AdisInsight
Query Springer Nature's drug-pipeline and trial database from an AI agent.
Query EMBL-EBI's bioactivity database.
Opens Agent Studio, where connecting is one click. The connector URL below works in any MCP client.
Paste it into any MCP client. Setup docs
The ChEMBL connector puts EMBL-EBI's curated bioactivity database inside Claude, ChatGPT and any MCP-compatible agent. Its six read-only tools search compounds by name, ChEMBL ID or chemical structure, look up biological targets by gene symbol, and return the quantitative activity measurements — IC50, EC50, Ki, Kd — that link the two. Anthropic builds and hosts it. It needs no sign-in.
Verified connector
Listed by Anthropic as an Anthropic-built connector in its Connectors Directory.
Connection checked by Agentman on .
Anthropic states this reflects the level of review a connector received, not a security audit.
Search for chemical compounds in ChEMBL database by name, ChEMBL ID, or molecular structure. WHEN TO USE compound_search vs drug_search: • compound_search: Use for ANY molecule lookup by name, ID, or structure (may lack clinical data) • drug_search: Use ONLY when searching by therapeutic indication (e.g., "drugs for diabetes") For a simple drug name lookup like "find aspirin", use compound_search. SEARCH STRATEGIES: • By name: Use 'name' for drug names, synonyms, or trade names (case-insensitive, partial match) • By ID: Use 'chembl_id' for direct lookup when you know the exact identifier (e.g., 'CHEMBL25' for aspirin) • By structure: Use 'smiles' with 'similarity_threshold' (70-100%) for finding structurally similar compounds • Substructure: Use 'smiles' without threshold to find compounds containing that substructure SIMILARITY SEARCH (uses Morgan fingerprints, radius 2, 2048 bits via FPSim2): • 70%: Loose similarity (finds diverse analogs) • 80%: Good starting threshold (finds close analogs) • 90%+: Very strict (finds nearly identical compounds) MAX_PHASE VALUES (clinical development stage): • 4 = Approved (marketed drug, e.g., FDA/EMA approved) • 3 = Phase 3 Clinical Trials • 2 = Phase 2 Clinical Trials (includes INN applications) • 1 = Phase 1 Clinical Trials (includes USAN applications) • 0.5 = Early Phase 1 • -1 = Unknown clinical phase • NULL = Preclinical compound (bioactivity data only) RETURNED DATA INCLUDES: • Molecular properties: MW, ALogP (lipophilicity), PSA, HBA, HBD, rotatable bonds, aromatic rings • QED (Quantitative Estimate of Drug-likeness): 0-1 scale, higher = more drug-like • Chirality: 0=racemic, 1=single stereoisomer, 2=achiral, -1=unchecked • Rule of Five: MW<500, ALogP<5, HBD<5, HBA<10 (compounds with 0-1 violations are Ro5 compliant) • ATC classifications, cross-references to external databases, molecule hierarchy (parent/salt forms) TIPS: • For approved drugs, add max_phase=4 • ChEMBL uses compound 'families' - salts map to parent compounds • Use get_bioactivity after finding compounds to get their target activities • Properties are calculated on parent (salt-free) form EXAMPLES: • Find aspirin: name='aspirin' or chembl_id='CHEMBL25' • Find kinase inhibitors: name='nib', max_phase=4 • Find structural analogs: Use SMILES with similarity_threshold=80
Retrieve bioactivity measurements (IC50, Ki, EC50, etc.) for compound-target interactions. WORKFLOW: Use target_search or compound_search first to get ChEMBL IDs, then query bioactivity. ACTIVITY TYPES (standard_type field): • IC50: Half-maximal inhibitory concentration (most common for inhibitors) • Ki: Inhibition constant (binding affinity, independent of substrate concentration) • Kd: Dissociation constant (direct binding measurement) • EC50: Half-maximal effective concentration (for agonists) • AC50: Half-maximal activity concentration • Potency/ED50: Effective dose measurements pChEMBL VALUE (recommended for filtering - standardized potency): • Definition: -log10(molar IC50/Ki/Kd/EC50/AC50/Potency/ED50) • Only calculated when: standard_relation='=' AND standard_units='nM' AND value>0 • pChEMBL 9 = 1 nM (highly potent, drug-like) • pChEMBL 7 = 100 nM (potent) • pChEMBL 6 = 1 µM (moderate) • pChEMBL 5 = 10 µM (weak) • pChEMBL 3 = 1 mM (very weak) • Use min_pchembl >= 6 for µM or better activity • Use min_pchembl >= 7 for sub-100nM potency (drug-like) ASSAY TYPES (assay_type field): • B (Binding): Direct target binding measurements (Ki, Kd) • F (Functional): Biological effect in cells/tissues (EC50, IC50 in