luckycat133 Gemini Assist
Extend the current task with Gemini assistance: a second opinion, grounded current-web evidence, visual or mixed-input understanding, or Deep Research. Use the user's task to pick one tool; do not tour the Gemini surface.
This surface deliberately exposes no history, Cookie, Scheduled, Gem, Prompt, manifest, or cleanup tools.
Choose The Tool
| User intent | Tool | What success means |
|---|---|---|
| Second opinion, critique, code or design review | gemini_ask |
the answer is compared or incorporated into the agent's own conclusion |
| Current-web question that should name sources | gemini_search |
answer plus observed sources and a truthful grounding_state |
| Understand one image or screenshot | gemini_understand_image |
analysis tied to that image and used in the surrounding task |
| Understand files, URLs, or mixed evidence together | gemini_understand |
every input keeps its id and per-input outcome plus one synthesized analysis |
| Multi-source investigation that yields a durable report | gemini_research |
one asynchronous start returns a preserved operation handle |
Server And Installation
The dedicated assistance entrypoint is gemini-mcp-assist; the MCP server name is gemini_assist_mcp.
uvx --from git+https://github.com/Luckycat133/gemini-web-mcp@main gemini-mcp-assist
The catalog is exactly five tools: gemini_ask, gemini_search, gemini_understand_image, gemini_understand, and gemini_research.
Models: flash-lite, flash (default), thinking, pro; thinking_level is standard or extended.
Live calls require Gemini Web account Cookies (GEMINI_PSID and the recommended matching GEMINI_PSIDTS); without them the tools return typed authentication errors instead of pretending to work.
Truthful Grounding
gemini_search reports one grounding_state:
grounded | answer_only | unavailable | failed
groundedrequires observed source URLs insources;- an answer without observed evidence is
answer_onlyand is never labeled grounded; - an empty answer is
unavailable; an errored search isfailed.
Do not relabel a source-free answer as grounded and do not invent source URLs. When a question needs evidence the quick search did not observe, escalate to gemini_research.
Information, Not Artifacts
Search and understanding return information to the calling agent, not files. Synthesize the answer or analysis into the user's task instead of dumping raw Gemini output. A completed Deep Research report stays in the retained Gemini chat; recover it through the preserved chat identity instead of expecting a downloaded Artifact from these tools.
Deep Research Starts Asynchronously
gemini_research starts one Deep Research run and returns immediately after the upstream research has started:
- the structured result carries an opaque
operation_idplus the preservedupstream_operation_idandupstream_chat_id; stateisqueuedorrunning; a started run is never a completed report;timeout_secondsbounds only the plan and start phases, never the report itself;- the research chat is retained by default so the report stays recoverable;
- preserve every returned identifier and never start a duplicate run because one call timed out.
Deep Research requires an AI Plus subscription; report the typed CAPABILITY_UNAVAILABLE result instead of retrying.
Boundaries
- Pure image, video, or music generation is a creation task, not assistance.
- Gemini account administration — history, Notebooks, Scheduled, Gems, Prompts, cleanup — needs explicit account surfaces.
- Developing the gemini-web-mcp repository itself is development work, not assistance.
- Do not call tools outside the five-tool catalog above; this surface has none.
Standard Workflow
- Identify the assistance intent and choose exactly one tool.
- Supply the smallest complete input: a prompt with optional context, one image, or one typed input list.
- Read the structured result before trusting the compatibility text.
- Continue the user's task with the returned information.
- For Deep Research, preserve the handle and recover the report later.