gemini-web-mcp

name	gemini-web-mcp
description	Operate an installed Gemini Web MCP server safely: inspect the tool manifest, choose the narrowest profile and read-only workflow, manage explicitly selected history/notebook/account tasks, and verify generated media artifacts. Use for MCP tool operation; do not use for repository implementation, tests, CI, packaging, or releases—use gemini-web-mcp-development instead.
license	MIT-0
compatibility	Requires Python 3.11+ and an installed server (verify with uvx --from git+https://github.com/Luckycat133/gemini-web-mcp@main gemini-mcp-onboarding). Live Gemini calls require account Cookies; image verification requires the image or all extra.
metadata	{"version":"0.1.1","openclaw":{"emoji":"♊️","homepage":"https://github.com/Luckycat133/gemini-web-mcp","requires":{"bins":["uvx"]},"primaryEnv":"GEMINI_PSID","envVars":[{"name":"GEMINI_PSID","required":false,"description":"Optional __Secure-1PSID Cookie for authenticated Gemini Web calls."},{"name":"GEMINI_PSIDTS","required":false,"description":"Optional matching __Secure-1PSIDTS Cookie recommended for session stability."},{"name":"GEMINI_PSIDCC","required":false,"description":"Optional __Secure-1PSIDCC Cookie forwarded when configured."},{"name":"GEMINI_PROXY","required":false,"description":"Optional HTTP or HTTPS proxy used by the MCP server."},{"name":"GEMINI_BROWSER_COOKIE_TIMEOUT_SECONDS","required":false,"description":"Optional bounded macOS browser-credential authorization wait for Cookie discovery."},{"name":"GEMINI_TOOLS","required":false,"description":"Optional primary-server tool profile such as model, core, or all."}]}}

Gemini Web MCP

Use this skill only to operate the installed primary or low-token MCP server. For source changes, tests, CI, packaging, compatibility probes, or releases, stop and use the separate gemini-web-mcp-development skill.

Start Here

1. Before configuring an account, verify the installed server with uvx --from git+https://github.com/Luckycat133/gemini-web-mcp@main gemini-mcp-onboarding; this calls a real auth-free text tool and strips Gemini Cookie variables from its child process.
2. Prefer gemini_get_tool_manifest before choosing primary-server tools. It is always exposed by src.server, including the narrow model profile.
3. Check manifest current_enabled, groups, and workflows; do not hard-code tool counts because the static manifest can include groups not loaded in the current process.
4. On the low-token server, prefer auth-free account(action="manifest") and account(action="capabilities") before account calls that initialize Gemini.
5. Prefer read-only discovery first: gemini_doctor, manifest/capabilities, gemini_probe_web_features, metadata-only history search, profile diagnostics, and inventory/list tools.
6. Treat privacy=reads_private_chat_text and other private text tools as explicit-user-intent tools: gemini_read_chat, gemini_export_chat, gemini_search_chats(scan_turns=true), and research-report create actions that read chat text.
7. Treat destructive tools as requiring explicit user intent: gemini_delete_chat, gemini_cleanup_test_artifacts(dry_run=false), gemini_delete_scheduled_action, gemini_reset_session, gemini_manage_gems(action="delete"), and prompt deletion.
8. gemini_reset_session changes only MCP/Gemini conversation state; it never changes agent memory or agent instructions.

Need the full tool/group/privacy/destructive map? See references/tool_surface.md. Load it on demand — the live gemini_get_tool_manifest remains the source of truth.

Tool Surfaces

• Primary MCP server: src.server
  • Use GEMINI_TOOLS=model or when an agent only needs to call Gemini models.