encode-toolkit 0.3.0b1__py3-none-any.whl

encode_connector/server/main.py

@@ -0,0 +1,1495 @@
+"""ENCODE Project MCP Server.
+
+Exposes ENCODE REST API as Claude-compatible tools for searching experiments,
+listing files, and downloading genomics data.
+
+All data stays local. Only connects to encodeproject.org over HTTPS.
+No telemetry, no analytics, no data sent elsewhere.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import json
+import logging
+from contextlib import asynccontextmanager
+from typing import Any, Literal
+
+from mcp.server.fastmcp import FastMCP
+from mcp.types import ToolAnnotations
+
+from encode_connector.client.auth import CredentialManager
+from encode_connector.client.downloader import FileDownloader
+from encode_connector.client.encode_client import EncodeClient
+from encode_connector.client.models import _human_size
+from encode_connector.client.tracker import (
+    ExperimentTracker,
+    parse_encode_pipelines,
+    parse_encode_publications,
+)
+from encode_connector.client.validation import (
+    clamp_limit,
+    validate_accession,
+    validate_data_export_format,
+    validate_encode_path,
+    validate_export_format,
+    validate_organize_by,
+    validate_reference_type,
+)
+
+logger = logging.getLogger(__name__)
+
+# --- Safety annotations for MCP tools ---
+# Read-only tools that query the ENCODE API (network access, no local writes)
+_READONLY_API = ToolAnnotations(
+    readOnlyHint=True,
+    destructiveHint=False,
+    idempotentHint=True,
+    openWorldHint=True,
+)
+# Read-only tools that read from local tracker DB only (no network)
+_READONLY_LOCAL = ToolAnnotations(
+    readOnlyHint=True,
+    destructiveHint=False,
+    idempotentHint=True,
+    openWorldHint=False,
+)
+# Tools that write to local tracker DB (no network)
+_WRITE_LOCAL = ToolAnnotations(
+    readOnlyHint=False,
+    destructiveHint=False,
+    idempotentHint=True,
+    openWorldHint=False,
+)
+# Tools that write to local tracker DB AND query ENCODE API (network)
+_WRITE_LOCAL_API = ToolAnnotations(
+    readOnlyHint=False,
+    destructiveHint=False,
+    idempotentHint=True,
+    openWorldHint=True,
+)
+# Tools that download files to disk (network + disk writes)
+_DOWNLOAD = ToolAnnotations(
+    readOnlyHint=False,
+    destructiveHint=False,
+    idempotentHint=True,
+    openWorldHint=True,
+)
+# Credential management (can clear credentials = destructive)
+_CREDENTIAL_MGMT = ToolAnnotations(
+    readOnlyHint=False,
+    destructiveHint=True,
+    idempotentHint=True,
+    openWorldHint=False,
+)
+
+# Global client instances (managed via lifespan)
+_client: EncodeClient | None = None
+_downloader: FileDownloader | None = None
+_credential_manager = CredentialManager()
+_tracker: ExperimentTracker | None = None
+_client_lock: asyncio.Lock | None = None
+
+
+def _get_client_lock() -> asyncio.Lock:
+    """Get or create the client lock in the current event loop."""
+    global _client_lock
+    if _client_lock is None:
+        _client_lock = asyncio.Lock()
+    return _client_lock
+
+
+@asynccontextmanager
+async def lifespan(server: FastMCP):
+    """Manage client lifecycle."""
+    global _client, _downloader, _tracker, _client_lock
+    # Initialize lock eagerly in lifespan so it's always bound to the
+    # correct event loop, rather than lazy-creating in _get_client_lock()
+    _client_lock = asyncio.Lock()
+    _client = EncodeClient(credential_manager=_credential_manager)
+    _downloader = FileDownloader(credential_manager=_credential_manager)
+    _tracker = ExperimentTracker()
+    try:
+        yield
+    finally:
+        if _client:
+            await _client.close()
+        if _tracker:
+            _tracker.close()
+
+
+mcp = FastMCP(
+    "ENCODE Project",
+    instructions=(
+        "Query and download genomics data from the ENCODE Project (encodeproject.org). "
+        "Search experiments by assay type, organism, organ, biosample, target, and more. "
+        "List and download files (FASTQ, BAM, BED, bigWig, etc.). "
+        "All data stays local - no telemetry or external data sharing."
+    ),
+    lifespan=lifespan,
+)
+
+
+async def _get_client() -> EncodeClient:
+    async with _get_client_lock():
+        if _client is None:
+            raise RuntimeError("ENCODE client not initialized")
+        return _client
+
+
+def _get_downloader() -> FileDownloader:
+    if _downloader is None:
+        raise RuntimeError("File downloader not initialized")
+    return _downloader
+
+
+def _serialize(obj: Any) -> Any:
+    """Serialize Pydantic models and other objects to JSON-compatible dicts."""
+    if hasattr(obj, "model_dump"):
+        return obj.model_dump()
+    if isinstance(obj, list):
+        return [_serialize(item) for item in obj]
+    if isinstance(obj, dict):
+        return {k: _serialize(v) for k, v in obj.items()}
+    return obj
+
+
+# ======================================================================
+# Tool 1: Search Experiments
+# ======================================================================
+
+
+@mcp.tool(annotations=_READONLY_API, title="Search ENCODE Experiments")
+async def encode_search_experiments(
+    assay_title: str | None = None,
+    organism: str = "Homo sapiens",
+    organ: str | None = None,
+    biosample_type: str | None = None,
+    biosample_term_name: str | None = None,
+    target: str | None = None,
+    status: str = "released",
+    lab: str | None = None,
+    award: str | None = None,
+    assembly: str | None = None,
+    replication_type: str | None = None,
+    life_stage: str | None = None,
+    sex: str | None = None,
+    treatment: str | None = None,
+    genetic_modification: str | None = None,
+    perturbed: bool | None = None,
+    search_term: str | None = None,
+    date_released_from: str | None = None,
+    date_released_to: str | None = None,
+    limit: int = 25,
+    offset: int = 0,
+) -> str:
+    """Search ENCODE experiments with comprehensive filters.
