Building
The AI implements every task from your plan. It writes code, runs your toolchain, validates against done whens, reviews its own work and fixes what it finds.
The build phase is where your plan becomes working code. An orchestrator takes the approved features and runs Claude Code inside a sandboxed cloud environment. It sets up the environment, builds each feature in order, validates every task against its done whens and runs fix rounds where needed.
You can watch the build in real time, but you don't need to intervene. A supervisor monitors the whole process and only pauses for a person when it hits something it can't resolve itself, such as a secret that needs a real value.
Most AI coding tools need constant guidance. Bynary takes a different approach: the agent gets the full plan as context and works through it autonomously, with checks around it rather than a person watching over it.
- •No babysitting — Start a build, walk away, and come back to something ready for review
- •Built-in validation — The AI checks its own work against your done whens
- •Self-healing — Genuine failures trigger a fix round with structured feedback
- •Full traceability — Every file change, command and decision is logged
The build page shows real-time progress as the AI works through each task.
Environment Setup
Feature-by-Feature Implementation
Validation
Review Agents
Fix Round (If Needed)
Completion
Key Features
Sandboxed Execution
Builds run in ephemeral cloud VMs that are destroyed afterwards. Your production environment is never touched.
Real-Time Progress
See files being created, commands being run and tasks completing as they happen.
Quick Look Preview
While building, the sandbox runs the dev server and gives you a link to the work in progress.
Preview Deployments
With Coolify configured, each build can deploy to a live preview URL for reviewers. This can be switched off per project.
Supervisor
An always-on monitor detects stuck or looping behaviour, diagnoses causes and pauses for a person instead of wasting fix attempts on an unfixable blocker.
Ephemeral Infrastructure
If the plan needs a database or cache, Bynary can provision a temporary one for the build and a separate fresh one for the preview.
Git Native
Each build attempt has its own branch. Nothing lives only in the sandbox, and everything reaches your main branch through a pull request.
MCP Server Support
The build agent can use MCP servers configured for your project, with per-server safety rules you review before the build starts.
Each of these is optional and switched on per project or cycle:
A designated reviewer must approve the cycle's scope before it moves from planning to building.
The agent writes an implementation plan first, and a person must approve it before any code is written.
Ask a running build to pause at the next feature boundary, then resume it when you are ready.
An engineer builds the planned branch in their own IDE, then Bynary runs selected review capabilities against it.
Code that lands on your default branch outside Bynary creates a retroactive 'synced from GitHub' cycle, keeping history and architecture docs accurate.
How long does a build take?
It depends on the number and complexity of tasks. Bynary shows a cost, token and duration estimate before you start, calibrated from your project's history once it has some.
Can I stop a build in progress?
Yes. You can pause a build at the next feature boundary or cancel it. A cycle can also be rewound to an earlier phase.
What if the build fails validation?
The cycle still moves to review, showing which tasks failed. You can flag these as issues and trigger a fix build, or defer them to the roadmap.
Does the AI have access to external services?
The sandbox has network access for package installation. The agent can also use MCP servers configured for your project. Database schema changes must still land as migration files in your repository.
How does the AI know about my codebase?
It works from a clone of your repository, your architecture docs, every feature and task in the cycle with their done whens, design briefs and context from previous cycles.
Can I continue a build that was interrupted?
Yes. Progress is pushed as the build runs, so a build can continue from the review page and pick up the incomplete work.
We've thought carefully about these areas. Some are intentional design choices, others are on our roadmap.
The build sandbox, dependency tooling and validation are all specific to this stack.
Bynary deploys previews and provisions temporary infrastructure. It doesn't manage your production environment.
Live preview URLs require a Coolify server configured by an administrator.