cellular context) • A (ADME): Absorption, distribution, metabolism, excretion assays (t1/2, bioavailability) • T (Toxicity): Cytotoxicity, hERG inhibition • P (Physicochemical): Solubility, stability (no biological material) • U (Unclassified): Cannot fit single category DATA VALIDITY FLAGS (data_validity_comment field): • 'Outside typical range': Value unusually high/low for activity type • 'Potential missing data': Incomplete data entry • 'Potential author error': Suspected error in publication • 'Manually validated': Curator confirmed accuracy • 'Potential transcription error': Values differ by 3 or 6 orders of magnitude (unit error) LIGAND EFFICIENCY METRICS (in results when available): • LE (Ligand Efficiency): Activity per heavy atom • BEI (Binding Efficiency Index): Activity per molecular weight • LLE (Lipophilic Ligand Efficiency): pActivity - LogP • SEI (Surface Efficiency Index): Activity per polar surface area TIPS: • Use with ADMET targets to get experimental ADMET data • Combine molecule_chembl_id + target_chembl_id for specific compound-target pairs • Check potential_duplicate flag - may indicate cited (not independent) measurements • activity_comment may indicate 'active'/'inactive' conclusions from depositor • document_chembl_id links to source publication for verification
Search for biological targets (proteins, enzymes, receptors, organisms) in ChEMBL database. SEARCH STRATEGIES: • By name: Use 'target_name' for protein names, families (e.g., 'kinase'), or receptors • By gene: Use 'gene_symbol' for exact gene symbol matches (e.g., 'EGFR', 'BRAF', 'TP53') • By ID: Use 'target_chembl_id' for direct lookup (e.g., 'CHEMBL203' for EGFR) • By organism: Filter results to specific species (e.g., 'Homo sapiens', 'Mus musculus') • By type: Filter by target_type to get specific categories TARGET TYPES (target_type field): • SINGLE PROTEIN: Individual protein (most common, highest confidence bioactivity data) • PROTEIN COMPLEX: Multi-subunit complex (e.g., ion channels, GPCRs with multiple subunits) • PROTEIN FAMILY: Homologous protein groups (broader search, lower specificity) • PROTEIN-PROTEIN INTERACTION: Two interacting proteins • CHIMERIC PROTEIN: Engineered fusion protein • SELECTIVITY GROUP: Panel of related targets for selectivity profiling • ORGANISM: Whole organism (bacteria, parasites, viruses, fungi) • TISSUE: Tissue-level target • CELL-LINE: Cell-based target (phenotypic screening) • NUCLEIC-ACID: DNA/RNA targets • SUBCELLULAR: Subcellular compartments • UNKNOWN: Unclassified targets TARGET CONFIDENCE SCORES (in bioactivity data): • 9: Direct single protein target (highest confidence, most reliable) • 8: Homologous single protein (inferred from related species) • 7: Direct protein complex subunits (multi-subunit target) • 6: Homologous protein complex • 5: Direct protein selectivity group • 4: Homologous selectivity group • 3: Protein not in target complex • 2: Non-protein organism target • 1: Non-molecular target (cell-line, organism, tissue) • 0: Default or uncurated TARGET RELATIONSHIPS (between targets): • EQUIVALENT TO: Same target in different contexts • OVERLAPS WITH: Partially shared components • SUBSET OF: Contains subset of components • SUPERSET OF: Contains additional components RETURNED DATA INCLUDES: • Target components: Protein subunits with accessions (UniProt) • GO annotations: Gene Ontology terms for molecular function, biological process, cellular component • Cross-references: Links to UniProt, PFAM, InterPro, IntAct, Reactome, etc. • Species information: Tax ID and organism name TIPS: • Use gene_symbol for exact matches - more precise than target_name • After finding a target, use get_bioactivity with target_chembl_id to find active compounds • For drug discovery, focus on SINGLE PROTEIN targets (confidence >= 7) • Filter by organism='Homo sapiens' for human targets • Check GO annotations to understand target function and localization
