Manual web tools and the agent workbench share one set of tools, tasks, and result management; the agent calls platform capabilities under tiered permissions, write confirmation, and audit, organizing supported analysis steps as needed — assisting analysis, never executing freely. * Figures are a static catalog count as of 2026-10-16, not a count of successful runs.
One tool catalog, task system, and result store backs both paths: manual submission for analyses with clear targets, agent assistance for natural-language exploration and interpretation; RStudio is an optional, standalone R execution stack for interactive R analysis.
Pick a tool, fill parameters, and submit tasks directly from the web UI. Verified native workflows (RNA-seq / ATAC-seq / ChIP-seq / scRNA-seq) submit directly as task chains, GWAS has a dedicated one-page frontend form, and the rest are routed to the Agent Workbench.
Work in natural language in the Agent Workbench. The agent organizes analyses via skill guidance and capability routing: it first matches a relevant skill (9 platform skills), then routes through four capability classes (execute directly / adapt then execute / needs more input / needs agent exploration) before a governed submission; read/write tiering, write confirmation, and full audit constrain every call.
Web platforms moved tools into the browser but lack context-aware agent collaboration; general LLM agents are flexible but lack structured, auditable access to domain tools and biological data. Plant Omics Agent unifies the two.
Traditional web platforms lack context-aware collaboration
General-purpose LLM agents lack structured, auditable access to domain tools and biological data
Arbitrary scripts run without oversight
Plant omics databases and regulatory resources are everywhere
No way to know which parameters produced an output
Key difference: one shared tool ecosystem, task system, and result store — usable manually, or by agents under tiered permissions and audit.
The upper layer is orchestration — how the agent organizes an analysis: skill guidance plus capability routing. The lower layer is governance — how execution is constrained: read/write tiering, write confirmation, result pagination, and full audit.
Ask a plant omics question in natural language in the Agent Workbench; the session is bound to user / project / session.
9 platform skills (batch RNA-seq, RStudio runtime, paper writing, etc.) provide method guidance for the task.
The capability catalog decides: execute directly / adapt then execute / needs more input / needs agent exploration.
Eligibility is checked before submission; the live tool schema is fetched and inputs are validated.
Real platform tools are invoked; the agent reads real outputs rather than paraphrasing.
Continue to the next step, verify evidence, or hand the key decision back to the researcher.
The backend writes MCP config dynamically per agent project, binding permission scope and write-confirmation state.
Read-only sessions can only access L1 read tools; subagent profiles further restrict the tool set.
Inputs are schema-validated; write / workspace file operations follow session scope gates.
Large results are paginated via the Result Store to avoid context overflow.
Every tool call enters MCP audit — with user, scope, confirmation state, and audit metadata.
Task logs, parameter snapshots, and artifact manifests are retained — any analysis can be revisited and reproduced.
Tool schemas, database search, gene ID mapping, motif-to-TF, task logs, result manifests
No confirmationSubmit action / native workflow, enrichment & plotting batch
ConfirmationFile access constrained to the configured workspace
Session scopeTwo MCP channels: compat (/api/mcp/*) and official (/api/mcp-official/*); static expansion counts 56 tools (48 read, 7 write, 1 file); runtime visibility adapts to session permissions.
From manually submitting tool tasks and native workflows, to agent-assisted evidence integration and one-page GWAS submission.
Given a candidate gene / DEG / motif table, query MaizeGDB / UniProt / KEGG / JASPAR / PlantRegMap via MCP, run gene ID mapping and motif-to-TF mapping, and consolidate everything into one evidence table — fully audited.
Ask in the Agent Workbench: the agent discovers tools via skill matching and capability routing, searches databases and workspace files, submits governed actions after confirmation, and interprets the real outputs.
Fill in sample metadata and reads in a form, then submit the task chain directly (QC → trim → align → quantify → DE → enrichment) and track progress in the task list; the run ends with a self-contained HTML report (layered writing model + hedging guard, claim strength aligned with evidence) and a project delivery ZIP.
