Repository navigation
PromptGuard CLI v2.3.0
·
34 commits
to main
since this release
PromptGuard CLI v2.3.0
Added
promptguard mcpspeaks MCP 2026-07-28 and 2025-11-25. A 2026-07-28
client needs noinitialize: each request names its revision in_meta,
server/discoveranswers with every revision the server speaks (2026-07-28,
2025-11-25, 2025-06-18, 2025-03-26, 2024-11-05), its capabilities and its
identity, and a revision it does not speak gets-32022
(UnsupportedProtocolVersionError) with that list. Those results carry
resultType, the server's identity in_meta, and, fortools/listand
server/discover,ttlMsandcacheScope;pingandlogging/setLevel,
which 2026-07-28 removed, are unknown methods there. Older clients still
open withinitialize, which now also agrees on 2025-11-25.promptguard_scan_texttakescontent_type(prompt, the default, or
response), so an agent can scan a model's output as well as a prompt. It is
sent to the API as the scantype; any other value is a tool error the model
can correct.
Security
promptguard mcpno longer takes an API key through the chat.
promptguard_authused to accept the key as a tool argument, which put it in
the model's context, the client's transcript and any log of tool calls. The
tool now only checks authentication and, when there is none, opens the
dashboard and tells the user to save the key withpromptguard loginor
PROMPTGUARD_API_KEY. A client still sendingapi_keyis refused, and
nothing is saved.promptguard_redactno longer returns the un-redacted text. It passed the
API's response through whole, including theoriginalfield, so the PII it
removed came straight back into the conversation. It now returnsredacted
andpiiFoundonly, as the hosted server does.
Changed
- The MCP server follows the current spec.
initializenegotiates the
protocol version (2025-06-18, 2025-03-26 or 2024-11-05) instead of always
answering 2025-03-26. It no longer advertises aloggingcapability it never
implemented. Every tool has atitleand behaviour annotations, and the scan,
redact, scan-project and status tools declare anoutputSchemaand return
structuredContent. Calling a tool that does not exist is now a JSON-RPC
-32602error rather than a tool result. - The shared MCP tools match the hosted server's exactly.
scan_text,
redactandstatusnow carry the hosted server's title, description,
parameters and annotations:scan_textandredactare read-only (each
call is still one metered Request, which their descriptions now say), and
tests/mcp-tool-definitions.json, a copy of the platform's published tool
manifest, fails a test on any drift. Every tool'sinputSchemasets
additionalProperties: false, and an undeclared argument is now a tool error
naming it instead of being ignored.
Drop-in LLM security for your applications - Built with Rust + Tree-sitter
Installation
One-line install (macOS/Linux):
curl -fsSL https://raw.githubusercontent.com/promptguardhq/promptguard-cli/main/install.sh | shManual install:
- Download the binary for your platform below
- Make it executable:
chmod +x promptguard-* - Move to PATH:
sudo mv promptguard-* /usr/local/bin/promptguard - Verify:
promptguard --version
Checksums
SHA256 checksums are provided for each binary (.sha256 files).
About PromptGuard CLI
- ✅ Real AST transformations using Tree-sitter
- ✅ TypeScript, JavaScript, Python support
- ✅ OpenAI, Anthropic, Cohere, HuggingFace providers
- ✅ Automatic backups with safe revert
- ✅ Single 5.3MB static binary
Supported Platforms
- macOS ARM64 (M1/M2/M3)
- macOS x86_64 (Intel)
- Linux x86_64
- Windows x86_64
📖 Documentation: https://docs.promptguard.co/cli
🏠 Homepage: https://promptguard.co
🐛 Issues: https://github.com/promptguardhq/promptguard-cli/issues