Retrieve mechanism of action (MoA) data for approved drugs and clinical candidates. PURPOSE: Understand how drugs interact with their targets - essential for drug repurposing, understanding polypharmacology, and target validation. ACTION TYPES (action_type field): • INHIBITOR: Blocks target activity (most common for small molecule drugs) • ANTAGONIST: Blocks receptor activation (prevents agonist binding) • AGONIST: Activates receptor (mimics natural ligand) • BLOCKER: Blocks ion channels or transporters • MODULATOR: Alters target activity (often allosteric mechanism) • POSITIVE ALLOSTERIC MODULATOR: Enhances agonist response without binding orthosteric site • NEGATIVE ALLOSTERIC MODULATOR: Reduces agonist response • OPENER: Opens ion channels (increases conductance) • ACTIVATOR: Increases enzyme activity • PARTIAL AGONIST: Partially activates receptor (submaximal efficacy) • INVERSE AGONIST: Reduces constitutive (basal) receptor activity • SUBSTRATE: Acts as substrate for enzyme (e.g., prodrugs) • RELEASING AGENT: Causes release of neurotransmitters • SEQUESTERING AGENT: Binds and removes target (e.g., antibodies) KEY RESULT FIELDS: • direct_interaction: true = drug binds directly to target; false = indirect effect • disease_efficacy: true = target is directly relevant to therapeutic effect • molecular_mechanism: Specific molecular action (e.g., 'Cyclooxygenase inhibitor') • binding_site_name/comment: Where drug binds on target (e.g., 'ATP binding site') • selectivity_comment: Notes on target selectivity vs related proteins • mechanism_refs: Literature references supporting the mechanism BINDING SITE INFORMATION: Results may include binding site details when known: • site_name: Named binding pocket (e.g., 'Colchicine site', 'ATP binding domain') • site_id: ChEMBL binding site identifier for further lookup WORKFLOW: 1. Find compound: compound_search(name='imatinib') 2. Get mechanisms: get_mechanism(molecule_chembl_id='CHEMBL941') 3. Validate with bioactivity: get_bioactivity(molecule_chembl_id='CHEMBL941', target_chembl_id='CHEMBL1862') TIPS: • MoA data is manually curated for approved drugs and advanced clinical candidates • Use target_chembl_id to find all drugs acting on a specific target (drug repurposing) • Check mechanism_refs for original publications supporting the mechanism • Combine with get_bioactivity to see quantitative potency data for mechanism targets • Parent compound IDs work even when mechanism is stored under salt form
Search for approved drugs and clinical candidates by therapeutic indication. PURPOSE: Find drugs used for specific diseases, identify approved treatments, explore drug repurposing opportunities. DRUG VS COMPOUND: • Drug: Compound with assigned INN/USAN name + clinical data (max_phase ≥ 1) • Compound: Any molecule in ChEMBL (may only have bioactivity data) • Use compound_search for broader chemical searches, drug_search for therapeutic applications MAX_PHASE VALUES (clinical development stage): • 4 = Approved (marketed drug, e.g., FDA/EMA approved) • 3 = Phase III Clinical Trials (large-scale efficacy trials) • 2 = Phase II Clinical Trials (proof of concept, includes INN applications) • 1 = Phase I Clinical Trials (safety, includes USAN applications) • 0.5 = Early Phase 1 (exploratory studies) • -1 = Unknown clinical phase (status uncertain) • NULL = Preclinical only (no human trials, compound_search more appropriate) SAFETY FLAGS IN RESULTS: • black_box_warning: 1 = has FDA black box warning (serious safety concern) • withdrawn_flag: true = withdrawn from one or more markets • withdrawn_reason: Why drug was withdrawn (e.g., hepatotoxicity, cardiac effects) • withdrawn_country/year/class: Details about market withdrawal INDICATION SEARCH: • Uses MeSH (Medical Subject Headings) disease terminology • Also searches EFO (Experimental Factor Ontology) terms • Partial matching supported (e.g., 'cancer' matches 'breast cancer', 'lung cancer') • Common indications: hypertension, diabetes, cancer, asthma, depression, arthritis INDICATION DETAILS IN RESULTS: • mesh_id/mesh_heading: MeSH disease classification • efo_id/efo_term: EFO disease ontology • max_phase_for_ind: Highest phase achieved for this specific indication TIPS: • Use only_approved=True for marketed drugs only • Check black_box_warning and withdrawn_flag for safety information • After finding drugs, use get_mechanism to understand their molecular targets • Use max_phase=3 to include drugs in late-stage trials (potential future approvals) WORKFLOW EXAMPLE: 1. drug_search(indication='hypertension', only_approved=True) → Find approved antihypertensives 2. get_mechanism(molecule_chembl_id='CHEMBL1200749') → Get targets for amlodipine 3. get_bioactivity(target_chembl_id='CHEMBL1940') → Find other L-type calcium channel blockers