One dedicated frontend form: genotype as a single VCF or PLINK BED triplet + a phenotype table, reference version B73 v4 / v5 / Others, with an intent review before submission. Five models in one submission: GEMMA (LMM / LM) + rMVP (GLM / MLM / FarmCPU); QC + kinship matrix + PCA with tiered significance thresholds (strict to relaxed); SnpEff effect annotation (B73 v5 default, v3 explicit opt-in); CMplot publication-grade Manhattan / QQ / density plots; same-build local gene annotation + optional GO/KEGG enrichment (B73v4 / v5 dual backgrounds); a self-contained HTML summary report + project delivery ZIP.
Each layer has clear responsibilities and boundaries; the manual and agent paths share the same capabilities.
19 plant-relevant public databases and regulatory resources, exposed to agents and manual users via unified MCP tools, plus a natural-language database query entry. Maize TF and motif evidence is backed by a versioned resource supply chain: the GRASSIUS 2.0 master list, iTAK v2, the CIS-BP 3.10 motif library, JASPAR 2026, PlantPAN 5.0, plus published evidence sets such as Huo2025 mGRN and Galli2025 DAP. The resource catalog is version-managed and retires what breaks: since 2026-09, 5 upstream-dead databases have been retired and PlantPAN upgraded to 5.0 with real form-based queries.
search_bio_database(s)gene_alias_resolvergene_id_mappertf_target_evidence_aggregationmotif_to_tf_mappingliterature_relation_extractionMaizeGDB ID / version / annotationFrom sequencing data to regulatory networks, evolution, epigenomics, GWAS and visualization — what matters is these capabilities are shared across both the manual and agent paths, not just the counts.
FastQC/MultiQC, Trimmomatic/fastp, HISAT2/STAR, featureCounts/StringTie, Salmon/Kallisto, DESeq2, GO/KEGG, GSVA/fgsea
FastQC, Trimmomatic, Bowtie2, SAMtools, MACS3, HOMER motif, TOBIAS, IDR, deepTools
Bowtie2, MACS3, deepTools
Cell Ranger, Scanpy, scVelo, CellChat
33 scientific plot templates (19 Python + 14 R) · maize GO/KEGG batch enrichment (B73v4 / v5 dual backgrounds) · eggNOG v5.0.2 built-in annotation library · GSVA / fgsea
WGCNA, GENIE3, GRNBoost2, SCENIC/SCENIC+, ATAC2GRN
MAFFT, IQ-TREE, OrthoFinder, MCScanX, WGDI, MUMmer4, SyRI
Bismark, methylKit, HiC-Pro, ChIA-PET2
GEMMA (LMM / LM) + rMVP (GLM / MLM / FarmCPU), five models in one submission · QC + kinship matrix + PCA · tiered significance thresholds (strict to relaxed) · SnpEff effect annotation (B73 v5 default, v3 explicit opt-in) · CMplot publication-grade Manhattan / QQ / density plots · local annotation + optional enrichment · HTML report + delivery ZIP · posthoc tiered exploration
Each area is rated as it stands: MCP governance and controlled execution are the core mechanism; the RNA-seq / ATAC-seq / ChIP-seq / scRNA-seq native workflows submit directly from the frontend, and GWAS has a dedicated one-page form plus posthoc; extended pipelines such as synteny route through the agent and remain preliminary. Ratings reflect catalog and run-verification status, not scientific conclusions.
Frontend/backend split, one-command Docker deployment, with local development and RStudio interactive analysis (optional standalone).
The landing page only gives the overview. Deployment, both usage paths in full, MCP governance, and database queries all live in the guide.
Open the platform now: submit a tool task, start a one-page GWAS submission, or begin an analysis with one sentence in the Agent Workbench.
Both entries point to the same Plant Omics Agent platform; choose by network.