Claude integration
Choose a local or browser-local evidence path.
DEEPBOM has two separate Claude delivery paths. The desktop extension runs the packaged analyzer on the user's computer. The remote MCP App requires the user to choose one file inside the app and returns only a bounded result to the conversation.
Local Claude Desktop extension
- Download the version-matched
.mcpbasset from the DEEPBOM GitHub release. - Install it in Claude Desktop and select one dedicated model directory as
allowed_root. - Ask Claude to inspect a supported artifact in that directory. Paths outside the selected root are rejected.
This path is intended for confidential, large, sharded, or package artifacts. It is a local Desktop extension and is not a Claude.ai or mobile execution path.
Remote MCP App for host testing
- Add
https://deepbom.org/mcp/claudeas a custom remote connector in a supported Claude host. - Ask to inspect a model artifact. The app opens a file picker; Claude attachments are not read automatically.
- Choose one supported file and select Analyze. The browser sandbox returns a bounded, hash-bound static result.
The endpoint is prepared for private host validation. This page does not claim Anthropic review, directory approval, or verified behavior on every Claude surface.
Data-flow boundary
| Desktop MCPB | The packaged analyzer reads only paths under the directory selected during installation. It does not upload model bytes. |
|---|---|
| Remote MCP App | The user deliberately selects one file in the app. Analysis runs in the host's browser sandbox. DEEPBOM's service receives the bounded result, not the selected model bytes. |
| Conversation | The result can include file identity, tensor or graph summaries, findings, and evidence limits. Those facts may still be confidential. |
Evidence boundary
DEEPBOM reports serialized-artifact observations and deterministic static derivations. It does not establish measured latency, energy, task accuracy, clinical validity, actual runtime placement, or regulatory compliance.
The browser-local path accepts one TFLite, ONNX, GGUF, SafeTensors, Core ML .mlmodel, or ExecuTorch .pte/.ptd file. Package and sidecar closure remain local-tool workflows.