Retrieve ADMET-related molecular properties for drug-likeness assessment. IMPORTANT: ChEMBL provides CALCULATED molecular properties (from structure), not experimental ADMET data. For experimental ADMET measurements, use get_bioactivity with specific ADMET target ChEMBL IDs. CALCULATED PROPERTIES (what this tool returns): • ALogP: Calculated lipophilicity (Wildman-Crippen LogP) - Optimal for oral drugs: 1-3 - < 0: Poor membrane permeability - > 5: Poor solubility, potential accumulation • Molecular Weight (full_mwt): Total molecular weight including salts • Molecular Weight (mw_freebase): Parent compound molecular weight only • H-bond Donors (HBD): Rule-of-5 limit < 5 • H-bond Acceptors (HBA): Rule-of-5 limit < 10 • Polar Surface Area (PSA): Topological PSA - < 140 Ų: Good oral absorption - < 90 Ų: Better CNS penetration (crosses BBB) • Rotatable Bonds (RTB): < 10 for good oral bioavailability • Heavy Atoms: Non-hydrogen atom count • Aromatic Rings: < 4 recommended for drug-likeness • Rule-of-5 Violations (num_ro5_violations): 0-1 preferred • Rule-of-3 Pass (ro3_pass): Y/N for fragment-like properties • QED Weighted: Quantitative Estimate of Drug-likeness (0-1 scale) - Higher = more drug-like profile - Based on MW, ALogP, HBD, HBA, PSA, RTB, aromatic rings, alerts DRUG-LIKENESS GUIDELINES: • Lipinski Rule-of-5: MW < 500, ALogP < 5, HBD ≤ 5, HBA ≤ 10 • Veber Rules: PSA ≤ 140 Ų, RTB ≤ 10 • Lead-like: MW < 450, ALogP -4 to 4.2, RTB ≤ 10 FOR EXPERIMENTAL ADMET DATA (use get_bioactivity with these targets): • hERG (CHEMBL240): Cardiac safety (K+ channel, IC50 > 10µM preferred) • CYP3A4 (CHEMBL340): Major metabolizing enzyme (avoid strong inhibition) • CYP2D6 (CHEMBL289): Polymorphic CYP (genetic variability concerns) • CYP2C9 (CHEMBL3397): Warfarin metabolism (drug interactions) • P-glycoprotein (CHEMBL4302): Drug efflux transporter (affects BBB, gut) WORKFLOW EXAMPLE: 1. Get calculated properties: get_admet(molecule_chembl_id='CHEMBL941') 2. Check hERG liability: get_bioactivity(molecule_chembl_id='CHEMBL941', target_chembl_id='CHEMBL240') 3. Check CYP inhibition: get_bioactivity(molecule_chembl_id='CHEMBL941', target_chembl_id='CHEMBL340')
Read from the server on 2026-08-18, including each tool's own safety annotations.
latest directory serves ChEMBL 37, prepared 1 May 2026. The gap is countable: v34 holds 2,431,025 compounds and 20,772,701 activities against v37's 2,921,148 and 24,527,044. Roughly 3.75 million activity records curated since are unreachable here.min_pchembl normalises across activity types onto a log molar scale but cannot normalise across experimental design. Read standard_type, units and target organism first.get_admet is calculated, not measured, and properties describe the parent compound. compound_search states properties are computed on the parent, salt-free form, so they may not describe the salt actually used in a study. Rule-of-Five and QED are prioritisation heuristics, not safety predictions.similarity_threshold has a schema minimum of 70, so broader similarity searching is not reachable through this connector.get_mechanism's description states the data is manually curated for approved drugs and advanced clinical candidates, so preclinical compounds usually return nothing — a fact about curation scope, not about the compound.prompts/get returned a rendering error for analyze_compound, compare_drugs and assess_admet on 2026-08-18, and the directory advertises none of them. assess_admet's description offers to "predict drug safety profile" while the tool beneath it returns structure-calculated descriptors; no vendor source qualifies that wording. Use the tools directly and treat any such output as triage, never a safety assessment.initialize, tools/list, prompts/list and resources/list handshake, plus three prompts/get calls that returned errors. Tool behaviour above comes from the server's own schemas, Anthropic's tutorial and EMBL-EBI's release notes; corpus counts come from EMBL-EBI directly, not through the connector.anthropics/life-sciences plugin manifest points chembl at a deepsense.ai host rather than the directory endpoint, and we could not handshake that host. See the setup section above.No. Every tool returns laboratory measurements and curated literature records, not therapeutic guidance. An IC50 or Ki value describes how a compound behaved in an assay, not whether it treats anyone or at what dose. Nothing on this server accounts for a person, a condition or a prescription. Treatment decisions belong with a qualified healthcare professional.