+
+    Examples:
+    - Find all Histone ChIP-seq on human pancreas tissue:
+      assay_title="Histone ChIP-seq", organ="pancreas", biosample_type="tissue"
+    - Find ATAC-seq on human brain:
+      assay_title="ATAC-seq", organ="brain"
+    - Find RNA-seq on GM12878 cell line:
+      assay_title="RNA-seq", biosample_term_name="GM12878"
+    - Find ChIP-seq targeting H3K27me3:
+      assay_title="Histone ChIP-seq", target="H3K27me3"
+    - Find all mouse liver experiments:
+      organism="Mus musculus", organ="liver"
+    - Free text search:
+      search_term="CRISPR screen pancreatic"
+
+    Common assay_title values: "Histone ChIP-seq", "TF ChIP-seq", "ATAC-seq",
+    "DNase-seq", "RNA-seq", "total RNA-seq", "WGBS", "Hi-C", "CUT&RUN",
+    "CUT&Tag", "STARR-seq", "MPRA", "eCLIP", "CRISPR screen"
+
+    Common organ values: "pancreas", "liver", "brain", "heart", "kidney",
+    "lung", "intestine", "skin of body", "blood", "spleen", "thymus"
+
+    biosample_type values: "tissue", "cell line", "primary cell",
+    "in vitro differentiated cells", "organoid"
+
+    WHEN TO USE: Use as the primary entry point when users want to find experiments.
+    Start with encode_get_facets if unsure what filters to use.
+    RELATED TOOLS: encode_get_facets, encode_get_metadata, encode_search_files
+
+    Args:
+        assay_title: Assay type (e.g., "Histone ChIP-seq", "ATAC-seq", "RNA-seq")
+        organism: Species (default: "Homo sapiens"). Also: "Mus musculus"
+        organ: Organ/tissue system (e.g., "pancreas", "brain", "liver")
+        biosample_type: Sample classification ("tissue", "cell line", "primary cell", "organoid")
+        biosample_term_name: Specific cell/tissue name (e.g., "GM12878", "HepG2", "pancreas")
+        target: ChIP/CUT&RUN target (e.g., "H3K27me3", "CTCF", "p300")
+        status: Data status (default: "released"). Also: "archived", "revoked"
+        lab: Submitting lab name
+        award: Funding project
+        assembly: Genome assembly (e.g., "GRCh38", "mm10")
+        replication_type: "isogenic", "anisogenic", or "unreplicated"
+        life_stage: "embryonic", "postnatal", "child", "adult"
+        sex: "male", "female", "mixed"
+        treatment: Treatment name if perturbation experiment
+        genetic_modification: Modification type ("CRISPR", "RNAi")
+        perturbed: True for perturbation experiments only
+        search_term: Free text search across all fields
+        date_released_from: Start date (YYYY-MM-DD) for date range filter
+        date_released_to: End date (YYYY-MM-DD) for date range filter
+        limit: Max results to return (default 25, use larger for comprehensive searches)
+        offset: Skip first N results (for pagination)
+
+    Returns:
+        JSON with experiment results, total count, and pagination info.
+    """
+    client = await _get_client()
+    limit = clamp_limit(limit)
+    result = await client.search_experiments(
+        assay_title=assay_title,
+        organism=organism,
+        organ=organ,
+        biosample_type=biosample_type,
+        biosample_term_name=biosample_term_name,
+        target=target,
+        status=status,
+        lab=lab,
+        award=award,
+        assembly=assembly,
+        replication_type=replication_type,
+        life_stage=life_stage,
+        sex=sex,
+        treatment=treatment,
+        genetic_modification=genetic_modification,
+        perturbed=perturbed,
+        search_term=search_term,
+        date_released_from=date_released_from,
+        date_released_to=date_released_to,
+        limit=limit,
+        offset=offset,
+    )
+    total = result.get("total", 0)
+    serialized = _serialize(result)
+    serialized["has_more"] = total > (offset + limit)
+    serialized["next_offset"] = offset + limit if total > (offset + limit) else None
+    if not result.get("results"):
+        serialized["suggestion"] = (
+            "Try broadening your search filters. Use encode_get_facets to see what data is available for your criteria."
+        )
+    return json.dumps(serialized, indent=2)
+
+
+# ======================================================================
+# Tool 2: Get Experiment Details
+# ======================================================================
+
+
+@mcp.tool(annotations=_READONLY_API, title="Get Experiment Details")
+async def encode_get_experiment(accession: str) -> str:
+    """Get full details for a specific ENCODE experiment by accession ID.
+
+    Returns complete experiment metadata including all associated files,
+    quality metrics, controls, replicate information, and audit status.
+
+    WHEN TO USE: Use when you have a specific accession and need full details
+    including files, quality metrics, and audit status.
+    RELATED TOOLS: encode_list_files, encode_track_experiment, encode_compare_experiments
+
+    Args:
+        accession: ENCODE experiment accession (e.g., "ENCSR133RZO", "ENCSR000AKS")
+
+    Returns:
+        JSON with full experiment details and file listing.
+    """
+    validate_accession(accession)
+    client = await _get_client()
+    result = await client.get_experiment(accession)
+    return json.dumps(_serialize(result), indent=2)
+
+
+# ======================================================================
+# Tool 3: List Files for Experiment
+# ======================================================================
+
+
+@mcp.tool(annotations=_READONLY_API, title="List Experiment Files")
+async def encode_list_files(
+    experiment_accession: str,
+    file_format: str | None = None,
+    file_type: str | None = None,
+    output_type: str | None = None,
+    output_category: str | None = None,
+    assembly: str | None = None,
+    status: str | None = None,
+    preferred_default: bool | None = None,
+    limit: int = 200,
+) -> str:
+    """List all files for a specific ENCODE experiment, with optional filters.
+
+    Examples:
+    - All BED files: experiment_accession="ENCSR133RZO", file_format="bed"
+    - FASTQs only: experiment_accession="ENCSR133RZO", file_format="fastq"
+    - Signal tracks: experiment_accession="ENCSR133RZO", output_category="signal"
+    - Default/recommended files: preferred_default=True
+    - Peaks from GRCh38: file_format="bed", output_type="IDR thresholded peaks", assembly="GRCh38"
+
+    Common file_format values: "fastq", "bam", "bed", "bigWig", "bigBed", "tsv", "hic"
+
+    Common output_type values: "reads", "alignments", "signal of unique reads",
+    "signal of all reads", "fold change over control", "IDR thresholded peaks",
+    "pseudoreplicated peaks", "replicated peaks", "gene quantifications",
+    "transcript quantifications", "contact matrix"
+
+    WHEN TO USE: Use to browse files within a known experiment. Use encode_search_files
+    instead to find files across experiments.
+    RELATED TOOLS: encode_search_files, encode_get_file_info, encode_download_files
+
+    Args:
+        experiment_accession: ENCODE experiment accession (e.g., "ENCSR133RZO")
+        file_format: Filter by format ("fastq", "bam", "bed", "bigWig", "bigBed", etc.)
+        file_type: Filter by specific type ("bed narrowPeak", "bed broadPeak", etc.)
+        output_type: Filter by output type ("reads", "peaks", "signal", etc.)
