0.1.4.0initial public release0.1.4.1inspectcommand, reworked output0.1.4.2added SSE support0.1.4.3added VSCode MCP support, better support for non-MacOS, improved error handling, better output formatting0.1.4.4-5fixes0.1.4.6whitelisting of tools0.1.4.7automatically rebalance command args0.1.4.8-10fixes0.1.4.11support for prompts and resources0.1.5semver compatible naming, npm release0.1.6updated help text0.1.7stability improvements (CI, pre-commit hooks, testing), --json, logging & error printing, bug fixes, server improvements0.1.8-9fixes0.1.10-12automatic releases, updated to newer mcp client0.1.13-14Bug fix for printing. Now consider full signature when analyzing a tool.0.1.15Added local-only mode for scanning.0.1.16Fix when handling tools or args that contain dots.0.1.17Fix single server would throw an error.0.2.1mcp-scan proxyfor live MCP call scanning, MCP guardrails; removed NPM support0.2.2Fix bug with non-serializable exceptions. Move cross reference analysis to server-side. Now supports fuzzy similarity. Supportstreamable-httpas transport.0.2.3Bug fix. A broken MCP server would compromise the output for all the other servers within the same client.0.3.0Add Toxic Flow analysis.0.3.1Add cosmetics. Now defaults toscanalso with--parameters.0.3.2Add user IDs, --control-server for managed use and --opt-out to disable tracking.0.3.3Fixinspectmode, bug fixes.0.3.4Various fixes. Scanning of built-in IDE tools.0.3.5Improving description of some built-in IDE tools.0.3.6Bug fix: tools without description were not working.0.3.7Added (experimental) feature to run mcp-scan as a mcp-server.0.3.8Argument Handling fix0.3.9Fixed bug inentity_to_tool, function exposed only whenmcp-scanis used as a library0.3.10Remove Trio dependency.0.3.11Binary builds.0.3.12Improved SSL handling; ability to pass headers for API interaction.0.3.13Get TFA labels from api. Bug fixes0.3.14Added new VS Code paths0.3.15-19Signed macos releases0.3.20Direct scanning, improved URL handling for other endpoints0.3.21Fix: handle paths with spaces; updated mcp library dependency0.3.22Fix: support more than 100 tools; allow specifying a version when scanning a package0.3.23Maintainace to ensure continued server compatability0.3.24Handle custom SSL certificates (via System/Truststore)0.3.25Improved error handling for control servers0.3.26Now pushing agent configurations once for the full machine, not once per agent0.3.27Comability updated for new server version and new python version0.3.28Support unknown MCP configurations0.3.29Addevocommand to push scan results to Snyk Evo0.3.30Add--skip-ssl-verifyflag to skip SSL cert verification0.3.31Add metadata on push and more detailed error logging0.3.32Adaptivly support sse urls.0.3.33Support server-level issues. Retry on analysis api timing out.0.3.34Error categories & server args. Enhanced redaction and traffic capturing.0.3.35Fixes. Send push-key to analysis server. This guarantees authentication for enterprise users. Empty config causing erros. Redacting secrets before pushing to analysis server.0.3.36Add fallback to check command in well known directories to reduce path not found errors.0.3.37Add possibility to use authentication tokens to scan authenticated mcp servers.0.3.38Fix--jsonfunctionality. Default to HTTP for servers with unknown types.0.3.39Add support to scan skills and clients without MCP servers.0.4Fix dependency issue, 0.4 version bump, include uv.lock0.4.1Add more agent support and rebalance command args earlier.0.4.2Add support for HTTP_PROXY and HTTPS_PROXY configuration.0.4.3Rename to agent-scan.0.4.4Use Snyk API Key for analysis.0.4.5Add possibility to scan agent on multiple users.0.4.6Add Snyk CLI support.0.4.7Bug fixes.0.4.8Bug fixes.0.4.9Bug fixes and improvements.0.4.10Add --ci flag and better support multi-user machines.0.4.11Bug fixes, Amazon Q support.0.4.12Add Amazon Q path.0.4.13Minor fixes.0.4.14'guard' hook installation flow for Claude Code and Cursor.0.4.15Support Windows and x86 builds0.4.16Bug fixes and improvements.0.4.17WSL support for well known client paths.0.4.18Add support for --ignore-issues-codes with --ci.0.5.0Add user consent flow for interactive commands and deprecate mcp-server and install-mcp-server options.0.5.1Dependency upgrades.0.5.2Dependency upgrades and enable skill scanning by default.0.5.3Improve redaction of args and env vars.0.5.4Fix bug in parsing of skills. Cosmetic improvements to the CLI. Security fixes. Added agent-guard support for codex.0.5.5Introduce bootstrap API and recursive check for mcp_config_globs and skills_dir_globs.0.5.6Per-agent discovery refactor with Claude Code support. Bundle all detect_secrets plugins in PyInstaller binary. Starlette security update.0.5.8Reverting changes.0.5.9Don't handshake (launch) stdio MCP servers by default. Add VSCode family discoverers (VSCode, Cursor, Windsurf, Kiro, Antigravity); only scan installed extensions and plugins. Upgrade aiohttp to fix CVEs.0.5.10Add Codex discoverer for MCP servers and skills. Add Claude Desktop support. Attach config-file path to each MCP server on upload. Fix remote URL handling. Stop probing host tool versions and enumerating home directories during bootstrap.0.5.11Discover Cursor's built-in/managed skills. Redact secrets from skill file contents. Extend guard hook change detection. Remove commands discovery from the Claude Code discoverer.0.5.12Start