The Synthesize Bio MCP Server enables AI agents to compare gene expression between sample groups, analyze bulk and single-cell RNA sequencing data, and run long-running bioinformatics workflows that return completed analysis reports.
Hermes is Nous Research's command-line agent. MCP servers are declared in its YAML config and reloaded without a restart.
This guide connects Synthesize Bio MCP to Hermes through MewCP, so Hermes can use Synthesize Bio MCP's 5 tools, like get_metadata_schema, resolve_sample_metadata and analyze_gene_expression, without your Synthesize Bio MCP credentials ever reaching the app.
Synthesize Bio MCP signs in through managed OAuth. MewCP stores the token encrypted, refreshes it before it expires and injects it only at call time. Your agent never holds it, so it can never leak it.
More apps
Manage courses, assignments, rosters and student activity within Google Classroom. Create learning workflows, track submissions and organize educational content across classrooms.
AI's impact on real-world sectors, as a structured record for agents: source-linked headlines placed in one of 18 sectors (healthcare, law, retail, government, defense and more), with outlet, country and time, since June 2026
PopHIVE provides AI agents with access to community-level U.S. public health data, enabling them to explore disease trends, compare health indicators across states and counties, analyze vaccination rates and chronic conditions, and investigate social determinants of health using data from multiple sources.
Search Microsoft Learn documentation and training content through MCP, helping AI agents find relevant technical guidance, tutorials, modules, and learning resources.
Other clients
Synthesize Bio MCP works the same way in every client below. Pick yours for its own guide.
Free to start.
Scope access to the exact Synthesize Bio MCP tools and accounts you allow, work and personal side by side. Anything outside that scope is refused at call time, even when the model asks for it.
Each call is logged: what ran, on which account, and whether it worked. Failures come back structured, so the agent knows whether to retry instead of guessing.
Every app sits behind one secure MCP URL, and there is no server for you to deploy or keep running. Add an integration in the dashboard and it is live on the next call, with no config to change and no new credentials to wire up. Tools load only when a task needs them, so your agent stays fast and focused however many apps you connect.
Edit config.yaml in your Hermes config directory (~/.hermes).
Put this under the top-level "mcp_servers:" key.
Run /reload-mcp in Hermes, or start a fresh session. Sign in when the browser window opens.
Start a new chat and ask: "List the tools you can use from MewCP." You should see your connected apps.
Returns the structured-metadata schema used to turn a natural-language experiment description into sample groups. Response fields include `group_schema` (the JSON shape of one sample group), the `fields` and `perturbation_fields` inventory for the active metadata version, and `instructions` for building the `groups` array. The resulting `groups` array is the input to resolve_sample_metadata. This tool makes no AI calls and consumes no usage budget.
Deterministically harmonizes already-structured sample `groups` to Synthesize Bio's controlled ontology vocabulary. Does not accept a natural-language prompt and makes no AI calls. Input: `groups` — a JSON array of sample-group objects matching get_metadata_schema's `group_schema`. Returns a resolution_id and a per-group breakdown including tissue, disease, cell type/line (with resolved ontology ids), sex, age, and full perturbation details (type, label/id, gene mechanism and mechanism type, dose, dose count, timepoint). Each resolved field also carries ranked candidate options for disambiguation. If a value resolves to the wrong id, correcting the term in `groups` and calling again produces a new resolution. The `warnings` array flags issues such as a described drug that did not match the ontology. When status is `resolving`, the same resolution_id identifies the pending request and can be passed again to poll. A completed resolution_id is required by analyze_gene_expression. Optional workspace_id selects which workspace the resolution is stored under. When omitted, a single-workspace account is assigned automatically. When the account has more than one workspace and workspace_id is omitted, the call is rejected with failure_kind `workspace_selection_required` and a `workspaces` list of `{ name, workspace_id }`; no metadata resolution runs until the user picks and the tool is retried with workspace_id. Quota and monthly-limit errors include a request-higher-limits URL and indicate remaining samples/cells when available.