+        output_category: Filter by category ("raw data", "alignment", "signal", "annotation")
+        assembly: Filter by genome assembly ("GRCh38", "hg19", "mm10")
+        status: Filter by status ("released", "archived", "in progress")
+        preferred_default: If True, return only default/recommended files
+        limit: Max files to return (default 200)
+
+    Returns:
+        JSON list of files with accession, format, size, download URL, and metadata.
+    """
+    validate_accession(experiment_accession)
+    limit = clamp_limit(limit)
+    client = await _get_client()
+    results = await client.list_files(
+        experiment_accession=experiment_accession,
+        file_format=file_format,
+        file_type=file_type,
+        output_type=output_type,
+        output_category=output_category,
+        assembly=assembly,
+        status=status,
+        preferred_default=preferred_default,
+        limit=limit,
+    )
+    return json.dumps(_serialize(results), indent=2)
+
+
+# ======================================================================
+# Tool 4: Search Files Across Experiments
+# ======================================================================
+
+
+@mcp.tool(annotations=_READONLY_API, title="Search Files Across Experiments")
+async def encode_search_files(
+    file_format: str | None = None,
+    file_type: str | None = None,
+    output_type: str | None = None,
+    output_category: str | None = None,
+    assembly: str | None = None,
+    assay_title: str | None = None,
+    organism: str | None = None,
+    organ: str | None = None,
+    biosample_type: str | None = None,
+    target: str | None = None,
+    status: str = "released",
+    preferred_default: bool | None = None,
+    search_term: str | None = None,
+    limit: int = 25,
+    offset: int = 0,
+) -> str:
+    """Search files across ALL experiments with combined experiment + file filters.
+
+    This is powerful for finding specific file types across many experiments.
+
+    Examples:
+    - All BED files from human pancreas ChIP-seq:
+      file_format="bed", assay_title="Histone ChIP-seq", organ="pancreas"
+    - FASTQs from mouse liver RNA-seq:
+      file_format="fastq", assay_title="RNA-seq", organ="liver", organism="Mus musculus"
+    - All IDR peak files for H3K27me3:
+      output_type="IDR thresholded peaks", target="H3K27me3"
+    - BigWig signal tracks from ATAC-seq on brain tissue:
+      file_format="bigWig", assay_title="ATAC-seq", organ="brain", biosample_type="tissue"
+
+    WHEN TO USE: Use to find specific file types across ALL experiments. More powerful
+    than encode_list_files for cross-experiment file discovery.
+    RELATED TOOLS: encode_list_files, encode_batch_download, encode_get_file_info
+
+    Args:
+        file_format: File format ("fastq", "bam", "bed", "bigWig", etc.)
+        file_type: Specific file type ("bed narrowPeak", "bed broadPeak", etc.)
+        output_type: Output type ("reads", "peaks", "signal", etc.)
+        output_category: Output category ("raw data", "alignment", "signal", "annotation")
+        assembly: Genome assembly ("GRCh38", "hg19", "mm10")
+        assay_title: Filter by assay type of parent experiment
+        organism: Filter by organism of parent experiment
+        organ: Filter by organ of parent experiment
+        biosample_type: Filter by biosample type ("tissue", "cell line", etc.)
+        target: Filter by ChIP/CUT&RUN target
+        status: File status (default: "released")
+        preferred_default: If True, only default/recommended files
+        search_term: Free text search
+        limit: Max results (default 25)
+        offset: Skip first N results (pagination)
+
+    Returns:
+        JSON with file results, total count, and pagination info.
+    """
+    client = await _get_client()
+    limit = clamp_limit(limit)
+    result = await client.search_files(
+        file_format=file_format,
+        file_type=file_type,
+        output_type=output_type,
+        output_category=output_category,
+        assembly=assembly,
+        assay_title=assay_title,
+        organism=organism,
+        organ=organ,
+        biosample_type=biosample_type,
+        target=target,
+        status=status,
+        preferred_default=preferred_default,
+        search_term=search_term,
+        limit=limit,
+        offset=offset,
+    )
+    total = result.get("total", 0)
+    serialized = _serialize(result)
+    serialized["has_more"] = total > (offset + limit)
+    serialized["next_offset"] = offset + limit if total > (offset + limit) else None
+    if not result.get("results"):
+        serialized["suggestion"] = (
+            "Verify assembly and file_format values. Use encode_get_metadata('file_formats') to see valid options."
+        )
+    return json.dumps(serialized, indent=2)
+
+
+# ======================================================================
+# Tool 5: Download Files
+# ======================================================================
+
+
+@mcp.tool(annotations=_DOWNLOAD, title="Download ENCODE Files")
+async def encode_download_files(
+    file_accessions: list[str],
+    download_dir: str,
+    organize_by: Literal["flat", "experiment", "format", "experiment_format"] = "flat",
+    verify_md5: bool = True,
+) -> str:
+    """Download specific ENCODE files by accession to a local directory.
+
+    Downloads files from ENCODE to your local machine. Supports MD5 verification,
+    concurrent downloads, and skip-if-already-downloaded.
+
+    WHEN TO USE: Use for downloading specific files by accession. For bulk downloads,
+    prefer encode_batch_download.
+    RELATED TOOLS: encode_batch_download, encode_search_files, encode_log_derived_file
+
+    Args:
+        file_accessions: List of file accessions to download (e.g., ["ENCFF635JIA", "ENCFF388RZD"])
+        download_dir: Local directory path to save files (e.g., "/Users/you/data/encode")
+        organize_by: How to organize downloaded files:
+            - "flat": All files in download_dir (default)
+            - "experiment": download_dir/ENCSR.../filename
+            - "format": download_dir/bed/filename
+            - "experiment_format": download_dir/ENCSR.../bed/filename
+        verify_md5: Verify file integrity with MD5 checksum (default True)
+
+    Returns:
+        JSON with download results for each file (path, size, success/error, MD5 status).
+    """
+    validate_organize_by(organize_by)
+    for acc in file_accessions:
+        validate_accession(acc)
+
+    client = await _get_client()
+    downloader = _get_downloader()
+
+    # Get file info for each accession
+    file_infos = []
+    errors = []
+    for acc in file_accessions:
+        try:
+            info = await client.get_file_info(acc)
+            file_infos.append(info)
+        except Exception as e:
+            errors.append({"accession": acc, "error": str(e)})
+
+    # Download all files
+    results = await downloader.download_batch(file_infos, download_dir, organize_by, verify_md5)
+
+    output = {
+        "downloaded": _serialize(results),
+        "errors": errors,
+        "summary": {
+            "total_requested": len(file_accessions),
+            "successful": sum(1 for r in results if r.success),
+            "failed": sum(1 for r in results if not r.success) + len(errors),
+            "total_size": sum(r.file_size for r in results if r.success),
+            "total_size_human": _human_size(sum(r.file_size for r in results if r.success)),
+        },
+    }
+    return json.dumps(output, indent=2)
+
+
+# ======================================================================
+# Tool 6: Get Metadata / Filter Values
+# ======================================================================
+
+
+@mcp.tool(annotations=_READONLY_API, title="Get Filter Values")
+async def encode_get_metadata(
+    metadata_type: Literal[
+        "assays",
+        "organisms",
+        "organs",
+        "biosample_types",
+        "file_formats",
+        "output_types",
+        "output_categories",
+        "assemblies",
+        "life_stages",
+        "replication_types",
+        "statuses",
+        "file_statuses",
+    ],
+) -> str:
+    """Get available filter values for ENCODE searches.
