zipnative-mcp tool explorer

The tools an AI assistant gets when you wire zipnative-mcp 1.0.0 into Claude Desktop, Claude Code, Cursor, VS Code or any other Model Context Protocol client. The server exposes 13 tools — the read side (inspect_zip, list_zip_entries, read_zip_entry, verify_zip, scan_zip_forward, sanitize_entry_paths), the write side (extract_zip, create_zip, modify_zip), the utilities (compute_crc32, inflate_raw) and the meta tools (describe_engine, draft_governance_issue) — plus 7 prompts and sandboxed zipnative://output/ resources. Each card shows a real example prompt, the MCP tool-call payload as the assistant would send it, and a Run button that executes the same operation on the published zipnative engine directly in your browser — no install, no LLM, nothing uploaded. The MCP guide holds the exact schemas.

1. Configure your client

Pick your client and paste the snippet into its MCP configuration. npx -y fetches zipnative-mcp on first use. ZIPNATIVE_MCP_OUTPUT_DIR is optional — without it the tools work on base64 only; with it, zipPath / sourcePath inputs, outputMode: "file" and the resources light up.

Other clients: Continue, Zed, Windsurf, Cline, Roo Code, the MCP Inspector and anything built on the MCP SDKs take the same command: "npx" / args: ["-y", "zipnative-mcp"] pair in whatever JSON shape they expect. The server speaks MCP 2026-07-28 and falls back to every 2024–2025 revision automatically.

2. Try the tools

Loading zipnative from the CDN… this happens once per page load.

archive under test

Drop any .zip here — it never leaves your browser

Every card runs the engine call behind its tool on the archive above and shows the result in the tool's output vocabulary — the shared inputs (strict, limits, verbosity, fields) are applied as the server applies them. What a browser cannot reproduce is stated on the card and refused with BROWSER_LIMITATION rather than ignored: the sandbox (zipPath, sourcePath, outputMode: "file", outputDir, resources), the worker pool behind create_zip.parallel, the operator ceilings (LIMIT_CEILING_EXCEEDED), the inline caps (ENTRY_TOO_LARGE), the response cache (_meta.cached) and the server half of describe_engine — all of which need the Node process. draft_governance_issue has no engine equivalent at all (it is the server's own policy engine) and offers Copy args only. The seven prompts are listed in section 3; they are text, not code.

3. The prompts, and what to ask

Seven read-only prompts teach the assistant the workflows and the rules — secure_extraction, reproducible_archive, incremental_update, forensic_scan, verify_before_trust, governance_contract, draft_issue_workflow. Once the config is pasted and the client restarted, ask it anything archive-shaped:

"Verify this build artefact, confirm it is reproducible with no encrypted or unsafe entries, then extract only the manifest."
"Build a deterministic EPUB from these three files — mimetype first, stored — and give me its SHA-256."
"This download is truncated. Scan it forward and tell me which entries survived, without trusting the result."
"Remove .env from this archive so that it is truly gone, not just unlisted."

Resources

MCP guide — configuration, every tool's inputs and outputs, prompts, resources, environment, the error and security models
zipnative-mcp on GitHub · on npm
CLI guide · CLI builder
Model Context Protocol specification