Starts a differential gene expression analysis using Synthesize Bio's AI platform. Requires the resolution_id returned by resolve_sample_metadata; raw natural-language prompts are not accepted. Requires `user_confirmed_metadata: true`. When the flag is missing or false, the call is rejected with failure_kind `user_confirmation_required`. Optional workspace_id selects which workspace owns the generated dataset. When omitted, the workspace from resolve_sample_metadata is used automatically. When the account has more than one workspace and the resolution has no workspace, the call is rejected with failure_kind `workspace_selection_required` and a `workspaces` list of `{ name, workspace_id }`. Returns a job_id immediately; get_analysis_results accepts that job_id and returns analysis status or results. The pipeline runs two steps: (1) GEM-1 — Synthesize Bio's Gene Expression Model inference; (2) Differential expression — GPU-accelerated DESeq2 (negative-binomial GLM with Wald test, Cook's outlier filter, and Benjamini-Hochberg padj). All genes are tested; pre-filtering is handled by DESeq2's independent filtering. If the query is unsupported, later polling responses include failure_kind `unsupported_query` and suggested_queries. Quota and monthly-limit errors include a request-higher-limits URL; a previous successful resolve does not grant an extra run when the account is out of budget.
Polls the status of a gene expression analysis. Each call waits server-side for a short bounded window and may return earlier if progress is detected. Responses always include a `structuredContent` object (declared by the tool's `outputSchema`); MCP clients read from `structuredContent` directly rather than re-parsing JSON out of the human-readable text. `structuredContent` always has `status` (one of `running`, `complete`, `failed`), `job_id`, and `steps_completed`. While running, it also has `step` (`gem_model` or `diff_expr`), `message`, and `progress_label`/`progress_percent`/`progress_bar`. Failed responses include `error`, and may also include `failure_kind`, `user_action_required`, and `suggested_queries`. When `status` is `complete`, `structuredContent` carries: `metadata` (prompt, modality, groups, plus summary counts such as `significant_genes`, `significant_up`, `significant_down`, `total_genes_tested`); `results` — up to 1000 differential expression rows (each with `gene_id`, `gene_symbol`, `log2FoldChange`, `pvalue`, `padj`, `neg_log10_padj` (pre-computed `-log10(padj)`, clamped to 300 if padj underflows), `direction`, `significant`) suitable for downstream analysis or visualization (e.g. a volcano plot with x = `log2FoldChange`, y = `neg_log10_padj`); `plot_results` — the top ~200 most significant rows (same per-row shape, pre-sorted most-significant-first), pre-sliced for charting; the full `results` array is better suited to tables, summaries, and analysis; `results_returned` and `results_total` for truncation accounting; `plot_results_returned` for the plotted subset size; `dataset_link` — `{ dataset_id, title, url }` for the Synthesize Bio platform dataset (or `null`). The accompanying `content[0].text` is a human-readable Markdown summary of the same data. For hosts that do not surface `structuredContent` (e.g. claude.ai), it inlines only the top ~200 most significant rows as an array of objects under a top-level `results` key — same per-row schema as `structuredContent.results`, including the pre-computed `neg_log10_padj` field — so chart-widget code can use those rows directly. The full result set remains available via `structuredContent.results` and the dataset link when present.
Returns a presigned S3 URL to download the raw gene expression counts data (JSON) produced by a completed or in-progress analysis job. Supported only in Claude Code or Codex with control of local terminal sessions; chat clients and hosted/browser-based code sandboxes are unsupported, even with Python or shell tools. The data is typically large (20,000+ genes by N samples) and is intended for local processing. Raw counts are optional for custom computations; DE results and contrast labels are available separately from get_analysis_results. The JSON has the following structure: { gene_order: string[] (Ensembl IDs), outputs: [{ counts: number[], metadata: object }], model_version: string }. Each entry in 'outputs' corresponds to one sample; 'counts' is aligned with 'gene_order'. The response also returns a second presigned URL to a small (~500 KB) Parquet file mapping every Ensembl `gene_id` in `gene_order` to its HGNC gene_name — both files join on gene_id to label genes by symbol. Both URLs expire after 1 hour; fresh responses contain fresh URLs. Available after analyze_gene_expression has completed the GEM-1 step.
