X-Mech Suite: twelve autonomous knowledge factories, one shared standard
The X-Mech suite brings together twelve autonomous knowledge factories that describe a microbe at every scale: its taxon and strain identity, habitat, community, traits, pathways, cellular structures, protein functions and unresolved domains, natural products, antibiotics, ingredients and culture media. Established corpora curate structured records with evidence and provenance; DUFMech begins with an InterPro/Pfam seed worklist. culturebotai-claw coordinates the eleven projects currently declared in its fleet manifest.
Each X-Mech is an autonomous knowledge factory that curates, validates, and connects scientific evidence to enable discovery, with human oversight. Its knowledge base is the collection of structured records it produces and maintains.
Entry totals combine the published record counts with DUFMech's 6,532 seed families. Seed families are identified candidates and historical DUFs, not curated mechanism claims. PR activity includes all twelve Mechs, CLAW and this website.
Shared vocabulary
The Mechs are joinable because they ground records in the same public ontologies. The table counts identifier occurrences per vocabulary in each Mech's record corpus; darker cells mean more. Scroll sideways to explore all 160 vocabularies, with DOI and PMID first, followed by vocabularies used by the most Mechs to the fewest. Rows run from the Mech using the most vocabularies to the fewest. Mech names stay visible as you scroll. Select any cell to see its occurrence count and matching records below, including cells with no occurrences. Click a Mech name to open it, or a column button to filter the graph.
This vocabulary census covers all twelve Mechs. Counts are prefix occurrences in structured records and seed worklists (one family per DUFMech worklist row; merged recipes for CultureMech, communities and isolate records for CommunityMech, habitat records for HabitatMech, species-level and infraspecific taxon records for TaxonMech, each carrying its lineage, so a higher taxon is counted once per record under it) as of 5 October 2026. DOI and PMID are the first two columns. All remaining columns are ordered by how many Mechs use each vocabulary, from most to least, then by total occurrences across the fleet and vocabulary name. Rows are ordered by how many vocabularies each Mech uses; ties retain the biological scale order. Every cell opens its matching-record summary; lists show up to 300 records. Column buttons filter the graph for vocabularies with a shared-term index. The shared terms behind each chord come only from what a record cites outside prose fields, so an identifier named only in a note, a change log or a quotation counts here but is not shared. ChEBI binds the chemistry arm (media, ingredients, antibiotics, proteins); NCBITaxon and ENVO bind the organism arm (habitat, community, taxa, traits); GO and METPO bridge phenotype, structure and protein.
The twelve Mechs
Each card carries its Mech’s identity color. Hover a card to trace its ties in the graph above; use “Show in graph” to select it. The small print under each headline number gives reviewed records and merged pull requests; a Mech whose schema has no field recording review shows only the pull-request count, rather than a zero that would claim more than it knows. Counts were checked on October 5, 2026, with all twelve Mechs included in the vocabulary census and repository statistics. DUFMech’s 6,532 seed families are labeled separately from curated mechanism records. Each Browse link is the source for its numbered card, except CultureMech, whose count comes from its repository’s corpus snapshot, and DUFMech, whose count comes from its linked seed snapshot. CommunityMech lists 456 communities online and keeps four additional isolate records in its repository. These published-browser totals may differ from the record-corpus census below. Fleet membership and capability declarations come from CLAW fleet manifest at ad3862e.
HabitatMech
HabitatFour habitat vocabularies harmonized into ontology-grounded records that keep every source's attestation.
686 reviewed · 876 merged PRs
CommunityMech
CommunityAutonomous knowledge factory for microbial communities, their interactions, cultivation conditions and evidence.
624 merged PRs
TaxonMech
Taxa and strainsMicrobial taxa and strains identified by NCBI Taxonomy, harmonized with GTDB, LPSN and BacDive, with evidence for strain-to-genome links.
0 reviewed · 27 merged PRs
TraitMech
TraitsAutonomous knowledge factory for microbial ecophysiological traits, seeded from METPO, with one curated YAML per trait and causal mechanism graphs.
427 reviewed · 912 merged PRs
PathwayMech
PathwaysMicrobial pathway mechanisms with explicit participants, reactions, causal edges and reference-backed evidence.
178 merged PRs
CellStructureMech
Cell structuresOrganelles, envelope layers, appendages, microcompartments and complexes: components, distribution, function and causal mechanism.
0 reviewed · 967 merged PRs
ProteinTraitsMech
Protein traitsProtein sequence, structure and function trait classes seeded from InterPro, Pfam, Rhea, CATH, SCOPe, CARD and more.
