Google Analytics, Vercel Web Analytics, Leadsy, the Ship or Die traffic check, the TikTok Pixel, the OpenAI Ads Measurement Pixel and Behavly stay off until you accept. The help chat starts once you have chosen. The scheduling button starts once you have chosen. Read our Privacy Policy.Google Analytics, Vercel Web Analytics, Leadsy, the Ship or Die traffic check, the TikTok Pixel, the Meta Pixel, Google Ads, the LinkedIn Insight Tag, Microsoft Advertising, the OpenAI Ads Measurement Pixel and Behavly stay off until you accept. The help chat starts once you have chosen. The scheduling button starts once you have chosen. Read our Privacy Policy.
The Coder turns your goals into tested code changes in its own workspace, then takes them as far as your house rules allow, with Approve on risky steps.
The Coder is the StellarFirm assistant that writes software. You give it a goal or a ticket, it works in a private workspace with a terminal and a browser, runs the tests, and shows you the exact changes. What happens next depends on your house rules: it can open a pull request for you to review, or merge and ship when your house rules allow it. A merge, a deploy, and anything else that reaches people or spends your money waits for your Approve.
You set a goal. Write it in chat: "Fix the login redirect" or "Pick up the top ticket on [repository]".
The CEO briefs the Coder. The CEO assistant routes the goal and passes along the details, such as the repository and the issue. If something is vague, the Coder asks before it starts.
The Coder clones the repository into its own workspace. It does not work in a copy on your own machine. The clone and the work happen in the assistant's private computer, using the login you connected under Integrations.
It works with a terminal and a browser. It reads the code, edits files, and runs commands. For a change you can see on screen, it opens the page in a real browser and checks it.
It runs the tests. It runs the project's own checks and reports what ran and what passed.
You see the changes. Open the job in Jobs to read the actual diff and the commands it ran. Open the session's Terminal tab to watch it live.
The last step depends on your company, not on a fixed setting of the Coder.
If your house rules say
The Coder
Nothing special (the default first run)
Pushes a branch and opens a pull request, by default a draft, for you to review. Nothing waits for your Approve.
Review only (the default merge policy)
Opens a pull request and leaves the merge to you or your team.
Merge after CEO approval
Opens a pull request, then merges once the CEO approves that step.
Auto merge on green
Merges once every check passes, with no extra ask.
Shipping
Follows its own rule: never (the default), after CEO approval, or automatically after a merge.
A step reaches other people, goes live, deletes, or spends your money
Waits for your Approve before it goes ahead.
Your house rules are the rules in your company brief plus what you allow without asking under Auto-approve in Settings. Nothing is auto-approved until you turn it on yourself. See approvals and the merge policy options.
Checkpoint
The first time you try the Coder, expect this path: you set a goal, the Coder works, and you get a pull request. A draft pull request is the default first-run outcome, and that is fine. Write your house rules when you want the Coder to go further.
A coding harness is the program the Coder codes with. StellarFirm's own is the default. In Settings you can choose another one for the whole company, or give one Coder a harness of its own, such as an open source runner that reaches an open model through StellarFirm. Only the CEO changes it, and using your own coding subscription needs a paid plan or the trial. A choice is never a permission: every Approve, budget and rule applies to every harness.
Read the transcript (43 seconds, no sound)
Pick how your Coder codes
Choose a coding harness for the company: built in, your subscription or open source.
Here, your own coding subscription becomes the default for every new Coder.
Or give one Coder an open source runner of its own.
It reaches an open model through StellarFirm.
A harness is never a permission. Every Approve, budget and rule still applies.
StellarFirm: AI Agents, Agentic Assistants for your work
Takes the oldest open issue, builds it, prepares the change
Before it pushes and opens a pull request
"Implement #42"
Builds that issue
Before it pushes and opens a pull request
"Fix the failing CI on PR #14"
Reads the failing check, fixes it in the smallest change
Before it pushes
"Build the pricing page"
Builds from your message
Before it pushes and opens a pull request
"Debug the flaky upload"
Investigates and finds the cause, with no write to your source host
No
"Review PR #14"
Reads the pull request and its checks, posts a summary
No, it only reads
"What is in the backlog?"
Lists open issues
No, it only reads
"Open an issue for the flaky test"
Files a well-formed issue
Before it opens the issue
"Merge PR #14"
Merges when your house rules allow it, otherwise asks you first
Per your house rules
Tip
Be specific about the outcome. "Fix the login redirect and open a pull request for review" tells the Coder where to stop. "Fix it and ship it" tells it you expect the house rules to carry it further.