+
+    Use this to discover valid values for search parameters.
+
+    WHEN TO USE: Use to discover valid filter values before searching. Helps prevent
+    typos in assay_title, organ, biosample_type etc.
+    RELATED TOOLS: encode_get_facets, encode_search_experiments
+
+    Args:
+        metadata_type: Type of metadata to retrieve. Options:
+            - "assays": Available assay types (Histone ChIP-seq, ATAC-seq, RNA-seq, etc.)
+            - "organisms": Available organisms (Homo sapiens, Mus musculus, etc.)
+            - "organs": Available organ/tissue systems (pancreas, brain, liver, etc.)
+            - "biosample_types": Biosample classifications (tissue, cell line, primary cell, etc.)
+            - "file_formats": File format types (fastq, bam, bed, bigWig, etc.)
+            - "output_types": Output data types (reads, peaks, signal, etc.)
+            - "output_categories": Output categories (raw data, alignment, signal, etc.)
+            - "assemblies": Genome assemblies (GRCh38, hg19, mm10, etc.)
+            - "life_stages": Life stages (embryonic, adult, child, etc.)
+            - "replication_types": Replication types (isogenic, anisogenic, unreplicated)
+            - "statuses": Experiment statuses (released, archived, etc.)
+            - "file_statuses": File statuses (released, archived, in progress, etc.)
+
+    Returns:
+        JSON list of valid values for the specified metadata type.
+    """
+    client = await _get_client()
+    try:
+        values = client.get_metadata(metadata_type)
+        return json.dumps({"metadata_type": metadata_type, "values": values, "count": len(values)}, indent=2)
+    except ValueError as e:
+        return json.dumps({"error": str(e)}, indent=2)
+
+
+# ======================================================================
+# Tool 7: Batch Download from Search
+# ======================================================================
+
+
+@mcp.tool(annotations=_DOWNLOAD, title="Batch Search and Download")
+async def encode_batch_download(
+    download_dir: str,
+    file_format: str | None = None,
+    output_type: str | None = None,
+    output_category: str | None = None,
+    assembly: str | None = None,
+    assay_title: str | None = None,
+    organism: str = "Homo sapiens",
+    organ: str | None = None,
+    biosample_type: str | None = None,
+    target: str | None = None,
+    preferred_default: bool | None = None,
+    organize_by: Literal["flat", "experiment", "format", "experiment_format"] = "experiment",
+    verify_md5: bool = True,
+    limit: int = 100,
+    dry_run: bool = True,
+) -> str:
+    """Search for files and download them all in batch.
+
+    First searches for files matching the criteria, then downloads them.
+    By default runs in dry_run mode to preview what would be downloaded.
+    Set dry_run=False to actually download.
+
+    WHEN TO USE: Use for searching and downloading files in one step. Always use
+    dry_run=True first to preview. For specific file accessions, use encode_download_files.
+    RELATED TOOLS: encode_download_files, encode_search_files
+
+    Examples:
+    - Download all BED files from human pancreas ChIP-seq:
+      file_format="bed", assay_title="Histone ChIP-seq", organ="pancreas",
+      download_dir="/data/encode", dry_run=False
+    - Preview FASTQ downloads for mouse brain RNA-seq:
+      file_format="fastq", assay_title="RNA-seq", organ="brain",
+      organism="Mus musculus", download_dir="/data/encode"
+    - Download IDR peaks for H3K27me3 in GRCh38:
+      output_type="IDR thresholded peaks", target="H3K27me3", assembly="GRCh38",
+      download_dir="/data/encode", dry_run=False
+
+    Args:
+        download_dir: Local directory to save files
+        file_format: File format filter ("fastq", "bam", "bed", "bigWig", etc.)
+        output_type: Output type filter ("reads", "peaks", "signal", etc.)
+        output_category: Output category ("raw data", "alignment", "annotation", etc.)
+        assembly: Genome assembly ("GRCh38", "mm10", etc.)
+        assay_title: Assay type ("Histone ChIP-seq", "ATAC-seq", "RNA-seq", etc.)
+        organism: Organism (default: "Homo sapiens")
+        organ: Organ/tissue ("pancreas", "brain", "liver", etc.)
+        biosample_type: Biosample type ("tissue", "cell line", "primary cell", etc.)
+        target: ChIP/CUT&RUN target ("H3K27me3", "CTCF", etc.)
+        preferred_default: If True, only download default/recommended files
+        organize_by: File organization ("flat", "experiment", "format", "experiment_format")
+        verify_md5: Verify downloads with MD5 checksums (default True)
+        limit: Max files to download (default 100, safety limit)
+        dry_run: If True (default), only preview what would be downloaded. Set False to download.
+
+    Returns:
+        JSON with download preview (dry_run=True) or download results (dry_run=False).
+    """
+    client = await _get_client()
+    downloader = _get_downloader()
+    validate_organize_by(organize_by)
+    limit = clamp_limit(limit)
+
+    # Search for files
+    search_result = await client.search_files(
+        file_format=file_format,
+        output_type=output_type,
+        output_category=output_category,
+        assembly=assembly,
+        assay_title=assay_title,
+        organism=organism,
+        organ=organ,
+        biosample_type=biosample_type,
+        target=target,
+        status="released",
+        preferred_default=preferred_default,
+        limit=limit,
+    )
+
+    files = search_result["results"]
+
+    if not files:
+        return json.dumps(
+            {
+                "message": "No files found matching the search criteria.",
+                "total": 0,
+                "has_more": False,
+                "next_offset": None,
+                "suggestion": "Try broadening your search filters. Use encode_get_facets to see what data is available for your criteria.",
+            },
+            indent=2,
+        )
+
+    if dry_run:
+        # Preview mode
+        search_total = search_result["total"]
+        preview = downloader.preview_downloads(files, download_dir, organize_by)
+        preview["message"] = (
+            f"Found {preview['file_count']} files ({preview['total_size_human']}). Set dry_run=False to download."
