Local agent setup

Let the agent run the audit where the artifact already lives.

DEEPBOM provides an Agent Skill and a zero-install npx path for Codex, Claude Code, and compatible local agents. The model stays in the local process; there is no hosted DEEPBOM analysis endpoint.

Use this route when the agent has shell and file access. For one attached file in ChatGPT, use the separately reviewed browser-sandbox integration.

Install the local Skill

Preview the exact managed files first. Add --apply only after reviewing the result.

npx -y deepbom@2.5.0 integrate codex
npx -y deepbom@2.5.0 integrate codex --apply

npx -y deepbom@2.5.0 integrate claude-code
npx -y deepbom@2.5.0 integrate claude-code --apply

The installer writes only the DEEPBOM Skill directory. It refuses to replace or remove files changed by the user.

Prefer a verified installation

The installed Skill includes scripts/verify-deepbom.mjs. It checks an explicit DEEPBOM_BIN, a package-bundled analyzer, and an executable already on PATH. It accepts only an engine that declares the compatible Agent and evidence contracts and passes a self-test, then records the exact engine version used. It does not contact npm unless --allow-download is supplied after network and installation permission is clear.

Engine releases are versioned independently from the stable public Agent contract. Parser and measurement fixes therefore do not change the reviewed tool surface by themselves.

node "<skill-root>/scripts/verify-deepbom.mjs"
node "<skill-root>/scripts/verify-deepbom.mjs" --allow-download

The JSON result records the selected command, version, whether network access was used, and the deterministic self-test. If resolution fails, report that no analysis ran.

Ask normally

Once the Skill is available, a supported request can be stated without memorizing the CLI.

Statically audit this ONNX deployment artifact.
Report MAC coverage, quantization contracts, unresolved evidence,
and the exact reproduction command.

The Skill selects DEEPBOM only for serialized deployment artifacts. It must not turn static estimates into claims about measured latency, accuracy, energy, or actual runtime assignment.

Run without installing a Skill

npx -y deepbom@2.5.0 capabilities --format agent-json
npx -y deepbom@2.5.0 self-test --compact
npx -y deepbom@2.5.0 audit "./model.onnx" --summary

Start with the bounded summary. Request a section, JSON Pointer, or evidence envelope only when the question needs more detail.

Optional local MCP

Transport	Local stdio using `npx -y deepbom@2.5.0 mcp`.
Tools	`deepbom_capabilities`, `deepbom_audit`, `deepbom_diff`, and `deepbom_explain_rule`.
Default response	Bounded human summary. Detailed evidence is explicit.
File boundary	The launch directory or roots declared through `DEEPBOM_MCP_ALLOWED_ROOTS`.

Gemini CLI can install the bundled local analyzer with gemini extensions install https://github.com/JunHwan-Kwon/deepbom-gemini. Start Gemini in your model directory and use /deepbom:inspect ./model.onnx. See Gemini CLI setup for permissions, data flow, updates, and validation.

Claude Desktop can install the same local server as a single desktop extension. Download deepbom-2.5.0.mcpb, then choose Settings, Extensions, Advanced settings, and Install Extension. Select the model directory the extension may read during installation.

For private testing on a Claude host that supports remote MCP Apps, add https://deepbom.org/mcp/claude as a custom connector. This is a separate browser-local path: the user selects one file inside the app, Claude attachments are not read automatically, and only a bounded result is returned to the conversation. See the Claude integration guide. Public discovery remains subject to Anthropic review.

Optional candidate review

DEEPBOM Review 0.1.0 is a separate local CLI and MCP package for single-file ONNX and TFLite candidates. An existing agent or Olive produces the candidate; DEEPBOM preserves the baseline/candidate analysis, semantic diff, fixed policy, external evaluation receipts, and an accept/reject/hold decision in a ZIP.

Olive rank records and completed MLflow runs can be imported with explicit model, dataset, evaluator and environment hashes. Missing or incomparable evidence stays unknown. These producer-declared measurements do not become DEEPBOM-executed benchmarks. Model bytes are optional; review commands do not train or deploy.

Use deepbom_review_capabilities before preparing a request. Keep criteria fixed across candidate revisions and retain each result. GitHub Copilot setup includes configurations for Copilot CLI and VS Code. The package also emits Gemini CLI settings and can be called from any agent with local shell access.

Review workflow, installation and adapter guides

Machine-readable discovery

DEEPBOM can be discovered or recommended in a plain chat. Actual analysis requires a connected local shell, Agent Skill, local MCP process, the DEEPBOM ChatGPT plugin for one authorized attachment, or the separate Claude MCP App with explicit in-app file selection.

The web workspace also offers an optional browser-observed runtime benchmark. That measurement path is separate from the local Skill, CLI static audit, and ChatGPT attachment tool; an agent must not describe static output as measured latency.