The server's own initialize response says ChEMBL Database v34, prepared 28 March 2024. EMBL-EBI has published three releases since: ChEMBL 35, 36, and ChEMBL 37 in May 2026, which the EBI FTP site serves as its latest directory. So results can lag the current database by roughly two years of curation, and v34 literature stops at July 2023.
No, and its own description says so first. It returns properties calculated from chemical structure — ALogP, molecular weight, polar surface area, hydrogen-bond counts, rotatable bonds, Rule-of-Five violations and a QED drug-likeness score. For measured ADMET endpoints the tool points you at get_bioactivity against specific assay targets such as hERG or CYP3A4.
Yes. compound_search accepts a SMILES string. Supply similarity_threshold between 70 and 100 for a similarity search using Morgan fingerprints at radius 2 and 2048 bits, per the tool description. Omit the threshold and the same SMILES runs as a substructure search instead. There is no InChI parameter on any tool.
No. We opened an anonymous handshake against the endpoint on 2026-08-18 and it returned all six tool definitions with no credential. The directory records the server as authless, and the host publishes no OAuth protected-resource descriptor at either the root or the RFC 9728 path-suffixed form. ChEMBL data is itself openly licensed.
ChEMBL data is released by EMBL-EBI under the Creative Commons Attribution-ShareAlike 3.0 Unported licence. That permits reuse including commercially, but attaches two conditions: attribute ChEMBL, and license any adapted database you distribute under the same terms. The ShareAlike condition follows anything an agent retrieves and redistributes as a derived dataset.
Anthropic lists it at anthropic tier in its Connectors Directory and hosts the endpoint at hcls.mcp.claude.com. The data is not Anthropic's. ChEMBL is built and manually curated by EMBL-EBI, the European Bioinformatics Institute, from medicinal-chemistry literature, patents and deposited screening datasets. Anthropic operating the connector says nothing about any measurement's accuracy.
claude.com/robots.txt is Allow: / with no Content-Signal line (retrieved 2026-08-18). · retrieved 2026-08-18initialize, tools/list, prompts/list and resources/list handshake against https://hcls.mcp.claude.com/chembl/mcp — six tools with descriptions, input and output schemas and annotations, plus three prompts, over an anonymous connection (2026-08-18). serverInfo reported chembl version 3.2.0 and the instructions string named "the EMBL-EBI ChEMBL Database v34". Source for the tool count, every tool description and parameter named on this page, the uniform read-only/non-destructive/idempotent annotations, the single openWorldHint: false on get_admet, the SMILES and Morgan-fingerprint detail, and the declared release. resources/list returned -32601 Method not found, consistent with the server declaring no resources capability. Three prompts/get calls returned rendering errors. No tool was called. · retrieved 2026-08-18https://hcls.mcp.claude.com/.well-known/oauth-protected-resource returned 404, as did the RFC 9728 path-suffixed form /.well-known/oauth-protected-resource/chembl/mcp and a per-tool probe at /.well-known/oauth-protected-resource/get_admet (all retrieved 2026-08-18). Consistent with the successful anonymous handshake and with the directory's authless record. robots.txt on that host also 404s. · retrieved 2026-08-18chembl_37. Source for the current release being ChEMBL 37 prepared 1 May 2026, its 2,921,148 compounds and 24,527,044 activities, and therefore for the release-lag figures under Limits. www.ebi.ac.uk/robots.txt carries a User-agent: * block with Crawl-Delay: 10 and no disallow covering /chembl, and no Content-Signal line (retrieved 2026-08-18). · retrieved 2026-08-18chembl/README.md, no chembl/.mcp.json and no bundled skill for this connector — a full tree listing returned only chembl/.claude-plugin/plugin.json under that path. · retrieved 2026-08-18https://mcp.deepsense.ai/chembl/mcp returned HTTP 403 from CloudFront, stating the distribution does not allow the request method used (retrieved 2026-08-18). Nothing could be established about what that host serves, and no claim is made about it. · retrieved 2026-08-18prompt_names array, anthropic tier, the life-sciences category, is_authless: true, Read permissions, streamable-HTTP transport, Anthropic as author, and the endpoint used throughout this page. It is the only ChEMBL entry in the 1,897-entry snapshot. · retrieved 2026-01-07Connect ChEMBL once and your agents call these tools on their own: on a schedule, in a workflow, with nobody at the keyboard.