CLI binary uploads to S3.0.5.13Add Cursor Plugins and Opencode support. Bug fix in skill serialization. Agent guard installer detects unauthorized modifications of the hooks script. Agent guard installer installs hooks only if requested agents are already installed on the machine.0.5.14Fix --insecure usage with the Snyk CLI. Remove unused API calls.0.5.15Fix false error triggering on skipping guard installations on missing agents.0.5.16Security fixes. Increase analysis request timeout. Add CLI reference docs and refresh JSON output docs.0.5.17Introduce support for async analysis. Security fixes related to dependencies.0.6.0Introduce risk-based MCP server and skill analysis with scored findings, updated scan output, and dedicated risk and operational-failure controls. Continue documenting and supporting the v0.5.x experience during the transition. Add--push-keyand--machine-idflags (replacing the deprecated--control-server-H x-client-idand--control-identifier),--config-fileto load CLI arguments from a YAML file (precedence: code defaults < config file < explicit CLI flags), and--show-analysis-resultsto force synchronous analysis so push-key scans display results. Retry transient async-analysis config and submission failures with exponential backoff.0.6.1Add Agent Guard MCP server discovery during hook installation and at Claude Code, Cursor, and Codex session starts, including project-scoped configurations. Introduce explicit machine identities for hook events, with temporary installation fallbacks for machine identity and the Agent Scan executable. Fix OpenCode MCP configuration discovery for explicit paths and project.opencodedirectories. Reduce skill secret-redaction CPU usage through cached detection and faster filtering. Clarify Agent Scan installation instructions.0.6.2Fix bugs with secrets redaction and increase timeouts for API calls.0.6.3Add CodexSubagentStarthook support. Include the Agent Scan CLI version in theX-Userheader for Agent Guard hook and discovery requests. Validate and scope install-time variable substitution in hook scripts.0.6.4Add GitHub Copilot support: discover its MCP servers, skills, and plugins, and install/uninstall/report Agent Guard hooks (user-level and managed). Fixguard install --managedbeing refused when the invoking account had no local agent config, even though managed installs don't depend on it. Accept Copilot CLI'stype: "local"for stdio MCP servers. Remove the Dangerous Words risk from Agent Scan. Discover Claude Code plugins from the install registry, including plugin-declared MCP servers and skills.0.6.5Ask for confirmation before contacting each configured remote MCP server in interactive scans; use--dangerously-run-mcp-serversto skip prompts and contact every configured MCP server, including remote URLs. Unattended scans keep their existing consent behavior. Stop following HTTP redirects for remote MCP connections and refuse link-local and known cloud-metadata addresses, while still allowing loopback and private network addresses. Resolve and verify macOS signed binaries configured by command name so they receive stable identities across machines.0.6.5.1Verify and start Codex MCP servers configured with a relative command and working directory, such as the built-in Computer Use server, so they receive stable signed-binary identities and complete scan results.0.6.5.2Discover Claude Desktop plugins on macOS, including their skills and MCP servers, from the installed-plugin list.0.6.6Include Claude Desktop cloud connectors (name, URL, tools, and instructions, with URL query values redacted) in Claude Code's Agent Guard install and session-start discovery events, so connectors that Claude Code only sees by connection UUID can be identified.0.6.7Discover Claude Desktop on Linux (the official beta) from~/.config/Claude, including its MCP config and the skills and MCP servers of installed Desktop plugins.0.6.8Republish release binaries and package after a release-pipeline mismatch on 0.6.7; no functional changes in this bump.0.6.9Discover MCP servers that VS Code extensions register in code throughregisterMcpServerDefinitionProvider(enabled for VS Code only, not for forks such as Cursor). Discover GitHub Copilot CLI session projects. Bound the Agent Guard hook request to 60 seconds and fail open with a warning on timeouts and transport errors. Require asuccess: trueresponse when sending thehooksConfiguredtest event duringguard install, and include the hook's diagnostic when it fails. Security fixes: upgrademcp,pyjwt,urllib3, andmultidict.0.6.10Addguard install --check-installedto skip agents that are not installed for any user on the machine when installing managed hooks; managed installs still do not check by default. Carry the tenant ID in the Windows PowerShell hook command soguard uninstallcan revoke the push key on Windows; hooks installed before this change needTENANT_IDset in the environment to be revoked on uninstall. Security fix: bumpcertifito 2026.10.19.0.6.11Stamp the CLI version into the Agent Guard hook scripts at release time instead of at install time, soguard installcopies the bundled scripts unchanged and the release pipeline can Authenticode-sign the Windows PowerShell hooks. Document the supported platforms for the standalone binaries. Detect standalone CLI installs and scan Antigravity 2.0's canonical global and bundled skill directories.