sieve-mcp

Purpose

sieve-mcp is the MCP server for Sieve. Sieve is a macOS app used to scan local AI coding tool history for secrets leaked into prompts. sieve-mcp is bundled with the Sieve Mac app, requires no separate installation, makes no network calls, and all operations run locally on the device.

Main Capabilities

sieve-mcp provides the following tools, all of which are safe by design—no tool returns plaintext secret values:

• sieve_health_status: Returns service health and finding counts.
• sieve_self_test: Verifies whether scanner rules and record source coverage are healthy.
• sieve_findings_list: Lists findings from the most recent scan. Input: { "limit": 50 } (optional). Here, providerExposureAssumed: true means a secret was found in AI tool records, and it should be assumed that the provider (Anthropic, OpenAI, etc.) can associate that prompt with your account identity.
• sieve_issue_list: Lists scanner health issues—unreadable sources, permission errors, parsing failures.
• sieve_scan_history: Returns summaries of past scan jobs. Input: { "limit": 10 } (optional).
• sieve_scan_status: Returns the status of the current or most recent scan. Values: "idle" · "running" · "completed" · "failed: <reason>".
• sieve_check_text: Checks whether text contains secrets. Returns only a boolean—never echoes the input or detected values. Input: { "text": "<content to check>" }. Can be used to gate content before sending it to an external API.
• sieve_redact_text: Redacts secrets in text, replacing values with sieve:// placeholders. Input: { "text": "<content>", "project_key": "myapp" }. The response never contains the original values.
• sieve_vault_run: Executes a command and injects secrets resolved from the macOS Keychain as environmentvariable injection. Secrets never pass through the MCP channel. Input: { "command": "npm run deploy", "fingerprints": ["<fingerprint>"] }. stdout/stderr content is never returned—only line counts.

How to Use

1. Download Sieve for Mac from the App Store.
2. Open Sieve → Settings → MCP → turn on the Enable MCP server switch.
3. Configure Claude Code: add the corresponding configuration block to ~/.claude/claude_desktop_config.json, restart Claude Code, and run /mcp to verify that sieve appears.
4. Configure Claude Desktop: add the same configuration block to ~/Library/Application Support/Claude/claude_desktop_config.json and restart.

sieve-mcp uses stdio transport: the sieve-mcp binary reads JSON-RPC 2.0 from stdin and writes to stdout. The MCP client (Claude Code, Claude Desktop) manages the process lifecycle.

Use Cases

• Before sending content to an external API, use sieve_check_text to check whether the text contains secrets.
• Use sieve_redact_text to redact secrets in text, replacing them with sieve:// placeholders.
• Use sieve_vault_run to execute commands and inject secrets as environment variables via the macOS Keychain, with secrets not passing through the MCP channel.
• View findings from recent scans, scanner health issues, scan history, and scan status.

Limitations

• macOS only. Requires installing the Sieve app from the Mac App Store.
• MCP is off by default and must be manually enabled in Sieve settings.
• Local-only operation: stdio transport, no network calls, no cloud sync.
• No plaintext secrets in responses: this is a hard invariant in the scheduler and is not configurable.
• Only fingerprints are persisted:The discovery database stores only HMAC fingerprints and redacted previews.
• Keychain-backed vault: key values are stored in the macOS Keychain with biometric access control.
• Unreadable sources are reported: SCAN_TARGET_UNREADABLE will appear in sieve_issue_list—it will not be silently skipped.

Supported AI tool sources

Sieve correlates findings across source types—if the same key appears in both records and .env files, it will be flagged as multi-source exposure.

Links

• Mac App Store:https://apps.apple.com/app/sieve-ai-secret-scanner/id6747506504
• Product website:https://sieve.app