Choose your setup.
Use a signed-in coding agent and configure the model profiles you want available. Choose manually, or opt into experimental Jev selection for a submitted task.
YOUR AGENTS · YOUR ACCOUNTSOPEN SOURCE. ON YOUR MACHINE.
Your agents.
One workspace.
Give your agent a task. Follow the work.
Know what happened before you move on.
Linux & WSL · Apache-2.0 · Bring your own agents
YOUR WORKSPACEC
CodexConfigured profile
Q
Qwen CodeConfigured profile
↗
Your next agentACP extension
Your choice. Optional Jev routing.
ILLUSTRATION · AVAILABLE AGENTS DEPEND ON YOUR SETUP
Local workspaceLive task visibilityExplicit permissionsChecks you can inspect
LESS GUESSING. MORE SEEING.
From a task
to a result you can check.
One real calculator task. An optional Jev choice among configured profiles. An agent edits the files. Urdia runs separate checks.
Captured from Urdia, not a simulated agent responseWhat this run proves
Read the demo description
The operator configures available execution profiles and connects Jev separately. They submit one calculator task with experimental automatic selection enabled. Jev returns a profile choice; Urdia validates it before admitting the task. The coding agent requests permission to edit two files. The operator approves that edit once. After the agent finishes, the operator explicitly runs the calculator checks and project tests. Their results appear separately from the agent's response.
A CLEAR PATH THROUGH THE WORK
Ask. Follow. Verify.
A small, deliberate loop. Agent completion and a checked result are different things.
Watch the task.
See streamed responses, reported tool activity and host timing. Review permissions, stop a turn, and inspect what the agent reported.
VISIBLE WORK · EXPLICIT CONTROLCheck the result.
Run an applicable Python or static frontend check with your consent. Inspect the files and tests before accepting the next step.
OBSERVED CHECKS · CLEAR LIMITSMORE THAN ONE WAY TO CHOOSE
Your choice.
Or a little help
from Jev.
Manual selection is the default. Jev can optionally choose a configured agent/model/effort profile and start one submitted task.
Laya and TypeLLM are optional advisory integrations. Coding agents and decision engines have separate roles.
Read how automatic Jev worksONE TASK, THREE SEPARATE RESPONSIBILITIES
01
The engine recommends.
A choice among host-prepared profiles, with an escalation path.
02
The host validates.
Urdia checks the exact profile and current task before dispatch.
03
You control permissions.
A routing decision never grants tool access or activates trust.
Experimental, opt-in, per-task routing. No claim of the best model or guaranteed savings.
SHOW THE RESULT. KEEP THE LIMITS.
One observed run.
No magic numbers.
See the recorded task and its separate checks above.
This is a development demonstration, not a benchmark or a guarantee for other projects. Model labels are agent-reported. Tokens and costs are shown only when actually observed.
BUILD SOMETHING. TELL US WHAT BROKE.
Take it for
a real run.
Start with the offline demo, then connect a signed-in coding agent to a small project. We want installation feedback and concrete bug reports.
A few honest answers.
Does Urdia include a model or an AI subscription?
No. Install and sign in to a supported coding agent yourself. Optional decision providers have their own configuration and may consume quota.
Can I bring a different coding agent?
Urdia supports configured external ACP agents. Compatibility depends on that agent's protocol behavior. See the provider-authoring guide.
Does Jev automatically run my whole plan?
No. Experimental automatic Jev mode can select a profile and start one task you submit. Reviewed plans do not automatically execute their remaining steps.
Does every completed task have verified output?
No. Agent completion is separate from applicable checks. Inspect your files and run suitable tests. Built-in checks cover specific Python and static frontend contracts, not every project.