Skip to main content
Glama
io.github.MohdSaleh/tinyfish-guided-research

io.github.MohdSaleh/tinyfish-guided-research

by MohdSaleh

TinyFish Guided Research MCP

An MCP server that adds a simple research workflow on top of TinyFish Search and Fetch.

The client model does the reasoning. The server keeps track of the research run, handles the search and fetch flow, stores the state, checks evidence and citations, and tells the client what should happen next.

There is no LLM running inside the server.

How it works

A research run usually follows this flow:

Research request
      ↓
Plan
      ↓
Search
      ↓
Review sources
      ↓
Fetch useful content
      ↓
Track claims and evidence
      ↓
Check what is supported or still missing
      ↓
Research the gaps
      ↓
Verify citations
      ↓
Finalize

Search results are treated as candidates first, not as evidence by default.

The server also keeps duplicate sources from being counted more than once. This includes cases where the same paper or source appears through different URLs or mirrors.

Quotes are checked against the fetched source content, while semantic decisions such as whether a passage actually supports a claim are left to the client model.

Conflicting evidence is kept in the research state instead of being ignored, and citations are checked before the research is finalized.

Related MCP server: genpark-multihop-research-query-decomposer-skill

What it handles

  • Research state across multiple steps

  • TinyFish Search and Fetch calls

  • Source screening and duplicate handling

  • Claim and evidence tracking

  • Quote checks against fetched content

  • Conflicting evidence

  • Research gaps and follow-up searches

  • Citation verification

  • Research budgets and stopping conditions

  • SQLite and PostgreSQL persistence

Quick start

Hosted

The hosted MCP endpoint is:

https://tinyfish-guided-research-mcp.fastmcp.app/mcp

Use it as a Streamable HTTP MCP server.

Local

Requires Python 3.11+ and a TinyFish API key.

TINYFISH_API_KEY="your-api-key" uvx tinyfish-guided-research-mcp

Example MCP client config:

{
  "mcpServers": {
    "tinyfish-research": {
      "command": "uvx",
      "args": ["tinyfish-guided-research-mcp"],
      "env": {
        "TINYFISH_API_KEY": "your-api-key"
      }
    }
  }
}

The compatibility entrypoint is also available:

uvx --from tinyfish-guided-research-mcp tinyfish-research-mcp

Storage

SQLite is fine for local or single-instance use:

export RESEARCH_DB_PATH="research_state.db"

For hosted or multi-instance deployments, use PostgreSQL:

export DATABASE_URL="postgresql://user:password@host:5432/database?sslmode=require"

PostgreSQL is the better option when more than one server instance can access the same research state.

Distribution

The server is available through PyPI, the official MCP Registry, and the hosted Horizon endpoint.

PyPI:

tinyfish-guided-research-mcp

MCP Registry:

io.github.MohdSaleh/tinyfish-guided-research

Hosted MCP:

https://tinyfish-guided-research-mcp.fastmcp.app/mcp

Development

Clone the repo and install the dependencies:

git clone https://github.com/MohdSaleh/tinyfish-guided-research-mcp.git
cd tinyfish-guided-research-mcp
uv sync --all-extras

Run it locally:

TINYFISH_API_KEY="your-api-key" uv run tinyfish-guided-research-mcp

Run the checks:

uv run ruff check .
uv run ruff format --check .
uv run pyright
uv run pytest
uv run python evals/run_evals.py
uv run pip-audit
uv build

The regular tests cover the implementation and storage layer. The research evals cover cases such as duplicate sources, weak evidence, quote mismatches, superseded claims, and citation coverage.

You can also inspect the MCP tools with:

npx @modelcontextprotocol/inspector \
  --cli uv run tinyfish-guided-research-mcp \
  --method tools/list

Deploying on Horizon

If you want to deploy your own instance with Prefect Horizon, use:

Entrypoint:
src/tinyfish_research_mcp/server.py:mcp

Dependencies:
pyproject.toml

Set:

TINYFISH_API_KEY
DATABASE_URL

Use PostgreSQL for hosted deployments instead of the local SQLite fallback.

License

MIT

Available Tools

20 tools
assess_claimsC

Compute deterministic claim resolution and return a complete next-step packet.

Inputs: research_id (required). Returns: data, status, agent_rules, next_action, +2 more

bind_evidenceA

Bind direct source quotations to existing claims and verify quote presence.

This tool does NOT accept semantic stance. Keeping quote binding separate from entailment judgment gives weak client models one bounded operation at a time and prevents a topical passage from becoming support merely because it was attached to a claim.

Inputs: bindings (required), research_id (required). Returns: data, status, agent_rules, next_action, +2 more

check_server_configA

Check API key, DNS, and live TinyFish Search/Fetch endpoint health.

Inputs: no named parameters. Returns: data, status, agent_rules, next_action, +2 more

dispatch_parallel_subagentsB

Execute independent retrieval tracks concurrently and return only a compact review shortlist.

Inputs: research_id (required), tasks. Returns: data, status, agent_rules, next_action, +2 more

finalize_researchC

Finalize into an auditable synthesis manifest; do not generate prose inside the MCP.

Inputs: research_id (required). Returns: data, status, agent_rules, next_action, +2 more

get_research_stateB

Debug/inspection tool. Returns compact protocol state, not full source text.

