URSA Dev
A development environment built around AI-assisted coding.
Use Claude Code, Codex, local models or other tools together. Decide which model does what, what it’s allowed to touch, what gets reviewed, what needs your approval and what counts as actually finished.
URSA Dev
Configurables
You don’t have to rebuild the process every time you switch models, start a new session or change the kind of work you’re doing.
01Model Selection and RoutingUse different models for different kinds of work.
One model handles implementation. Another reviews. Local models take routine work. Stronger models handle the work that actually needs them.
You decide which models get used where, when the system switches and what happens if the preferred model isn't available.
02Multi-Model RolesDifferent models don't have to do the same job.
One model plans. Another builds. Another challenges the work. Another tests it.
Those roles stay in the development setup instead of being recreated every time.
03Automatic Review GatesSet rules for what has to happen before work moves forward.
Another model reviews the change, tests pass, you approve it or the workflow requires some combination of the three.
The model that did the work doesn't automatically get to declare its own work finished.
04Risk-Based EscalationRoutine work moves normally. Higher-risk work gets more scrutiny.
A UI adjustment moves through normally. A change touching authentication, architecture or sensitive data automatically brings in another reviewer, stronger verification or a human approval step.
You decide what counts as higher risk and what happens when the system sees it.
05Human Approval PointsDecide where the AI keeps going and where it has to stop and ask you.
Models inspect, write and test within the boundaries you set. Deployment, destructive operations or major architecture changes stop for approval when that's the rule.
The boundary stays in the development process instead of depending on you remembering it in the moment.
06Project-Specific RulesKeep the project's own rules inside the development process.
Coding conventions, architecture decisions, naming rules, approved dependencies, protected areas and things the models simply aren't allowed to do stay with the project.
You don't have to repeat the same rules every time a new session starts.
07Context and ContinuityKeep important project history available across sessions.
Projects have history. Decisions were made. Things were tried. Some worked. Some didn't. Requirements changed. Certain areas became off-limits.
URSA Dev keeps that history available so every new session doesn't start by trying to reconstruct the project from scratch.
08Scope ControlsLimit models to the parts of the system they actually need.
Set the files, folders, branches, worktrees, repos or types of changes that belong inside the task.
A documentation job doesn't need authority over the whole application. A small repair doesn't need access to everything either.
09Verification RequirementsDecide what has to be checked before something counts as done.
Tests, builds, runtime checks, diffs, repository state and whatever else the project depends on become part of completion.
The model saying something worked isn't the same thing as establishing that it worked.
10Cost / Token PolicyYou shouldn't need the most expensive model for every little thing.
Routine inspection, repetitive work and lower-risk jobs go to local, free or cheaper models when they're sufficient. Hard architecture, debugging or adversarial review goes somewhere stronger.
You decide where the money is worth spending.
11Provider Independence and FallbackChanging models shouldn't mean rebuilding the whole development process.
Use Claude today, Codex tomorrow and local models for routine work.
The project's rules, review process, verification and history don't disappear because the model changed.
12Recovery and Rollback RulesWhen something fails, recover cleanly instead of stacking patches.
Go back to a known state, keep the evidence from what failed and decide how to approach it again.
Use another model or another strategy when that makes sense. Don't blindly patch the patch until nobody knows what state the repo is in.
13Persistent Specialist RolesKeep specialist reviewers available when the work needs them.
Security, architecture, testing, documentation, adversarial review or another role specific to the project stays available as part of the environment.
Work that always needs the same kind of scrutiny automatically gets the right reviewer.
14Reusable Workflow ProfilesDifferent kinds of work follow different processes.
A bug fix, a refactor, a deployment and a production emergency don't need the same workflow.
Set up each process once, then reuse it when that kind of work comes up again.
15Parallel Development CoordinationRun several models or operators without losing track of who is changing what.
Separate worktrees, repos, responsibilities and integration points keep parallel work from turning into one shared mess.
You get the speed of parallel development without pretending that all of those lanes are the same state.
16Audit / Evidence TrailKeep track of what actually happened.
When several models work quickly, it gets hard to remember what was proposed, what changed, what got reviewed, what passed and what actually reached the repo.
URSA Dev keeps those states separate so the record doesn't depend on somebody remembering the conversation later.
17Privacy / Local-First RulesDecide what goes to cloud models and what stays local.
Cloud-eligible work goes to cloud models. Local-only code, files or project information stays on the machine.
Those rules stay in the development setup instead of being decided ad hoc every time.
18Operator InterfaceDifferent people use different interfaces without changing the rules underneath.
A developer works directly in the terminal and repo. Someone less technical uses a guided interface.
The review, scope, verification and approval rules underneath stay the same.
One setup
One setup looks like this.
Claude Code builds.
Codex challenges meaningful changes.
A local model handles routine work.
Higher-risk changes automatically get more review.
Tests and runtime checks show what actually worked.
You approve the steps that still belong to you.
Different project, different setup.Model choice
Model choice shouldn’t control the whole system.
Use Claude today and Codex tomorrow. Move routine work local. Use several models at once.
The development process shouldn’t fall apart because the model changed.
Human control
Human control stays where you put it.
AI handles more of the implementation without quietly taking over more of the decision-making.
You decide what it’s allowed to do, what needs review and where it has to stop.
We show the major pieces: model routing, review, risk handling, verification, scope, recovery, cost control, evidence and human approval. We don’t publish the exact rules and machinery that make those pieces work together.