Scout

Problem

Agents that work with private files have two common deployment models.

• Hosted agents require uploading that data to a third party.
• Self-hosted agents require the user to write connectors, configure permissions and trust each agent with broad access to the system.

Local models have made hosted inference optional, but most of the integration and permission problems are still unsolved. I cofounded Scout so that users can download, run and manage permissions for agents in one local platform.

Permissioned execution

Scout is a native macOS app with an embedded Python backend. Agents are distributed through the Agent Store, an app store for agents where every listing declares the permissions it requires.

An agent runs in a sandbox and cannot touch the file system or network itself. It sends each action it wants to perform to Scout, and Scout carries it out only if the user has granted the matching permission.

An agent never acts on the system directly. Scout performs each action only under a permission the user has granted.

Knowledge graph

Every agent needs context about the user's files and the applications it works with. When each agent maintains its own connector and index, the same data is parsed many times, every agent sees a different slice of it, and the user manages access agent by agent.

I built Scout's local knowledge graph over the folders and mailboxes the user approves, and every agent on the platform reads from it. Documents and emails are embedded for semantic search, and a language model extracts the entities they mention and the relations between them into a linked graph. As a result, an agent can follow connections across files, email and applications without a connector of its own.

Scout's first two agents are built on this graph: a Search Agent that retrieves files and facts from the indexed data in response to natural-language queries, and a Report Generator that compiles a report on a given topic from the user's email and documents.

SDK

Agents do not get direct access to the graph. Through Scout's SDK, a developer describes what the agent needs in a YAML query language, and Scout compiles the query, checks the agent's permissions and runs it. A single query can combine graph relationships, metadata filters and semantic search, which covers most of what an agent needs to know about a user's data.

The result is that developers build agents without writing connectors or running a backend, and users see and control what every agent can access from one place.