+        )
+        preview["search_total"] = search_total
+        preview["has_more"] = search_total > limit
+        preview["next_offset"] = limit if search_total > limit else None
+        return json.dumps(_serialize(preview), indent=2)
+
+    # Actually download
+    results = await downloader.download_batch(files, download_dir, organize_by, verify_md5)
+    search_total = search_result["total"]
+
+    output = {
+        "downloaded": _serialize(results),
+        "summary": {
+            "total_found": search_total,
+            "total_downloaded": len(results),
+            "successful": sum(1 for r in results if r.success),
+            "failed": sum(1 for r in results if not r.success),
+            "total_size": sum(r.file_size for r in results if r.success),
+            "total_size_human": _human_size(sum(r.file_size for r in results if r.success)),
+        },
+        "has_more": search_total > limit,
+        "next_offset": limit if search_total > limit else None,
+    }
+    return json.dumps(output, indent=2)
+
+
+# ======================================================================
+# Tool 8: Store/Manage Credentials
+# ======================================================================
+
+
+@mcp.tool(annotations=_CREDENTIAL_MGMT, title="Manage API Credentials")
+async def encode_manage_credentials(
+    action: Literal["store", "check", "clear"],
+    access_key: str | None = None,
+    secret_key: str | None = None,
+) -> str:
+    """Manage ENCODE API credentials for accessing restricted/unreleased data.
+
+    Most ENCODE data is public and requires no authentication.
+    Credentials are only needed for unreleased or restricted datasets.
+
+    Credentials are stored securely in your OS keyring (macOS Keychain,
+    Linux Secret Service, Windows Credential Locker) and never in plaintext.
+
+    WHEN TO USE: Use only for accessing unreleased/restricted ENCODE data.
+    Public data requires no authentication.
+    RELATED TOOLS: encode_search_experiments
+
+    Args:
+        action: What to do:
+            - "store": Save new credentials (requires access_key and secret_key)
+            - "check": Check if credentials are configured
+            - "clear": Remove stored credentials
+        access_key: Your ENCODE access key (only for action="store")
+        secret_key: Your ENCODE secret key (only for action="store")
+
+    Returns:
+        JSON with action result.
+    """
+    if action == "store":
+        if not access_key or not secret_key:
+            return json.dumps(
+                {
+                    "error": "Both access_key and secret_key are required to store credentials.",
+                    "help": "Get your access key pair from your ENCODE profile at https://www.encodeproject.org/",
+                },
+                indent=2,
+            )
+
+        location = _credential_manager.store_credentials(access_key, secret_key)
+        # Reset client to pick up new credentials
+        global _client
+        async with _get_client_lock():
+            if _client:
+                await _client.close()
+            _client = EncodeClient(credential_manager=_credential_manager)
+
+        return json.dumps(
+            {
+                "success": True,
+                "message": f"Credentials stored securely in: {location}",
+                "note": "Credentials are encrypted and never stored in plaintext.",
+            },
+            indent=2,
+        )
+
+    elif action == "check":
+        has_creds = _credential_manager.has_credentials
+        return json.dumps(
+            {
+                "credentials_configured": has_creds,
+                "message": (
+                    "Credentials are configured. You can access restricted data."
+                    if has_creds
+                    else "No credentials configured. You can still access all public ENCODE data. "
+                    "Use action='store' with your ENCODE access key pair to access restricted data."
+                ),
+            },
+            indent=2,
+        )
+
+    elif action == "clear":
+        _credential_manager.clear_credentials()
+        # Reset client
+        async with _get_client_lock():
+            if _client:
+                await _client.close()
+            _client = EncodeClient(credential_manager=_credential_manager)
+
+        return json.dumps(
+            {
+                "success": True,
+                "message": "All stored credentials have been removed.",
+            },
+            indent=2,
+        )
+
+    else:
+        return json.dumps(
+            {
+                "error": f"Unknown action: {action}. Use 'store', 'check', or 'clear'.",
+            },
+            indent=2,
+        )
+
+
+# ======================================================================
+# Tool 9: Get Live Facets (Dynamic Filter Discovery)
+# ======================================================================
+
+
+@mcp.tool(annotations=_READONLY_API, title="Explore Available Data")
+async def encode_get_facets(
+    search_type: str = "Experiment",
+    assay_title: str | None = None,
+    organism: str | None = None,
+    organ: str | None = None,
+    biosample_type: str | None = None,
+) -> str:
+    """Get live filter counts from ENCODE to discover what data is available.
+
+    Returns faceted counts showing how many experiments/files exist for each
+    filter value. Useful for exploring what's available before searching.
+
+    WHEN TO USE: Use to explore what data exists before searching. Shows counts
+    per filter value. Best first step for unknown datasets.
+    RELATED TOOLS: encode_get_metadata, encode_search_experiments
+
+    Examples:
+    - What assays are available for pancreas?
+      organ="pancreas"
+    - What organs have Histone ChIP-seq data?
+      assay_title="Histone ChIP-seq"
+    - What targets are available for mouse brain ChIP-seq?
+      assay_title="Histone ChIP-seq", organism="Mus musculus", organ="brain"
+
+    Args:
+        search_type: Object type ("Experiment" or "File")
+        assay_title: Pre-filter by assay type
+        organism: Pre-filter by organism
+        organ: Pre-filter by organ
+        biosample_type: Pre-filter by biosample type
+
+    Returns:
+        JSON with facet names and their term counts.
+    """
+    client = await _get_client()
+    filters = {}
+    if assay_title:
+        filters["assay_title"] = assay_title
+    if organism:
+        filters["replicates.library.biosample.donor.organism.scientific_name"] = organism
+    if organ:
+        filters["biosample_ontology.organ_slims"] = organ
+    if biosample_type:
+        filters["biosample_ontology.classification"] = biosample_type
+
+    facets = await client.search_facets(search_type=search_type, **filters)
+
+    # Simplify output - show most useful facets
+    useful_facets = {}
+    for field, terms in facets.items():
+        # Only include facets with reasonable number of terms
+        if len(terms) <= 200:
+            useful_facets[field] = terms[:50]  # Cap at 50 terms per facet
+
+    return json.dumps(useful_facets, indent=2)
+
+
+# ======================================================================
+# Tool 10: Get File Info
+# ======================================================================
+
+
+@mcp.tool(annotations=_READONLY_API, title="Get File Details")
+async def encode_get_file_info(accession: str) -> str:
+    """Get detailed information about a specific ENCODE file.
+
+    WHEN TO USE: Use when you need detailed metadata for a specific file
+    (size, md5, assembly, biological replicate info).
+    RELATED TOOLS: encode_download_files, encode_list_files
+
+    Args:
+        accession: File accession ID (e.g., "ENCFF635JIA")
+
+    Returns:
+        JSON with file metadata including format, size, download URL, MD5, assembly, etc.
