Building with agents 0/15

The AI tools San Francisco engineers use

What Claude Code, Codex, ChatGPT, Grok Bot, Hermes, Nuphos and Aside each do, one request for each with what comes back, and when another tool suits the job better.

1,909 words · 9 min read

What an agent is

This lesson covers the AI tools that WhiteBox’s founder, a founding engineer in San Francisco, chose from day-to-day work: Claude Code, Codex, ChatGPT, Grok Bot, Hermes, Nuphos and Aside. Product details come from each tool’s own site or documentation as of November 2026, linked under Primary references. These products change quickly, so check the linked page before you rely on a detail.

Most of them are agents. An agent is a program that gives a language model tools, such as reading a file, running a command or clicking a button on a web page, and then carries out the actions the model chooses, one after another, until the task is done or it needs you. The tools below differ mainly in where that loop runs and what it can reach: a project folder on your laptop, your logged-in accounts in a browser, or a computer in the cloud that keeps working after you close the laptop.

Coding agents in the terminal and editor

A coding agent works in a project folder. It reads the code, edits files and runs commands such as your tests. The terminal is the text window where you type commands; both tools in this group start there with one word, and both also run inside code editors and desktop apps.

Claude Code

Claude Code is Anthropic’s coding agent. It reads your codebase, edits files, runs commands and works with git, and it runs in the terminal, in VS Code and JetBrains editors, in a desktop app and in the browser at claude.ai/code. Most of these need a Claude subscription or an Anthropic Console account.

Its documentation lists the everyday jobs: writing tests for untested code, fixing lint errors across a project, resolving merge conflicts, writing commit messages and opening pull requests. A CLAUDE.md file in the project root holds instructions it reads at the start of every session, such as how to run the tests or which libraries the team prefers.

claude on its own opens a conversation in the current folder. claude -p sends one request, prints the answer and exits, and it reads whatever you pipe into it. This line hands it the files your branch changed and asks for a security review:

Bash
cd your-project
git diff main --name-only | claude -p "review these changed files for security issues"
One request with piped input, from the Claude Code overview

git diff main --name-only prints the names of the files that differ from main. Claude Code receives that list, reads those files, prints its review in the terminal and exits. Because it exits, the same line also works in a script or a CI job, the checks a repository runs on every pull request.

If a teammate needs to watch or steer the same session, run Claude Code inside Nuphos, covered below, which shares one session with the team.

Codex

Codex is OpenAI’s coding agent. It runs in the ChatGPT desktop, web and mobile apps, in the terminal as the Codex CLI, as an editor extension and as Codex Cloud, and it can sign in with your ChatGPT account.

In the terminal it inspects files, makes edits and runs the tools already installed on your machine. /permissions sets when it may edit files or run commands without asking, codex exec runs it from scripts and pipelines, and codex cloud hands work to Codex Cloud and brings the result back for you to apply locally. Run a review before each commit:

Bash
curl -fsSL https://chatgpt.com/codex/install.sh | sh
cd your-project
codex
/review
Install, start in a project and review uncommitted work

The first line installs the CLI on macOS or Linux. codex opens a session in the project, and /review reviews your uncommitted changes, a commit or the difference from a base branch. It reports what it found in the session and leaves your working tree unchanged.

Claude Code and Codex do the same job, so the practical choice is the subscription you already pay for: Codex signs in with ChatGPT, Claude Code with a Claude subscription or an Anthropic Console account. For a quick question with no project folder, ChatGPT’s chat is faster to open.

Chat models

A chat app sends your message and any files you attach to a language model and shows its reply. It sees only what you attach or the apps you connect; it has no access to your project folder.

ChatGPT

ChatGPT is OpenAI’s chat app, on the web, on phones and as a desktop app for macOS and Windows. OpenAI’s overview splits it into Chat for questions and learning, Work for documents, slides, spreadsheets and connected apps such as Gmail and Slack, and Codex for code.

For an engineer, it is the quickest place for a question that fits in one message: explaining an error, comparing two approaches, reading an attached spreadsheet or translating a file from one language to another. This request is one of the examples on OpenAI’s overview page:

Text
Translate the attached chartUtils.py file into JavaScript.
A chat request with one Python file attached

The reply contains a JavaScript version of the file. Nothing runs against your project: you copy the code across, run your own tests and fix what they catch.

When a change spans several files or the tests have to run, give it to a coding agent instead. Codex is part of the same ChatGPT account.

Agents that run on their own

These agents keep working when you are not watching: on a schedule, when a webhook or alert arrives, or from a message you send in Slack or from your phone. A webhook is an HTTP request another service sends when something happens, such as a monitor reporting that a site is down. Because nobody approves each step as it happens, limit what such an agent can reach before you hand it a recurring job.

Grok Bot

Grok Bot is xAI’s agent for work across apps. Each Bot works on a computer in the cloud with a browser, a file system and a terminal, shared with your other Bots, and it keeps your preferences, files and logins between tasks. You talk to it in its desktop app for macOS, Windows and Linux, on mobile, in Slack and on X. It signs in with a Cursor account and needs a paid X, Grok or Cursor plan.

