Google Gemini Deep Research as an MCP server. Durable background runs, spend control, context-safe report reading, citation verification, and private-corpus grounding.
Strong tool naming (verb_noun pattern throughout) and comprehensive descriptions (avg 250+ chars). All 10 tools have clear, actionable descriptions explaining WHAT, WHEN, and consequences. Input schemas present with types and constraints (enums, min/max). However, output schemas are not explicitly documented in the source, responses are described in prose but lack formal JSON Schema definitions. Error handling guidance is present in descriptions (e.g., 'budget refusal comes back as a result rather than an error') but not systematized. Tool composition is excellent: each tool has one clear responsibility, and chaining is well-supported (e.g., create_research returns runId for get_research_status). Security is strong: no secrets in parameters, destructive operations clearly marked. Naming is consistent and unambiguous (e.g., list_research_runs vs list_reports vs get_research_status are all distinct).
Ask whether a run of a given cost band would fit inside the per-provider-key budgets, BEFORE anything is spent. Free, and it reserves nothing — asking does not hold the money. Answers either that it fits, or every provider budget that would refuse it, with what remains and the date that period resets. Budgets are per key, not one global cap, so four keys is four independent answers.
Start a Dossier research run on one of your machines or in the cloud. THIS SPENDS MONEY: a run costs roughly $1-3 (fast) or $3-7 (max), charged against the provider-key budgets on whichever machine executes it, and the estimate is reserved at the HIGH end of the band BEFORE the run starts. Returns a run handle immediately and never waits — a run takes 4-60 minutes and continues with or without you. If a budget would be exceeded the run is refused before any spend, and that refusal comes back here as a result (every offending provider, what remains, and the date the period resets) rather than as an error. Poll `get_research_status`; the report URL and its markdown files appear there when it lands. Call `check_budget` first if the answer would change what you do — that one is free.
Your per-provider-key budgets and what remains. Budgets are per key, not one global cap — four keys is four independent limits. Returns every key you have set, what you have spent this period, what remains, and the date the period resets. If a key has no budget set, it is not returned.
Your account credits and what they are for. Returns the balance and the date it expires, if it does.
Output schemas not formally documented. Responses described in prose (e.g., 'Returns state, cost band, panel members, report URL') but no JSON Schema definitions visible. LLMs cannot reliably parse unstructured output or plan downstream tool calls without explicit response schemas.
Error handling not systematized. Descriptions mention error cases (e.g., 'budget refusal comes back as a result') but no formal error response schema or recovery guidance. LLMs need explicit error classification (retryable, user-fixable, fatal) to plan recovery.
Pagination not fully specified. list_research_runs and list_reports accept 'limit' but no 'offset', 'cursor', or 'total_count' documented. Large result sets risk context window exhaustion without clear pagination semantics.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 66 | 2026-07-28+ | v2 |
Your machines and their provider keys. Returns every device you have paired, its id, its name, which providers it has keys for, and whether it is online. The cloud is always online.
Where a run has got to. Returns its state — running, retrying, awaitingAnswer, paused, stalled, done, failed or cancelled — how long it has been going, its cost band, its panel members and, once it has landed, the published report and its markdown files.
List the reports you have read. Returns their ids, titles, questions, when they landed, and their cost. Does not return the markdown — call `read_report` to get it.
List your runs, newest first. Returns their ids, states, questions, cost bands, and how long they have been running. Does not return reports — call `get_research_status` on a run to read its report once it lands.
Read a report's markdown. Returns either the distilled written report or the source record — the whole gathered evidence, structured as a registry of claims and their citations. Both are markdown. Distilled is what you read; source is what you verify against. The report is streamed in chunks; if you hit the character ceiling mid-chunk, the chunk is returned whole and you can call again with a higher ceiling or a different file.
DESTROYS WORK IN FLIGHT: end a run that is running or retrying. The run cannot be restarted — it is gone. Nothing is refunded. Returns the run's final state and cost. Stopping a run that is already done, failed or cancelled is a no-op and returns its state.
Confirmation pattern missing for destructive operations. stop_research destroys work in flight with no dry-run or confirmation step. Agents can accidentally cancel long-running research without a safety gate.
Tool annotations present (toolAnnotations=true) but not visible in source. Verify that readOnlyHint, destructiveHint, and idempotentHint are correctly applied to all tools (e.g., stop_research should be destructive, get_* should be readOnly).