39,591 reviewed · 383 merged PRs
DUFMech
Unknown functionsEvidence gathering for domains and protein families whose functions remain unresolved.
The worklist separates 4,533 unknown candidates from 1,999 historical DUFs. These families are candidates for investigation, not curated mechanism records.
28 merged PRs
NaturalProductMech
Natural productsOne record per natural product structure: who makes it, from which gene cluster, what it does, and the evidence for all three.
0 reviewed · 361 merged PRs
AntibioticMech
AntibioticsOne record per antimicrobial chemical structure, harmonizing ChEBI and CARD with targets, mode of action, resistance and evidence.
80 reviewed · 648 merged PRs
MediaIngredientMech
IngredientsLLM-assisted curation of media-ingredient ontology mappings with full audit trails; owns ingredient identity for the fleet.
396 merged PRs
CultureMech
MediaAutonomous knowledge factory curating versioned, ontology-grounded culture-media recipes from MediaDive, TogoMedium, KOMODO and the major collections.
265 merged PRs
Explore the current sites
Alongside record browsing, the Mechs publish complementary ways to explore their data:
- Pathway mechanisms: browse PathwayMech’s 152 pathway summaries, labeled mechanistic edges and cited evidence. Its repository records include participants, reactions and cited evidence, alongside KGX nodes and edges and SSSOM source mappings.
- Unknown protein families: explore DUFMech’s InterPro/Pfam worklist. Its family identifiers connect the worklist to the other Mechs through the shared-vocabulary graph and heatmap.
- Taxa and strains: TaxonMech lists the 20,648 taxa with a type strain, each naming the BacDive deposit that LPSN records as the type. Taxon pages carry strain-level NCBI, GTDB, BV-BRC/PATRIC, IMG and AllTheBacteria genome identifiers, plus StrainInfo references and their evidence. The source catalogue describes provenance.
- Community, trait, cell structure and media similarity: explore CommunityMech, TraitMech, cell structures, ingredients and culture media through their embedding browsers.
- Protein traits, proteins and sequences: ProteinTraitsMech provides a text-embedding corpus map, a map of 68,657 Swiss-Prot proteins by their traits, and an ESM-2 sequence map of canonical-example proteins.
- Chemical structures: compare compounds in the AntibioticMech chemical map and the NaturalProductMech structure map. Both Mechs also offer a corpus map that places compounds by their record text rather than their structure: AntibioticMech and NaturalProductMech.
- Habitat meaning: HabitatMech offers a semantic text map alongside its ontology-grounded record browser.
How the Mechs reference each other
Beyond shared vocabulary, Mechs name one another directly, in record fields, in schema slots, and in the curation practices they adopt from each other. The graph draws biological data links and complementary coverage; this detailed list also includes shared software and record formats. Arrows point at the Mech that consumes, or at the one a scope decision defers to. Taxa tie TaxonMech to much of the fleet through shared NCBI Taxonomy identifiers, while DUFMech connects to ProteinTraitsMech records through shared Pfam and InterPro families. Those links appear as chords in the graph’s shared-vocabulary layer. Direct-reference arrows require a committed record, schema or curation practice naming the other Mech.
