Hosted MCP server

Connect a Streamable HTTP client to https://mcp.genomicintelligence.ai/mcp. Anonymous calls use a shared, rate-limited demo principal. For own quotas and job access send Authorization: Bearer <GI_API_KEY> (or the documented X-GI-Key header). Local gi-mcp is a thin client: sequences used for predictions still travel to the hosted REST inference service.

Reviewed hosted discovery on 2026-10-01: 15 tools and six static resources. Initialize/retain the MCP session and let the host negotiate protocol support; the read-only probe negotiated 2025-11-25. Do not assume another server's protocol version or HTTP framing. The reviewed local package is gi-mcp 0.1.0a21 (Python >=3.10); source below a18 has a minus-strand expression-window one-base shift. Pin a reviewed release when using a local deployment.

Acquisition and handles

Acquisition returns a GI envelope **{data: {ref, name, length, ...}, meta}**. Use data.ref as the next tool's sequence_ref; it is not a top-level ref. At the MCP transport layer this is inside result.structuredContent on the reviewed server, also serialized in text content. Check MCP isError and any GI error object before reading a success payload: an HTTP 200 can be a tool failure. Hosted demo job listing, for example, returns isError: true and code: unavailable, outside the REST error-code enum.

Handles are server/session-local, in memory, TTL/capacity-limited, and disappear on restart or eviction. They are not durable sequence accessions. Preserve acquisition parameters and sequence provenance so a lost handle can be recreated. The hosted tool list excludes load_local_fasta; that tool is local-only. The handle summary contains only a sequence preview and no checksum. In this handle-only workflow verify returned length/TSS/strand/source metadata and tool errors; do not claim to have checked every base or computed a sequence hash. For exact byte-level provenance, separately acquire and hash the full reference bases from the recorded assembly/region, checking that source provenance matches.

Despite acquisition tool names, bases can come from the bundled cache, UCSC, or Ensembl. meta.source, meta.locus_source, meta.assembly, meta.catalog_release, and optional stale metadata reveal this; data.source may be only the acquisition channel. See sequence acquisition.

Prediction

predict_promoter, predict_splice, predict_enhancer, predict_chromatin, predict_expression each take exactly one of sequence_ref or sequence, plus optional model and sequence_name. With a handle, its stored name wins. Expression also needs description and, for longer-than-9,198-bp input, tss_index. These two semantic requirements are described by the tools even where JSON Schema does not mark the arguments required.

Current hosted prediction schemas do not expose REST options.threshold, site_types, format, or an async preference. Use REST when those controls are needed; do not invent corresponding MCP keyword arguments.

Annotation is find_genes, not predict_annotation. It takes a sequence or handle, not a region. find_genes_and_predict_expression likewise takes a sequence/handle plus required experimental description, with no region or tss_index argument. Both run async internally. wait=True (default) returns the result or a timeout error, not a detached job handle.

**The anonymous demo disables get_job, list_jobs, and wait=False.** With an own-key session, wait=False returns {data: {job_id, status: "submitted"}} to poll with get_job(job_id). This MCP status differs from REST submission's "accepted". list_jobs(limit=...) accepts 1–100 (default 20), whereas REST allows up to 500. Neither surface exposes pagination.

Discovery and resources

list_models(task) returns the flat {task, default_model, models} object. Discover and record the resolved model ID, organism, window and request bounds. Resources are gi://models, gi://docs/tasks, gi://openapi.json, gi://sequences, gi://jobs/recent, and gi://account. The account resource identifies demo/byok mode; recent jobs is unavailable in demo mode. A resource being listed does not prove that the current principal may read its content.

Worked sequence of tool calls

This is host pseudocode, not a Python SDK. Discovery/acquisition was exercised; the inference calls below are illustrative and consume inference quota.

h = fetch_region(region="chr11:5,225,000-5,235,000")
find_genes(sequence_ref=h.data.ref, wait=True)
find_genes_and_predict_expression(sequence_ref=h.data.ref,
    description="polyA plus RNA-seq; Homo sapiens K562", wait=True)

w = fetch_gene_for_expression(gene="HBB")
# Check w.data.length == 9198 and w.data.tss_source == "canonical-transcript".
# Retain w.meta assembly/source and w.data strand/TSS/region.
predict_expression(sequence_ref=w.data.ref,
    description="polyA plus RNA-seq; Homo sapiens K562")

# No inference: inspect/load the public positive-control sequence.
d = load_demo_sequence(name="promoter_tp53")
# Inference only when requested:
predict_promoter(sequence_ref=d.data.ref)