+    """
+    validate_accession(accession)
+    client = await _get_client()
+    info = await client.get_file_info(accession)
+    return json.dumps(_serialize(info), indent=2)
+
+
+# ======================================================================
+# Tracker helper
+# ======================================================================
+
+
+def _get_tracker() -> ExperimentTracker:
+    if _tracker is None:
+        raise RuntimeError("Experiment tracker not initialized")
+    return _tracker
+
+
+# ======================================================================
+# Tool 11: Track Experiment
+# ======================================================================
+
+
+@mcp.tool(annotations=_WRITE_LOCAL_API, title="Track Experiment Locally")
+async def encode_track_experiment(
+    accession: str,
+    fetch_publications: bool = True,
+    fetch_pipelines: bool = True,
+    notes: str = "",
+) -> str:
+    """Track an ENCODE experiment locally with its publications, methods, and pipeline info.
+
+    Fetches full experiment metadata from ENCODE and stores it in a local SQLite
+    database along with any associated publications (PMIDs, DOIs, authors, journal)
+    and pipeline/analysis information (software versions, methods).
+
+    This is like adding an experiment to your "library" - similar to Endnote for papers.
+
+    WHEN TO USE: Use to save an experiment to your local library with publications
+    and pipeline info. Required before compare or citations.
+    RELATED TOOLS: encode_compare_experiments, encode_get_citations, encode_export_data
+
+    Args:
+        accession: ENCODE experiment accession (e.g., "ENCSR133RZO")
+        fetch_publications: Also fetch and store publications/citations (default True)
+        fetch_pipelines: Also fetch and store pipeline/analysis info (default True)
+        notes: Optional notes to attach to this experiment
+
+    Returns:
+        JSON with tracking result including publications and pipeline info found.
+    """
+    client = await _get_client()
+    tracker = _get_tracker()
+    validate_accession(accession)
+
+    # Get full experiment data from ENCODE
+    exp_data = await client.get_experiment_raw(accession)
+
+    # Track the experiment
+    from encode_connector.client.models import ExperimentDetail
+
+    detail = ExperimentDetail.from_api(exp_data)
+    result = tracker.track_experiment(detail.model_dump(), raw_metadata=exp_data)
+
+    if notes:
+        tracker.add_note(accession, notes)
+
+    output: dict[str, Any] = {"tracking": result}
+
+    # Auto-link cross-references from dbxrefs (GEO, PMID, etc.)
+    dbxrefs = exp_data.get("dbxrefs", []) or []
+    auto_linked = []
+    for xref in dbxrefs:
+        if not isinstance(xref, str):
+            continue
+        if xref.startswith("GEO:"):
+            ref_result = tracker.link_reference(
+                accession,
+                "geo_accession",
+                xref.replace("GEO:", ""),
+                "Auto-extracted from ENCODE dbxrefs",
+            )
+            if ref_result.get("action") == "linked":
+                auto_linked.append({"type": "geo_accession", "id": xref.replace("GEO:", "")})
+        elif xref.startswith("PMID:"):
+            ref_result = tracker.link_reference(
+                accession,
+                "pmid",
+                xref.replace("PMID:", ""),
+                "Auto-extracted from ENCODE dbxrefs",
+            )
+            if ref_result.get("action") == "linked":
+                auto_linked.append({"type": "pmid", "id": xref.replace("PMID:", "")})
+    if auto_linked:
+        output["auto_linked_references"] = auto_linked
+
+    # Fetch publications
+    if fetch_publications:
+        refs = exp_data.get("references", [])
+        # If references are paths, fetch them (validate to prevent SSRF)
+        resolved_refs = []
+        for ref in refs:
+            if isinstance(ref, str):
+                try:
+                    safe_path = validate_encode_path(ref)
+                    ref_data = await client.get_json(safe_path, {"format": "json"})
+                    resolved_refs.append(ref_data)
+                except Exception as e:
+                    logger.warning("Skipping ref %r: %s", ref, e)
+            elif isinstance(ref, dict):
+                resolved_refs.append(ref)
+
+        pubs = parse_encode_publications(resolved_refs)
+        pub_count = tracker.store_publications(accession, pubs)
+        output["publications_found"] = pub_count
+        output["publications"] = pubs
+
+    # Fetch pipeline info
+    if fetch_pipelines:
+        analyses = exp_data.get("analyses", [])
+        resolved_analyses = []
+        for analysis in analyses:
+            if isinstance(analysis, str):
+                try:
+                    safe_path = validate_encode_path(analysis)
+                    a_data = await client.get_json(safe_path, {"format": "json"})
+                    resolved_analyses.append(a_data)
+                except Exception as e:
+                    logger.warning("Skipping ref %r: %s", analysis, e)
+            elif isinstance(analysis, dict):
+                resolved_analyses.append(analysis)
+
+        pipelines = parse_encode_pipelines(resolved_analyses)
+        pipe_count = tracker.store_pipeline_info(accession, pipelines)
+        output["pipelines_found"] = pipe_count
+        output["pipelines"] = pipelines
+
+    return json.dumps(output, indent=2)
+
+
+# ======================================================================
+# Tool 12: List Tracked Experiments
+# ======================================================================
+
+
+@mcp.tool(annotations=_READONLY_LOCAL, title="List Tracked Experiments")
+async def encode_list_tracked(
+    assay_title: str | None = None,
+    organism: str | None = None,
+    organ: str | None = None,
+) -> str:
+    """List all experiments you've tracked locally, with optional filters.
+
+    Shows your local library of tracked ENCODE experiments, their metadata,
+    publication counts, and derived file counts.
+
+    WHEN TO USE: Use to see all experiments in your local library. Filter by assay,
+    organism, or organ.
+    RELATED TOOLS: encode_summarize_collection, encode_export_data
+
+    Args:
+        assay_title: Filter by assay type (partial match)
+        organism: Filter by organism (partial match)
+        organ: Filter by organ (partial match)
+
+    Returns:
+        JSON with tracked experiments metadata table and tracker stats.
+    """
+    tracker = _get_tracker()
+    experiments = tracker.list_tracked_experiments(
+        assay_title=assay_title,
+        organism=organism,
+        organ=organ,
+    )
+
+    # Build metadata table
+    table = tracker.get_metadata_table([e["accession"] for e in experiments] if experiments else None)
+
+    # Remove raw_metadata from output
+    for row in table:
+        row.pop("raw_metadata", None)
+
+    return json.dumps(
+        {
+            "experiments": table,
+            "count": len(table),
+            "stats": tracker.stats,
+        },
+        indent=2,
+    )
+
+
+# ======================================================================
+# Tool 13: Get Experiment Publications & Citations
+# ======================================================================
+
+
+@mcp.tool(annotations=_READONLY_LOCAL, title="Get Citations")
+async def encode_get_citations(
+    accession: str | None = None,
+    export_format: Literal["json", "bibtex", "ris"] = "json",
+) -> str:
+    """Get publications and citations for tracked experiments.