habitatmech: prefix, and TaxonMech's curation-event helper and closed-schema write gate are ported from HabitatMech's and CellStructureMech's; no TaxonMech record has a causal graph yet.Schema · TaxonMech schema CausalNodeTypeEnum and prefixes; src/taxonmech/curate/curation_event.py, src/taxonmech/validation/write_validated.pyrender_pages.py, including its paged category pages that filter through a JSON index, and its stylesheet is ported from HabitatMech's. Together they render AntibioticMech's whole browser. Its curation-event helper and closed-schema write gate are adapted from HabitatMech's copies; their docstrings still credit TraitMech, which HabitatMech ported them from.Practice · AntibioticMech scripts/render_pages.py; src/antibioticmech/templates/style.css; src/antibioticmech/curate/curation_event.pyresearch_community_edison.py and deep-research-community skill. Its METPO proposal skill takes CommunityMech's v1 proposal cohort as the worked example to follow, and keeps TraitMech's placeholder ID blocks clear of that cohort's.Practice · TraitMech scripts/research_trait_edison.py; .claude/skills/deep-research-trait and metpo-proposalcommunitymech-unmapped and cites its CommunityMech source id (communitymech.ingredient:name). 10 records are still live; the other 7 were merged into existing records.Record data · MIM data/ingredients (synonym source communitymech-unmapped)/ground-or-propose-metpo curation command is adapted from TraitMech's command of the same name, and treats community interaction graphs as the analog of TraitMech's causal graphs. Its strict validator, writer audit and curation-event helper also name TraitMech as a source, alongside CultureMech and MediaIngredientMech.Practice · CommunityMech .claude/commands/ground-or-propose-metpo.md; scripts/validate_strict.py, scripts/audit_writers.pytraitmech: prefix; no TaxonMech record has a causal graph yet.Schema · TaxonMech schema CausalNodeTypeEnum and prefixesnew_history_record.py.Schema · ProteinTraitsMech conf/traitmech_category_vocabulary.yaml; scripts/audit_cross_mech_categories.py; scripts/new_history_record.py/ground-or-propose-ingredient curation command is ported from TraitMech's /ground-or-propose-metpo, re-aimed at chemical identity instead of phenotype.Practice · TraitMech docs/SCHEMA.md, README.md, scripts/seed_from_metpo.py; MIM .claude/commands/ground-or-propose-ingredient.md (MIM#92)associated_traits links to TraitMech or METPO terms.Practice · CellStructureMech docs/CURATION.md (record-scope table); schema descriptionproteintraitsmech: prefix, but no record uses it.Practice · CellStructureMech docs/CURATION.md; schema components slotdownload.yaml source catalogue follows the shape of ProteinTraitsMech's, and its check_sources.py validator is adapted from ProteinTraitsMech's. Unlike the original, it fails a source that records no licence.Practice · TraitMech download.yaml; scripts/check_sources.py; justfile sources-checkmedia_ingredient_mech_id), but they use MIM's retired MediaIngredientMech:NNNNNN scheme, so the ids no longer resolve. Its environment-matching tools read MIM's ingredient records and SSSOM mappings. They join on the exactMatch CHEBI term, which is what MIM's CURIE standard recommends; MIM wrote that standard for CommunityMech (MIM#119).Record data · CommunityMech schema GrowthMediaComponent, src/communitymech/cross_repo_environment.py; MIM docs/CURIE_STANDARD.mdmediaingredientmech_chebi_term, about 70% of all ingredient rows, and older records keep legacy MediaIngredientMech:NNNNNN ids. CultureMech also resolves ingredient labels against a pinned copy of MIM's label index, types its role slots with MIM's vendored role enums and took over stock solutions from MIM's complex-media list, while MIM's schema leaves recipe bodies to CultureMech.Record data · CultureMech schema mediaingredientmech_chebi_term, mim_roles.yaml; src/culturemech/ingredients/mim_label_index.pyculturemech_id, and CommunityMech's media linker copies CultureMech recipe components into them. CommunityMech also ported CultureMech's schema-validated writer and its Edison deep-research runner.Record data · CommunityMech schema GrowthMedia/RelatedMedia culturemech_id; scripts/link_growth_media.py; src/communitymech/validation/write_validated.pyCultureMech:NNNNNN ids, and to turn labels it cannot resolve into new records and synonyms. CultureMech leaves questions about an ingredient's chemical form to MIM, and MIM's validated record writer and curation-event helper are ported from CultureMech.Record data · MIM mappings/culturemech_recipe_membership.tsv, scripts/refresh_occurrence_statistics.py; CultureMech docs/DATA_LAYERS.mdOrchestration: culturebotai-claw
culturebotai-claw is the fleet’s coordinator. It does not hold science of its own; it holds the definition of the fleet, the artifacts every Mech must share byte-for-byte, and tooling for cross-repository curation and validation.
Fleet manifest
fleet.yaml is the single source of truth for which repositories form the fleet. The manifest declares eleven of the twelve projects shown here, including PathwayMech. DUFMech is listed here and is not yet declared in the manifest. Every member declares every capability exactly once as enabled, disabled or not applicable, and every disabled or not-applicable entry gives a reason, so nothing is silently off.
Vendored governance
Shared LinkML modules (mech_shared.yaml, history.yaml), validators and behavioral contracts live in claw and are vendored into its declared members byte-identically, pinned to one immutable claw commit and checked in CI. The canonical registry defines 20 artifacts, with each member receiving the contracts that apply to it.
Pipelines and skills
CLAW provides pipeline discovery, repository checks, shared curation tools and validated dry runs. Agent and pipeline execution through openclaw-cli is not implemented; environment-curation and unified ingredient-mapping apply modes are disabled pending transactional writers. Deep-research tooling supports provider triage and dry-run result capture; provider execution is not implemented.
