An operations team wants an AI assistant to assemble a report from PDFs and spreadsheets. IT wants to know which files the assistant can access, where their contents go, and who reviews the output. You can answer those questions with a small FileHop pilot.
Start with one workflow and one AI client. FileHop handles supported file operations on the computer. The client decides which model receives prompts and tool results, and the team decides which inputs, outputs, and actions to authorize.
Choose where the model runs
- On the employee’s device. Use a supported local model and local FileHop tools. To require offline operation, test the complete client configuration with external network access blocked after installing the model. Include embeddings, sync, web tools, and telemetry in that test.
- On a company-hosted endpoint. Use a client that supports your model endpoint, authentication, and tool calling. Requests still travel from the employee’s device to the model service. Decide who can access the endpoint and how it logs requests.
- With an approved cloud provider. Use a provider and account approved for the data involved. The client may send document excerpts, previews, filenames, and tool results to that provider even when FileHop converts or edits the files locally.
Bring your own model (BYOM) describes choosing a model or model endpoint for the client. Bring your own key (BYOK) describes using your own provider credentials. Ask where inference happens and which account, endpoint, and retention settings apply to the actual request.
Compare clients that can launch local MCP tools
FileHop uses local MCP over stdio. The client starts the installed FileHop executable on the same computer. A remote MCP URL field alone cannot launch that process. The options below are based on vendor documentation. Test the exact client, model, FileHop build, and workflow before rolling it out.
- Jan Desktop
- A desktop chat interface for operations, research, or document work
- Model choice. Run a model on the device, or configure a supported cloud provider or compatible endpoint. Choose a model with tool calling.
- FileHop connection. Add FileHop by its local launch command in MCP settings.
- Model setupMCP setup
- AnythingLLM Desktop
- Teams that want document workspaces and a desktop agent
- Model choice. Choose a local runtime such as Ollama or LM Studio, an approved provider, or a compatible endpoint.
- FileHop connection. Configure FileHop as a local MCP server and use its tools through the desktop agent. This client supports MCP tools, so use tool discovery when a resource is unavailable.
- Model providersDesktop MCP
- LM Studio
- A desktop chat for evaluating on-device models
- Model choice. Download or load a model supported by the local runtime. Model size and tool-calling ability affect the pilot.
- FileHop connection. Add a local server in the app’s MCP configuration. Keep the model API local to the device for a single-device pilot.
- Local modelsMCP setup
- VS Code with Copilot chat
- Teams already comfortable with an editor
- Model choice. Configure a supported provider or local model for chat. Check the organization policy for bringing your own model key and choose a model supported by the agent.
- FileHop connection. Use a local stdio server in VS Code’s MCP configuration. Review workspace trust, enabled tools, and approvals for the pilot.
- Model selectionMCP setup
- Codex CLI
- Technical teams that already use Codex
- Model choice. Use the organization’s configured model or a supported local provider through the CLI’s open-source model mode where available.
- FileHop connection. Configure a local stdio server and review the agent’s shell access and approval settings separately.
- Codex MCP configuration
- Claude Code
- Teams already running Claude through an approved account or cloud deployment
- Model choice. Use Claude through an approved deployment such as Anthropic or Amazon Bedrock. Treat this as a managed Claude path.
- FileHop connection. Use a local stdio server. Administrators can manage which servers are allowed.
- Deployment optionsMCP setup and controls
Run a pilot with one folder and one job
- Pick a representative task. Put synthetic or approved sample files in a dedicated input folder. For example, ask the agent to make a Word briefing from two sample PDFs and save it to a review folder.
- Set access in both places. In FileHop’s Agents settings, grant read access to the inputs and write access to the separate review folder. Enable the tool groups the pilot needs. In the AI client, configure shell and filesystem access, disable unused connectors, and require approval for writes or external sharing where the client supports it. FileHop grants apply to FileHop tools. The client’s shell and other connectors have their own permissions.
- Verify the boundaries. Launch FileHop using the client’s stdio setup. Ask the agent to inspect an approved sample, then confirm that a FileHop read of a harmless test file outside the grant and a FileHop write to the read-only folder are refused. Check the client’s other access routes and confirm where the model request goes. Treat instructions inside a document as source content and keep external sends under human control.
- Review the output in FileHop. Where the installed build supports it, have the agent open the saved output in a FileHop window and attach it to a session. Compare the briefing with the sources, then review FileHop’s Agent activity log for calls, refusals, and output paths. Have a person approve the deliverable before it leaves the review folder.
FileHop’s activity log covers operations through its tools. Review the AI client, model provider, and other connectors separately for their activity and retention. Use the vendor’s organizational controls wherever the team needs to enforce the same settings across devices.