Inputs: research_id (required). Returns: data, status, agent_rules, next_action, +2 more

get_source_contextA

Return a few relevant passages from persisted full source content.

Use when the client needs a better direct quote without re-fetching the URL.

Inputs: source_id (required), research_id (required), query. Returns: data, status, agent_rules, next_action, +2 more

init_researchB

Start a research session with an explicit latency/coverage profile.

FAST is the default and preserves the same hard evidence gates while fetching/reviewing fewer candidates and stopping resolved claims early.

Inputs: topic (required), mode, contested, max_rounds. Returns: data, status, agent_rules, next_action, +2 more

judge_evidenceB

Assign bounded semantic relation labels to quote-verified evidence.

The server owns evidence IDs and quote integrity. The client does exactly one semantic task here: classify the relationship between the displayed quote and its claim.

Inputs: judgments (required), research_id (required). Returns: data, status, agent_rules, next_action, +2 more

override_atomicityC

Explicitly override an atomicity warning for a maximally atomic claim.

Inputs: reason (required), claim_id (required), research_id (required). Returns: data, status, agent_rules, next_action, +2 more

plan_researchB

Validate and store the research plan before any retrieval.

Inputs: plan (required), research_id (required). Returns: data, status, agent_rules, next_action, +2 more

register_claimsA

Register new claims only; does not accept evidence.

This separation prevents claim creation and evidence judgment from being conflated in one call. Server generates every claim_id.

Inputs: claims (required), research_id (required). Returns: data, status, agent_rules, next_action, +2 more

research_unknownsC

Run targeted retrieval only for claims that still need work, then return one compact review packet.

Inputs: gaps (required), research_id (required). Returns: data, status, agent_rules, next_action, +2 more

review_candidatesA

Unified claim/source review: relevance + direct quote + semantic relation.

Initial broad retrieval still uses screen_sources before claims exist. Once claim IDs exist, this is the preferred path for both initial evidence and targeted gap evidence; bind_evidence/judge_evidence remain low-level compatibility tools rather than the normal happy path.

Inputs: reviews (required), research_id (required). Returns: data, status, agent_rules, next_action, +2 more

review_claim_tensionsA

Record cross-claim tensions after the claim graph is stable.

Mandatory once for contested sessions. Tensions may describe genuine theoretical disagreement even when both claims accurately represent different positions.

Inputs: tensions (required), research_id (required). Returns: data, status, agent_rules, next_action, +2 more

revise_claimA

Replace an incorrect/overbroad active claim without leaving the parent unresolved.

The old claim is superseded automatically. Use this for semantic revision; use split_claim when one claim must become multiple children.

Inputs: claim_id (required), replacement (required), research_id (required). Returns: data, status, agent_rules, next_action, +2 more

screen_sourcesD

Initial semantic source gate over only the server-provided review shortlist.

Inputs: screenings (required), research_id (required). Returns: data, status, agent_rules, next_action, +2 more

split_claimC

Split a compound claim. The parent is superseded automatically.

Inputs: children (required), research_id (required), parent_claim_id (required). Returns: data, status, agent_rules, next_action, +2 more

verify_citationsA

Server-owned citation integrity + coverage audit.

Normal use: pass research_id only. The server automatically selects the strongest quote-verified winning evidence for each strongly resolved claim. Optional citations are accepted only as explicit client overrides.

Inputs: research_id (required), citations. Returns: data, status, agent_rules, next_action, +2 more

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    B3.3/5.0

    Scored across 20 tools

    Disambiguation5/5

    Each tool has a clearly distinct role in the research pipeline: planning, health checks, initialization, parallel retrieval, discovery, screening, claim lifecycle (register, split, revise, override), evidence binding, judging, assessment, gap targeting, unified review, context retrieval, tension recording, citation audit, finalization, and state inspection. Even overlapping tools like review_candidates vs bind_evidence are explicitly distinguished by the description, leaving no ambiguity for an agent.

    Naming Consistency5/5

    All tool names follow a consistent snake_case verb_noun pattern (e.g., plan_research, dispatch_parallel_subagents, verify_citations). The naming convention is uniform and predictable, making it easy for an agent to infer action and target from each name.

    Tool Count4/5

    With 20 tools, the surface is larger than the typical 3–15 range, but the complexity of the guided research workflow—spanning planning, retrieval, claim management, evidence handling, and finalization—justifies the count. Each tool has a specific purpose and the number is proportionate to the domain's needs, though it is on the heavier side.

    Completeness5/5

    The tool set covers the entire research lifecycle: from plan validation and server health checks to parallel retrieval, source screening, claim registration/splitting/revising, evidence binding and judging, claim assessment, targeted gap filling, unified review, tension recording, citation verification, finalization, and state inspection. No obvious dead ends or missing operations are apparent; the surface is comprehensive and well-integrated.

    Maintenance

    ActivityStale
    ResponsivenessNo recent community issues

    Related MCP Connectors

    Related MCP Servers

    • F
      license
      Not graded
      quality
      B
      maintenance
      Enables decomposition of complex research queries into multi-hop sub-queries and synthesis DAGs, with scientific consensus analysis, citation credibility verification, and deterministic JSON outputs for MCP-compliant clients.
      8
      -
    • A
      license
      Not graded
      quality
      B
      maintenance
      Enables traceable research workflows by searching uploaded sources, verifying citations, routing uncertain or high-risk requests to human review, and streaming answers with an audit trail.
      MIT