+
+    Returns publication data with authors, journal, DOI, PMID.
+    Can export as BibTeX or RIS (Endnote/Zotero/Mendeley compatible).
+
+    WHEN TO USE: Use to get publication data for tracked experiments. Supports
+    BibTeX and RIS export for reference managers.
+    RELATED TOOLS: encode_track_experiment, encode_link_reference
+
+    Args:
+        accession: Specific experiment accession. If None, returns all publications.
+        export_format: Output format:
+            - "json": Structured data (default)
+            - "bibtex": BibTeX format for LaTeX
+            - "ris": RIS format (Endnote, Zotero, Mendeley)
+
+    Returns:
+        Publications in the requested format.
+    """
+    validate_export_format(export_format)
+    if accession:
+        validate_accession(accession)
+    tracker = _get_tracker()
+
+    if export_format == "bibtex":
+        bibtex = tracker.export_citations_bibtex([accession] if accession else None)
+        return bibtex if bibtex else "No publications found."
+
+    if export_format == "ris":
+        ris = tracker.export_citations_ris([accession] if accession else None)
+        return ris if ris else "No publications found."
+
+    # JSON format
+    if accession:
+        pubs = tracker.get_publications(accession)
+    else:
+        # Get all from all tracked experiments
+        experiments = tracker.list_tracked_experiments()
+        pubs = []
+        for exp in experiments:
+            exp_pubs = tracker.get_publications(exp["accession"])
+            pubs.extend(exp_pubs)
+
+    return json.dumps(
+        {
+            "publications": pubs,
+            "count": len(pubs),
+        },
+        indent=2,
+        default=str,
+    )
+
+
+# ======================================================================
+# Tool 14: Compare Experiments (Compatibility Analysis)
+# ======================================================================
+
+
+@mcp.tool(annotations=_READONLY_LOCAL, title="Compare Experiments")
+async def encode_compare_experiments(
+    accession1: str,
+    accession2: str,
+) -> str:
+    """Analyze whether two ENCODE experiments are compatible for combined analysis.
+
+    Compares organism, genome assembly, assay type, biosample, organ, target,
+    replication strategy, and lab to identify potential issues.
+
+    Both experiments must be tracked first (use encode_track_experiment).
+
+    WHEN TO USE: Use to check if two experiments are compatible for combined analysis.
+    Both must be tracked first.
+    RELATED TOOLS: encode_track_experiment, encode_list_tracked
+
+    Args:
+        accession1: First experiment accession (e.g., "ENCSR133RZO")
+        accession2: Second experiment accession (e.g., "ENCSR000AKS")
+
+    Returns:
+        JSON compatibility report with verdict, issues, warnings, and recommendations.
+    """
+    validate_accession(accession1)
+    validate_accession(accession2)
+    tracker = _get_tracker()
+    result = tracker.analyze_compatibility(accession1, accession2)
+    return json.dumps(result, indent=2)
+
+
+# ======================================================================
+# Tool 15: Log Derived File (Provenance)
+# ======================================================================
+
+
+@mcp.tool(annotations=_WRITE_LOCAL, title="Log Derived File")
+async def encode_log_derived_file(
+    file_path: str,
+    source_accessions: list[str],
+    description: str = "",
+    file_type: str = "",
+    tool_used: str = "",
+    parameters: str = "",
+) -> str:
+    """Log a file you've derived from ENCODE data for provenance tracking.
+
+    Use this when you create new files from ENCODE data (e.g., running a pipeline,
+    filtering peaks, merging samples). This creates a provenance record linking
+    your derived file back to the original ENCODE source data.
+
+    WHEN TO USE: Use after creating files from ENCODE data (filtered peaks, merged
+    signals). Creates provenance chain back to source.
+    RELATED TOOLS: encode_get_provenance, encode_download_files
+
+    Args:
+        file_path: Path to the derived file you created
+        source_accessions: List of ENCODE accessions this file was derived from
+            (experiment or file accessions, e.g., ["ENCSR133RZO", "ENCFF635JIA"])
+        description: What this derived file contains
+        file_type: Type of file (e.g., "filtered_peaks", "merged_signal", "differential")
+        tool_used: Tool/software used to create it (e.g., "bedtools intersect", "DESeq2")
+        parameters: Parameters or command used
+
+    Returns:
+        JSON with the provenance record ID.
+    """
+    # Validate each source accession is a valid ENCODE identifier
+    for acc in source_accessions:
+        validate_accession(acc)
+
+    tracker = _get_tracker()
+    row_id = tracker.log_derived_file(
+        file_path=file_path,
+        source_accessions=source_accessions,
+        description=description,
+        file_type=file_type,
+        tool_used=tool_used,
+        parameters=parameters,
+    )
+    return json.dumps(
+        {
+            "success": True,
+            "record_id": row_id,
+            "file_path": file_path,
+            "source_accessions": source_accessions,
+            "message": "Provenance logged. Use encode_get_provenance to view the full chain.",
+        },
+        indent=2,
+    )
+
+
+# ======================================================================
+# Tool 16: Get Provenance
+# ======================================================================
+
+
+@mcp.tool(annotations=_READONLY_LOCAL, title="Get File Provenance")
+async def encode_get_provenance(
+    file_path: str | None = None,
+    source_accession: str | None = None,
+) -> str:
+    """Get provenance information for derived files.
+
+    Shows the chain from your derived files back to original ENCODE data,
+    including what tools and parameters were used.
+
+    WHEN TO USE: Use to trace a derived file back to original ENCODE data.
+    Shows tools and parameters used.
+    RELATED TOOLS: encode_log_derived_file
+
+    Args:
+        file_path: Get provenance for a specific derived file
+        source_accession: List all files derived from a specific ENCODE accession
+
+    Returns:
+        JSON provenance chain or list of derived files.
+    """
+    tracker = _get_tracker()
+
+    if file_path:
+        chain = tracker.get_provenance_chain(file_path)
+        return json.dumps(chain, indent=2, default=str)
+
+    derived = tracker.get_derived_files(source_accession)
+    return json.dumps(
+        {
+            "derived_files": derived,
+            "count": len(derived),
+        },
+        indent=2,
+        default=str,
+    )
+
+
+# ======================================================================
+# Tool 17: Export Tracked Data
+# ======================================================================
+
+
+@mcp.tool(annotations=_READONLY_LOCAL, title="Export Tracked Data")
+async def encode_export_data(
+    format: Literal["csv", "tsv", "json"] = "csv",
+    assay_title: str | None = None,
+    organism: str | None = None,
+    organ: str | None = None,
+) -> str:
+    """Export tracked experiments as a table (CSV, TSV, or JSON).
