Follow a real agent
from its map to the code.
Explore Local Deep Researcher through six real Ultra 13 views: architecture, workflow, tools, model evidence, coverage and a clearly labelled security beta preview.
These screenshots show a pinned public Python repository using LangGraph. They are product captures, not an interactive agent or execution trace.
Start with the connections.
The architecture view brings orchestration, model roles, tools, state and deployment declarations into one place. In this snapshot, the Ollama deep researcher workflow connects to SummaryState, a declared local model role and source-declared search calls.
Use it to orient yourself in an unfamiliar repository and identify which components deserve a closer look.
Review the pinned source fileFollow a step into its source.
Selecting web research shows its source registration, tool-call paths and connected routes. The diagram distinguishes fixed routes, conditional routes and loops. Tavily, Requests, DuckDuckGo, HTTPX and SearXNG calls are visible in the captured source.
These are possible source paths. The screenshot does not establish which branch ran or which service was called in a live session.
Review the pinned source fileSee what each tool declaration means.
The tools view lists the five captured call declarations. Opening the Tavily entry reveals the search call and its query, max_results and include_raw_content arguments, with a source path and file hash.
Tool registration does not establish effective permissions. Credentials, IAM enforcement and production network access remain unverified.
Review the pinned source fileCheck the model configuration behind the map.
The model panel shows the local_llm declaration, its default of llama3.2, and the workflow steps bound to that role. The expanded configuration excerpt lets a reviewer check the declaration directly.
A declared default is a starting point for investigation. Active runtime configuration can differ.
Review the pinned source fileKeep the scope and the gaps in view.
The source view identifies the repository, exact commit, capture time and selected variant. This snapshot captured ten files. The framework version is unresolved, and the coverage view distinguishes mapped declarations from runtime verification.
A map is easier to review when its provenance and remaining uncertainty travel with it.
Review the pinned repositorySecurity findings are coming soon.
This beta screenshot previews a source-review candidate that points to model-message construction in graph.py. It includes the scanner rating, source evidence, recommended review action and an explicit statement that successful prompt injection has not been demonstrated.
Beta preview only. Security findings are coming soon and are outside the initial architecture-mapping release. Static candidates require review; they do not demonstrate runtime safety or a successful attack.
Review the pinned source fileYour repository.
A clearer starting point.
Join the waitlist for free access during the initial launch.