Deep Research — MCP

Turn a fuzzy question into a sourced research brief with open questions and next checks — not a final verdict. (MCP-oriented tool + prompt pack (read-only first).)

Id: mcp/deep-research · Slug: deep-research · Status: Untested · Untested on this machine — recipe reviewed for safety (no secrets, no destructive commands, no auto-spend). Mark tested after you run it locally.

What is this

MCP for MCP. Turn a fuzzy question into a sourced research brief with open questions and next checks — not a final verdict. (MCP-oriented tool + prompt pack (read-only first).)

How this runtime fits

This MCP pack is an MCP integration / tool-enabled agent surface for the parent AI Team. Multi-agent language applies only when a host actually composes multiple agents around these tools.

Parent AI Team: deep-research · What is an AI Team?

What it does

Turn a fuzzy question into a sourced research brief with open questions and next checks — not a final verdict. (MCP-oriented tool + prompt pack (read-only first).)

Who for

Builders working on Research jobs who can run MCP themselves.

Need to run

You will run this locally or on infra you control.
No production credentials in prompts or committed files.
MCP-aware client
Ability to allowlist tools

How to use

Review server_hints.md and implement only allowlisted read-only tools.
Load prompt.md as the client system prompt.
Run one Example in query with tools off; then enable library_search only.
Confirm path allowlist before any file read tool.

Limitations

Not a substitute for professional legal, medical, or investment advice.
Outputs can be wrong; human review required before publish or spend.
Stage-1 Library items are free recipes — marketplace ready-to-use teams remain separate.
Local models may omit sections; re-prompt once if needed.

Test status

Status: Untested · Untested on this machine — recipe reviewed for safety (no secrets, no destructive commands, no auto-spend). Mark tested after you run it locally.

Files

server_hints.md — MCP server/tool design notes
prompt.md — Client system prompt when tools are available

Full prompt / config / code

# MCP pack — Deep Research

## prompt.md (client)
You are a careful research desk. Goal: produce a brief, not a verdict.

RULES:
- Use only facts present in the user paste or clearly marked as [UNVERIFIED].
- Separate: Known / Inferred / Unknown / Next checks.
- Cite sources as the user labeled them (URL, doc name, date). Do not invent citations.
- No spending, no shell commands, no credential requests.
- Stop after the brief. Ask one clarifying question only if blocking.

OUTPUT:
1) One-sentence scope
2) Findings (bullets)
3) Open questions
4) Next checks (max 5)
5) Confidence: low|medium|high + why

## server_hints.md
Expose only read-only tools for stage 1, for example:
- `library_search` — search local catalog snippets the user installed
- `fetch_local_file` — read a path under an allowlisted directory

Tool design rules:
- No tool that sends email, posts to social, or charges payment.
- No tool that reads `~/.ssh`, `.env`, or browser cookie DBs.
- Arguments must be schema-validated; reject path traversal.
- Log tool calls for the human operator.

Example tool schema (illustrative JSON):
```json
{
  "name": "library_search",
  "description": "Search local BotShelf library snippets",
  "inputSchema": {
    "type": "object",
    "properties": {
      "query": {"type": "string", "minLength": 1, "maxLength": 200}
    },
    "required": ["query"]
  }
}
```

Wire this pack into an MCP-aware client with tools disabled until the allowlist is reviewed.

Example in

Question: Why did our signup conversion drop last week?
Sources:
- [A] analytics export 2026-09-01..09-07 (CSV summary pasted)
- [B] changelog: checkout button copy tweak on 09-03
Ask for a brief with next checks.

Example out

Scope: Explain likely drivers of signup conversion drop using only pasted analytics + changelog.
Findings:
- Drop concentrated on mobile checkout step (from A)
- Timing aligns with 09-03 copy tweak (B)
Open questions:
- Was A/B still running?
Next checks:
1) Diff mobile CTA copy pre/post
2) Error rate on payment iframe
Confidence: medium — single-week window.

Model / runtime notes

Protocol-level pack. Prefer read-only tools. Never ship credentials inside MCP configs.

Canonical Team

Parent: deep-research · implementation_id: mcp/deep-research

GitHub source

teams/deep-research/mcp · commit eb87e8e049fdf9908e438305ff827c7b617505b5 · license: free-use-at-own-risk

Structural check: PASS · env: structural (structural only — does not set Verified)

Related Library items

Related marketplace Team pages

Soft thematic links only — not identity merges. Marketplace Teams ≠ Library AI Team records.

Research Desk (marketplace Team page — separate entity)
Source Desk (marketplace Team page — separate entity)