The current support matrix distinguishes supported tools from planned execution. The autonomous knowledge factory label describes each Mech’s curation model with human oversight; it does not imply that every CLAW workflow runs unattended.
Capability declarations from CLAW record adoption of shared contracts; they do not certify complete validation of every record field or that every failed check blocks a merge. Every disabled entry records a reason, such as no download.yaml or no source queue yet. DUFMech is included with undeclared capabilities because it is not yet a manifest member:
| Mech | ID label validation | Curation history | Strict validation | Schema sync | Release management | Build coordination | Documentation | Testing | Refactoring | Coordination hooks | Deep research | Edison key discovery | Environment coverage | Knowledge-gap scan | Vendored sync | Unmapped inventory input | Source catalogue | Corpus statistics | Causal-graph coverage | Page budgets | KGX export | Source queue | METPO proposal | SSSOM export | Writer audit | Site contract |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| HabitatMech | ||||||||||||||||||||||||||
| CommunityMech | ||||||||||||||||||||||||||
| TaxonMech | ||||||||||||||||||||||||||
| TraitMech | ||||||||||||||||||||||||||
| PathwayMech | ||||||||||||||||||||||||||
| CellStructureMech | ||||||||||||||||||||||||||
| ProteinTraitsMech | ||||||||||||||||||||||||||
| DUFMech | ||||||||||||||||||||||||||
| NaturalProductMech | ||||||||||||||||||||||||||
| AntibioticMech | ||||||||||||||||||||||||||
| MediaIngredientMech | ||||||||||||||||||||||||||
| CultureMech |
What makes a Mech
The established Mechs share curation conventions, with adoption recorded per capability in the table above. DUFMech currently provides a seed worklist; the record and schema conventions below describe the structured corpora. The fleet standard is documented in CLAW’s MECH_STANDARD.md:
- One YAML record per entityA recipe, an ingredient, a community, a taxon, a trait, a pathway, a structure, a protein trait, a natural product, an antibiotic, a habitat. The file is the unit of curation, review and history.
- A LinkML schema per MechRecords validate against the Mech's schema; shared modules such as
mech_shared.yamlare vendored where applicable. PathwayMech defines its own closed record schema. - Ontology-grounded identityRecords are keyed by a public CURIE (ChEBI, GO, METPO, ENVO, NCBITaxon) where one exists, otherwise by a local identifier whose form varies by Mech.
- The ID–label invariantOntology-aware identifier–label checks compare configured fields with their source vocabulary; coverage and CI enforcement vary by corpus.
- Evidence and provenanceEvidence cites its source, such as a publication (PMID or DOI), a database record or a web page, as each Mech's schema allows. Mechs that adopt the source-catalogue or source-queue contract record each source's licence before ingesting it.
- Causal mechanism graphsTraits, pathways, protein traits, structures, natural products, antibiotics, habitats and taxa can carry directed, evidence-backed graphs of mechanism, and community records chain their ecological interactions into causal interaction graphs. Each of these Mechs defines its own node types; most share a core set, such as chemical, pathway and biological process. DUFMech's seed worklist supplies candidate families for evidence gathering.
- Append-only curation historyMechs adopting the history contract record changes as append-only events, preserving provenance across agent and human curation.
- A static browser and an open licenceEstablished corpora publish GitHub Pages browsers; DUFMech currently links to its repository. The suite licensing policy is CC BY 4.0 for data and BSD-3-Clause for code. Redistributed source material retains its applicable terms.
Related resources
- kg-microbe - The central knowledge graph several Mechs draw from; MediaIngredientMech provides candidate ingredient mappings for it
- MicroGrowAgents - Multi-agent media design built on kg-microbe
- Resources - The full tool catalogue, including MicroMediaParam and the kg-microbe utilities
- METPO - The Microbial Ecophysiological Trait and Phenotype Ontology that seeds TraitMech and grounds trait references across the fleet
Bibliography
- Santangelo BE, Hegde H, Caufield JH, Reese J, Kliegr T, Hunter LE, Lozupone CA, Mungall CJ, Joachimiak MP. KG-Microbe — Building Modular and Scalable Knowledge Graphs for Microbiome and Microbial Sciences. GigaScience. 2026;giag077. doi:10.1093/gigascience/giag077
- Naseem S, Miller MA, Martinez-Gomez NC, Sun N, Joachimiak MP. MicroGrowAgents: An Agentic AI System for Microbial Cultivation Engineering. bioRxiv. 2026. doi:10.64898/2026.06.04.729985
See the full bibliography on the Publications page.