Run a benchmark. Keep the benchmark app open, because the next steps continue in the same browser tab.