AI coding agent for early-stage startups

Quick answer

For a startup of about five people, Nori adds an engineering teammate that lives in Slack. Anyone on the team can ask for a fix, a small feature or an answer about the codebase and get a pull request back. The Team plan is a flat $250 per month for five machines and unlimited members, model usage runs on the Claude or ChatGPT plan you already pay for, and setup takes an afternoon.

The problems a five-person team has with coding agents

Most early teams already use an agent on a laptop. The friction starts when they try to make it a shared resource.

The chores have no owner

Dependency updates, the flaky test, the customer-reported typo, the admin script someone promised sales. None of it is worth pulling an engineer off the roadmap, so it waits. With five people there is no platform team to hand it to.

The agent only works while a laptop is open

A local agent stops when its owner goes into a customer call or closes the lid for the night. The work that could run unattended is tied to the calendar of the busiest people in the company.

Usage-based bills are hard to plan

A startup budget is a spreadsheet with few rows. Tools that charge per seat plus credits, or resell model usage at their own rate, turn a good week of shipping into a surprise on the invoice. Teams that already pay for Claude or ChatGPT do not want to pay for the same tokens twice.

The context lives in two people's heads

Why the billing table has that odd column, which test suite is safe to skip, how a deploy is rolled back. The founders know. The agent does not, and neither does the next hire. Each person's agent config captures a different fragment.

Nobody has a week for tooling

Building your own cloud runner, Slack bot and credential handling is a real project. A team with eight months of runway should not be maintaining it.

How Nori helps an early-stage team

Flat pricing you can put in the budget

The Team plan starts at $250 per month. It includes five machines and unlimited members, so the co-founder who writes no code and the contractor who joins for a month do not change the bill. A solo founder can start on Developer at $99 per month with three machines. Both begin with a 30-day free trial and no card.

Your existing Claude or ChatGPT plan

Connect the subscription you already have, or API keys, or an LLM gateway. Nori never charges for tokens. The background on why that matters is in do cloud coding agents mark up token costs.

Slack-first, set up in an afternoon

Install the Slack app and the GitHub app, connect your model account, pick your repositories. Then anyone writes @nori in a thread. Each thread gets its own isolated cloud machine, and the reply is a pull request. More on the workflow in AI coding agent for Slack that opens pull requests.

A teammate that is ready on day one

Nori ships Jiro, its own engineering teammate. During onboarding Jiro reads your repositories, asks how the team ships, and writes that down as the team's skillset. If you prefer Claude Code, Codex, Gemini CLI or any ACP agent, Nori runs those in the same machines.

Shared memory instead of tribal knowledge

Instructions, skills and memories are stored for the organization, not per laptop. When a founder explains the rollback process once in a thread, the agent can save it, and every later session starts with it. So does the next engineer you hire.

Schedules and webhooks cover the nights

A cron line turns a chore into a standing job. A webhook from your error tracker turns an exception into a proposed fix. See scheduled coding agents and how to turn a Sentry webhook into a reviewed fix PR.

A worked example: the first afternoon

A team of five: two founders, two engineers and a designer. One repository for the API, one for the web app. They already pay for Claude. The names below are illustrative.

1. 1:00 pm. A founder starts the Team trial, installs the Nori Slack app, and installs the GitHub app on acme/api and acme/web.
2. 1:15 pm. She connects the team's Claude subscription. No API key is created and no card is entered.
3. 1:30 pm. Jiro reads both repositories and posts what it found: the stack, the test commands, how deploys run. It asks three questions the code cannot answer. She answers in the thread and confirms the setup script.
4. 2:30 pm. The designer tries a real task in #product, with a screenshot attached.
5. 3:00 pm. An engineer reviews and merges the pull request.
6. 3:30 pm. The founder adds two schedules for work nobody owns.
7. 4:00 pm. Done. Nothing was installed on anyone's laptop.

The first request

Mika  2:31 pm  #product
@nori the empty state on the Projects page still says "No proejcts yet"
and the button is misaligned on mobile. Screenshot attached.
[projects-empty-mobile.png]

Nori  2:38 pm
PR: fix(web): Projects empty state copy and mobile alignment (#207)
- web/src/pages/Projects/EmptyState.tsx: typo fixed
- EmptyState.module.css: button was absolutely positioned; now in the
  flex column, centered below 480px
- Preview: before and after screenshots are in the PR description

Jon  2:55 pm
Looks right. Merged. !done

The two schedules

0 7 * * 1     Update minor and patch dependencies in acme/api and acme/web.
              Run the tests. Open one PR per repository. Post a summary
              to #eng.

30 8 * * 1-5  Read yesterday's merged PRs and open issues. Post a five-line
              standup summary to #eng: what shipped, what is in review,
              what is blocked.

From the next morning, the same thread pattern covers most small work: the founder asks how a customer's data is stored before a sales call, an engineer hands off a migration at the end of the day, and review comments on an open pull request go back to the agent that wrote it (GitHub PR comment agent).

What it costs a team of five

Five machines means five pieces of work can run at the same time. A machine is released when a thread is finished, so a team of five rarely has all of them in use. Your model provider's own plan limits still apply to the usage you send through it.

What to hand over first

• Customer-reported small bugs. Paste the report and a screenshot into a thread. Review the pull request.
• Questions about your own code. How is this calculated, where is this stored, what would break if we changed it. Useful for the founder who is not in the codebase every day.
• The Monday chores. Dependency updates, a standup summary, a check for failing tests on main.
• End-of-day handoffs. Describe the next step in a thread before you leave. Read the pull request in the morning.
• Internal scripts. The one-off export or backfill that sales or support asked for, with the query shown for review before it runs.

Keep product decisions, architecture choices and anything that touches production data with a person. The agent proposes through pull requests, and your branch protection decides what merges.

When the company grows, the same setup carries forward: more machines on the Team plan, custom agents per team, and for companies with strict requirements, the full stack in your own AWS VPC or Azure VNet with SAML and SCIM.

Frequently asked questions