For a Coder with the IT administrator role: what is down, who has which account and laptop, and the IT ticket queue
Novu
Reading your notification workflows, subscribers, sent notifications and Novu agents; triggering a workflow waits for your Approve, and deleting and cancelling wait for it too unless you turned on auto-approve for deletions or closing in Settings; building notification systems with the Novu SDK is code in your repository
For a Coder with the machine learning role: reading training runs and metrics, and reading the logs of your rented GPU machines when you give their ids. Both read only
Pointing a domain at your site with DNS records or nameservers, after your Approve. It registers a domain only when your company has no Product owner.
Connect your source host under Integrations, on stellarfirm.ai or in the StellarFirm desktop app. Connect your code lists which token to create and the permissions each step needs. Only the Coder can change code.
Coder, pick up the top open ticket on [repository]. Build it, run the checks, and show me the changes. Follow our house rules for what happens next.
PromptImplement a specific issue
Coder, implement issue #[number] on [repository]. Keep the change small, add tests, and open a pull request for review.
PromptBuild a feature from a description
Coder, on [repository], add [feature]. It should [behavior]. Check it in the browser and send me screenshots with the pull request.
PromptFix a failing check
Coder, the checks on [pull request or branch] in [repository] are failing. Find the cause, fix it in the smallest change, and tell me what you changed.
PromptDebug a bug from a report
Coder, users report that [symptom] on [repository]. Find the root cause first and explain it. Then fix it with a regression test.
PromptReview a pull request
Coder, review pull request #[number] on [repository]. Summarize what it changes, point out risks and missing tests, and say whether you would approve it.
PromptWrite tests for a module
Coder, write tests for [file or module] in [repository]. Cover the edge cases and one bug-style regression. Do not change the behavior.
PromptPlan a test strategy
Coder, look at [repository] and propose a test strategy. Rank the riskiest areas first and tell me what to test, what to skip, and why.
PromptRefactor safely
Coder, refactor [function or module] in [repository] to remove duplication. Behavior must not change, and the existing tests must still pass.
PromptWeigh a design decision
Coder, we need to decide between [option A] and [option B] for [problem]. Compare them for our codebase, recommend one, and write no code yet.
PromptSet up CI
Coder, set up CI on [repository] so every pull request runs lint, tests, and the build. Use caching and keep the permissions minimal.
PromptImprove a pull request description
Coder, rewrite the description of pull request #[number] on [repository] with a summary, the changes, how it was tested, and the risks.
PromptOpen a clear issue
Coder, open an issue on [repository] for [problem]. Include steps to reproduce, what you expected, and what happened.
PromptVerify a page in the browser
Coder, open [page or flow] from [repository] in the browser, click through it, and report anything broken with screenshots.
You can run more than one Coder at once. Give each one its own name and focus, for example a UI Coder and an API Coder, and its own repository and login. Each works in its own private workspace, so they do not collide.
Address one by name: "API Coder, add the endpoint."
Or write the goal without a name, and the CEO assistant picks the Coder whose focus or repository fits. If it cannot tell, the default Coder takes it.
Approvals and retries go back to the Coder that ran the job.
A Coder can also take on up to three roles, such as frontend, backend, devops, mobile, QA, security, IT administrator, WordPress, game developer, AI engineer, machine learning engineer, or healthcare app developer, so it loads the skills that fit. Add a separate Coder when you need a fourth.
The machine learning engineer role prepares training data, builds fine-tuning projects for language models, vision models and tabular models, plans evaluations, and sets up serving, with skills for each of those and for estimating the GPU a run needs. No StellarFirm computer has a GPU, so the Coder proves a project with a tiny run on the CPU, pushes a branch, and writes a RUN.md with the GPU class, the count, and the hours the real run needs, plus a stop rule in hours. It does not hold a GPU provider's token on StellarFirm, so you start the real run from RUN.md on your own machine or your own provider account. If you have given a Coder's own computer a provider token, starting a paid run from it waits for your Approve, and so does uploading to the Hugging Face Hub. Anything the Coder would pay a model provider for with your own key, even a small test, is asked about first, with the model and the number of items named. Connect Hugging Face, Weights & Biases, and RunPod so it can read your models, runs, and GPU logs. All three are read only.
Set Coders up in Settings. Assistants spend credits when they work, so see Usage for what your assistants used.
The Coder follows your house rules, so check them before you allow merging. A rule such as "merge small fixes after checks pass" is powerful. Start narrow and widen it as you build trust.
Name the repository. The Coder needs to know where to work. Connect your code host first, see Connect your code.
State when to stop. Say "open a pull request for review" or "merge once checks pass". If you say nothing, your house rules decide.
Keep goals small. One clear change per goal gives you a diff you can read in minutes.
Read the diff. Open the job in Jobs before you approve a risky step.
Work happens on its own branch. The Coder never force-pushes, and your source host's own branch protection still applies.
Secrets stay out of prompts. Never paste passwords or tokens into chat. Connect services through Integrations.
Vague goals get questions. If the Coder asks, it is saving you a wrong turn. Answer in one line.