Its documented jobs are office work: sorting email, keeping a calendar in order, drafting Slack replies and Notion updates, and building a new deck in the format of past ones. A task you walk a Bot through can be saved as a skill, and a task you repeat can become a routine that runs on a schedule. Grok Bot’s getting-started page suggests this first request:

Text
Sort my unread email into urgent, needs a reply, and FYI. Draft replies to the urgent ones and don’t send anything.
A first request to an inbox Bot

The documented result is your urgent threads first, each with a drafted reply waiting for your approval. The request says not to send anything, so you send every reply yourself.

Its examples cover email, calendars and documents rather than a code repository. For a change to code on your machine, a coding agent working in your project folder is the direct route.

Hermes

Hermes Agent is an open-source agent from Nous Research, released under the MIT license. You install it on your own computer, on macOS, Linux or Windows, or run a hosted copy from Nous Portal. It works with a model provider you choose, such as Nous Portal, OpenRouter or OpenAI, and you can talk to it from the terminal or from Telegram, Discord, Slack, WhatsApp, Signal or email, with one memory across all of them.

It runs scheduled reports and briefings, starts subagents with their own conversations for side tasks, and runs commands in one of five sandboxes: local, Docker, SSH, Singularity or Modal. A sandbox is a separate environment that limits what the agent’s commands can touch. Its quickstart begins like this:

Bash
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup --portal
hermes --tui
Install Hermes, connect a model and open the terminal interface

hermes setup --portal connects a model, plus web search, image generation, speech and a browser, with one Nous Portal login. hermes --tui opens the terminal interface. One of the first requests the quickstart suggests is “Summarize this repo in 5 bullets and tell me what the main entrypoint is.” Run from a project folder, it asks for five points about the project and the file where the program starts; the wording depends on the model you connected. hermes gateway setup later connects a messaging platform, so the same agent answers you in Telegram or Slack.

Hermes suits someone who wants to choose the model and keep the agent on their own hardware. If you would rather not run anything yourself, Grok Bot keeps its computer in the cloud for you.

Nuphos

Nuphos, from Zeabur, is an open-source workspace where a team runs coding agents together. It runs Claude Code, Codex, OpenCode, Grok Build or Antigravity under each person’s own provider subscription, on your computer, on a server you manage or on a cloud computer Nuphos manages. Teammates can open the same session and see the same messages, tool calls and results.

A team connects tools such as GitHub, Linear, Kubernetes, Grafana, cloud accounts and databases, then chooses which of them each conversation may use. Agents can also work in Slack and Discord threads. The quickstart’s first task is this request:

Text
Work in ~/code/my-app. Read the README, explain how to run the project, and identify one small improvement.

Show me the proposed change before committing anything.
The first task in the Nuphos quickstart

The transcript shows each file the agent reads and each command it runs. You get the steps to run the project and one proposed change to look at before anything is committed. Share the session and a teammate can read along or send their own corrections.

Triggers start agent work without anyone opening a conversation: a schedule such as 0 9 * * MON, which runs every Monday at 09:00 UTC (schedules use UTC, not your local time), or a webhook, for example one wired to a monitoring alert. Conversations started this way begin with full access, so limit the credentials and permissions a trigger’s agent can use instead of relying on instructions in its prompt.

Working alone on one laptop, you can run Claude Code or Codex directly, without a workspace around it.

Browsers and research tools

Some work only exists behind a login on a web page: an admin dashboard, a support inbox, a vendor console with no API. A browser agent does that work in the accounts you are already signed in to.

Aside

Aside is a web browser for macOS and Windows with an agent built in. It signs in to your email, dashboards and internal tools and works there directly: replies, comments, follow-ups, research, spreadsheets and documents on your computer. It runs on a ChatGPT or Claude subscription or on your own API key.

Its password manager fills credentials into the website without showing them to the agent and logs every use. Payments, posts and messages wait for your confirmation.

Engineers can also drive it from the terminal. aside runs a task described in plain language, and aside mcp lets a coding agent use the browser. MCP, the Model Context Protocol, is an open standard for connecting AI tools to outside tools and data; Claude Code and Codex both support it.

Bash
aside "Open localhost:3000 and run a smoke test"
A task from the terminal, from Aside’s developer page
JSON
{
  "mcpServers": {
    "aside": {
      "command": "aside",
      "args": ["mcp"]
    }
  }
}
The MCP client configuration from Aside’s developer page

localhost:3000 is a web app running on your own machine, and a smoke test is a quick check that its main pages load and work. The task runs in Aside as a session that aside session resume continues later. With the configuration above in a coding agent’s MCP settings, the agent can use Aside’s browser while it works on the code.

Aside acts in web pages; it does not edit your repository. Pair it with a coding agent: Claude Code or Codex changes the code, and Aside checks the result in the browser.

Primary references