+
+    Creates a tabular export of all tracked experiments with metadata,
+    publication counts, PMIDs, and derived file counts. Useful for loading
+    into Excel, R, pandas, or sharing with collaborators.
+
+    PMIDs in the output can be directly used with PubMed MCP tools for
+    further literature analysis.
+
+    WHEN TO USE: Use to create shareable tables of tracked experiments (CSV, TSV, JSON).
+    Good for manuscripts and reports.
+    RELATED TOOLS: encode_list_tracked, encode_summarize_collection
+
+    Args:
+        format: Output format:
+            - "csv": Comma-separated values (default, for Excel/spreadsheets)
+            - "tsv": Tab-separated values (for R, pandas)
+            - "json": JSON array (for programmatic use)
+        assay_title: Filter by assay type (partial match)
+        organism: Filter by organism (partial match)
+        organ: Filter by organ (partial match)
+
+    Returns:
+        Formatted table data in the requested format.
+    """
+    validate_data_export_format(format)
+    tracker = _get_tracker()
+    result = tracker.export_tracked_data(
+        format=format,
+        assay_title=assay_title,
+        organism=organism,
+        organ=organ,
+    )
+    if not result:
+        return json.dumps({"message": "No tracked experiments found matching filters."})
+    return result
+
+
+# ======================================================================
+# Tool 18: Summarize Collection
+# ======================================================================
+
+
+@mcp.tool(annotations=_READONLY_LOCAL, title="Summarize Collection")
+async def encode_summarize_collection(
+    assay_title: str | None = None,
+    organism: str | None = None,
+    organ: str | None = None,
+) -> str:
+    """Summarize your tracked experiment collection with grouped statistics.
+
+    Provides an overview of tracked experiments grouped by assay type, target,
+    organism, organ, biosample type, and lab. Shows total counts for publications,
+    derived files, and external references.
+
+    Useful when tracking 10+ experiments and needing a bird's-eye view of your
+    research data collection.
+
+    WHEN TO USE: Use for a bird's-eye view of tracked experiments grouped by assay,
+    target, organ. Best for 10+ tracked experiments.
+    RELATED TOOLS: encode_list_tracked, encode_export_data
+
+    Args:
+        assay_title: Filter by assay type (partial match)
+        organism: Filter by organism (partial match)
+        organ: Filter by organ (partial match)
+
+    Returns:
+        JSON summary with experiment counts grouped by multiple dimensions.
+    """
+    tracker = _get_tracker()
+    summary = tracker.summarize_collection(
+        assay_title=assay_title,
+        organism=organism,
+        organ=organ,
+    )
+    return json.dumps(summary, indent=2)
+
+
+# ======================================================================
+# Tool 19: Link External Reference
+# ======================================================================
+
+
+@mcp.tool(annotations=_WRITE_LOCAL, title="Link External Reference")
+async def encode_link_reference(
+    experiment_accession: str,
+    reference_type: Literal["pmid", "doi", "nct_id", "preprint_doi", "geo_accession", "other"],
+    reference_id: str,
+    description: str = "",
+) -> str:
+    """Link an external reference to a tracked ENCODE experiment.
+
+    This is the cross-server bridge. Attach PubMed IDs, bioRxiv DOIs,
+    ClinicalTrials.gov NCT IDs, GEO accessions, or any external identifier
+    to your tracked experiments for provenance and cross-referencing.
+
+    After finding a relevant paper with PubMed MCP or a preprint on bioRxiv,
+    link it to the ENCODE experiment for future reference.
+
+    WHEN TO USE: Use to attach external IDs (PMID, DOI, GEO, NCT) to tracked
+    experiments for cross-referencing.
+    RELATED TOOLS: encode_get_references, encode_get_citations
+
+    Args:
+        experiment_accession: ENCODE experiment accession (e.g., "ENCSR133RZO")
+        reference_type: Type of external reference:
+            - "pmid": PubMed ID (e.g., "32728249")
+            - "doi": DOI (e.g., "10.1038/s41586-020-2493-4")
+            - "nct_id": ClinicalTrials.gov ID (e.g., "NCT04567890")
+            - "preprint_doi": bioRxiv/medRxiv DOI
+            - "geo_accession": GEO accession (e.g., "GSE123456")
+            - "other": Any other identifier
+        reference_id: The actual identifier value
+        description: Optional description of why this reference is linked
+
+    Returns:
+        JSON with linking result.
+    """
+    validate_accession(experiment_accession)
+    validate_reference_type(reference_type)
+    tracker = _get_tracker()
+    result = tracker.link_reference(
+        accession=experiment_accession,
+        reference_type=reference_type,
+        reference_id=reference_id,
+        description=description,
+    )
+    return json.dumps(result, indent=2)
+
+
+# ======================================================================
+# Tool 20: Get External References
+# ======================================================================
+
+
+@mcp.tool(annotations=_READONLY_LOCAL, title="Get Linked References")
+async def encode_get_references(
+    experiment_accession: str | None = None,
+    reference_type: Literal["pmid", "doi", "nct_id", "preprint_doi", "geo_accession", "other"] | None = None,
+) -> str:
+    """Get external references linked to tracked experiments.
+
+    Returns PMIDs, DOIs, NCT IDs, GEO accessions and other identifiers
+    linked to experiments. These identifiers can be directly passed to
+    PubMed, bioRxiv, ClinicalTrials.gov, or other MCP tools.
+
+    WHEN TO USE: Use to retrieve external references linked to experiments.
+    PMIDs can be passed to PubMed MCP tools.
+    RELATED TOOLS: encode_link_reference, encode_get_citations
+
+    Args:
+        experiment_accession: Filter by specific experiment (optional)
+        reference_type: Filter by reference type (optional):
+            "pmid", "doi", "nct_id", "preprint_doi", "geo_accession", "other"
+
+    Returns:
+        JSON with linked external references.
+    """
+    if experiment_accession:
+        validate_accession(experiment_accession)
+    tracker = _get_tracker()
+    refs = tracker.get_references(
+        accession=experiment_accession,
+        reference_type=reference_type,
+    )
+    return json.dumps(
+        {
+            "references": refs,
+            "count": len(refs),
+        },
+        indent=2,
+        default=str,
+    )
+
+
+# ======================================================================
+# Entry point
+# ======================================================================
+
+
+def main():
+    """Run the MCP server."""
+    mcp.run()
+
+
+if __name__ == "__main__":
+    main()