Skip to guide

Welcome to Aditya documentation

Ask about this guide

A practical guide to the current source build: choose an interface, understand the runtime, and start with a project you can inspect.

Updated 9 sections · Current source build

On this page · 9 sections

Get started

Aditya is a research project being built around useful language and project work. These guides describe the local agent runtime and the interfaces already represented in the repository. The source-build guides are not a promise that every research direction is ready for use. For the paid cloud API, read the Billing and Pricing guides.

Start with a small, recoverable workspace. Confirm where the model runs and what the selected interface is allowed to do before giving it a task.

Choose a surface

Aditya Code is the coding agent’s terminal interface; Aditya is the general model. Start with aditya-code login, then choose a project with --cwd. The local server supplies the runtime for the web and desktop workspaces. Those paths share concepts, but their startup requirements and permission settings can differ.

  • CLI: begin in a selected project directory and inspect the task from the terminal.
  • Local server: provide the runtime used by the visual interfaces.
  • Model workspace: see the intended interaction on the public website. Its screens are illustrations, not live runs.

Give the task a reviewable boundary

A useful task describes the question, the permitted scope and the evidence you expect at the end. ‘Improve this project’ leaves all three open. ‘Explain why this test fails, without changing files’ gives the first run a boundary you can review. You can authorize a repair separately after inspecting the explanation.

For a change request, name the behavior that should be different and what must stay the same. For example, an accessible navigation fix should preserve link destinations, keyboard access and the mobile menu. A screenshot alone may show alignment, but it cannot establish those other requirements.

  • Question: what should the agent investigate or change?
  • Scope: which project and operations are authorized?
  • Acceptance: what observation would establish success?
  • Stop condition: when should the agent report a limitation instead of continuing?

Separate the interface from model inference

The public website demonstrates an intended interaction. Its example responses and workspace screens are not connected to your local source build. The CLI and local server need their own configured model service; opening either interface does not grant access to Aditya’s private model.

Check the endpoint before selecting a repository. A loopback address points to a service on your machine; a remote address changes the data destination. The model name must also correspond to a model available to that service. These guides do not download weights, provision an account or supply an API key.

Three things that should not be confused
SurfaceWhat it establishes
Public model previewAn illustration of the interaction, not a live answer.
Running local interfaceThe interface started; inference may still be unavailable.
Successful model requestThe configured service answered that request, not that the task is correct.

Read the guides in context

Setup, execution and verification answer different questions. A command can start the program without establishing that the model service is reachable. A tool can finish without proving that the requested task is complete. The guides keep those distinctions explicit.

Internal configuration names such as AROXA_PORT and the .aroxa folder remain unchanged for compatibility. Use the documented aditya-code command for Aditya Code and keep those exact internal identifiers where required.

Read a run as a sequence of observations

A task can involve several proposals, operations and checks. Read the record in that order: what the agent intended, what was actually permitted, what executed and what observation came back. A proposed test is not an executed test, and a successful file edit is not a passing test.

Preserve the distinction when reporting a result. ‘The route was inspected’ is narrower than ‘authentication is fixed.’ A timeout, declined approval or unavailable dependency belongs in the summary because it limits what the run can establish. The next task should address that missing evidence rather than repeat the same completion claim.

Keep the first change small and recoverable

Use a disposable copy or a version-control checkpoint before allowing writes. Inspect the starting state so you know which edits already belonged to you. After the run, review the changed files and the checks alongside the request; unrelated modifications should not disappear inside a broad cleanup.

Inspection tasks can still create local run metadata. In Aditya Code, local task and run records are stored under the selected workspace’s .aroxa directory. A request not to edit application code is therefore different from running a process that makes no filesystem writes at all. Keep sensitive data out of your first test and review records before sharing them.

Review a Git workspace
git status --short
git diff --stat
git diff

What is still research?

NIF explores how capability can be composed, checked, retained and withdrawn. The code contains runtime and adapter-management mechanisms, but broader transfer and consolidation claims require dedicated experiments. Read the research pages for the distinction between implementation and a measured result.

Access to the live model is limited to internal-team testing, not open for public access. Email hello@adityalab.in about access. API-key calls are paid per token from a prepaid balance. See the Billing and Pricing guides for rates, balance checks and recovery from billing errors.

Choose the next guide by the question you have

Read Quickstart if you need a first source-build run. Use the Aditya Code guide for account sign-in, workspace flags and shared sessions, or the runtime guide when a visual interface cannot reach its local server. Permissions and data flow explains the boundaries to check before allowing a consequential operation.

When something fails, troubleshoot the earliest failed dependency first. A missing Python module, an unreachable model and an unverified repair are different problems. Sharing the exact entry point and a redacted error makes a diagnosis more useful than reporting that ‘the AI does not work.’

Search covers these 8 guides. Press Esc to close.
Documentation