External genomics analyses and MCP: parameters and notes

See the feature guide for purposes, methods and examples, and the source FAQ for papers, authors and license evidence.

Discovery and catalog status

search_projects and list_analyses accept optional query and domain filters; all query terms must match. describe_analysis(analysis) accepts a registered analysis ID and returns input keys, parameter types and ranges, installation instructions, outputs and sources.

reference_only means a discovered project without a registered project-specific interface. For available, check whether the record represents a native function, executable adapter, MCP bridge or bundled skill. Inclusion does not mean all upstream methods were implemented or scientifically validated. Initial gap fields describe the repository before this expansion; remaining_gaps describes current limits.

External analysis parameters and execution

Prepare the software, references, model weights and credentials required by the selected analysis separately. Harmonize reference builds, sample identifiers and allele orientation. Input prefixes must refer to complete file sets, such as PLINK BED/BIM/FAM.

Parameter	Contract
analysis	Required; an ID returned by list_analyses.
inputs	Required; mapping with exactly the keys declared by describe_analysis, and file paths or prefixes as values.
output_directory	Required; a directory that does not yet exist.
parameters	Optional; registered parameters only. Defaults and ranges come from describe_analysis; arbitrary command-line arguments are not accepted.
threads	Default 1; integer in 1–1024.
executable_paths	Optional program-to-path overrides, such as {"plink2": "/path/to/plink2"}.
timeout_seconds	Execution only; default 3600.0, range (0, 86400] seconds.
max_artifact_bytes	Execution only; default 1000000000, limiting each artifact file in bytes.
verbose	Default True; controls runtime messages and progress display.

plan_analysis validates inputs and builds commands without writing files; missing executables remain explicit in the plan. run_analysis requires the backend and a new output directory. It saves commands, inputs, parameters, sources, logs and dnakit-run.json, including failed runs. A completed process/file contract still requires scientific QC.

The four PLINK adapters were exercised with real PLINK 2 and deterministic synthetic data. Each additional adapter records its interface and synthetic-backend execution status in validation; see the 335-project review. Valid empty FastANI/Barrnap/Cas-OFFinder results are explicitly allowed; other required empty outputs fail. modkit-pileup requires a pre-existing FASTA .fai.

MCP installation and connection parameters

Install MCP client support in a source checkout:

python -m pip install -e ".[agent]"

Install and configure each third-party service separately from its upstream instructions. list_mcp_servers returns installation and launch recipes without reading credentials or starting services; the feature guide retains the per-service catalog. The client does not install servers, and catalog inclusion does not establish local availability or scientific validation.

Parameter	Contract
server_id	Required; a service catalog ID.
tool_name	Required for a single tool call; obtained from the live tool list.
arguments	Required for a single tool call; a JSON object matching the tool's inputSchema.
transport	Default "stdio"; "http" uses only cataloged author-documented endpoints.
executable_path	Optional path to the separately installed server executable or interpreter.
workdir	Optional working directory, commonly required for source checkouts.
environment_variables	Optional mapping from server variable names to existing local environment variable names, never credential values.
timeout_seconds	Default 60.0 seconds; request or complete workflow session timeout.
max_response_bytes	Default 4194304; response limit after SDK decoding.
max_tools	Default 5000; maximum tool list size.
verbose	Default True; controls runtime messages and progress display.

list_mcp_tools starts a stdio service or connects to a registered HTTP service to obtain live schemas. call_mcp_tool validates arguments before invocation; remote isError is retained as remote_error, and known forwarded credential values are redacted. Each standalone call opens and closes one session, so in-memory state does not persist between calls. Response size is checked after SDK decoding, not as a wire-level traffic bound.

Tests exercise real local stdio discovery, argument validation, session state and failure handling; six author-hosted HTTP services also passed live tool discovery in the supplemental review. See the review FAQ for the limits of scientific execution, account and model-environment validation.

Calls within one session

call_mcp_workflow requires server_id and calls, with the same connection options as above. calls contains 1–100 entries, each with exactly tool_name and arguments. Arguments are explicit: no scripts run and no previous output is substituted automatically.

Schemas are refreshed and validated before every step. A remote error stops execution and returns all executed steps, including the failed step. Configuration and connection errors raise exceptions. Earlier uploads or computation are not rolled back. timeout_seconds bounds the entire session, and the response size limit applies to accumulated decoded results. Consult the selected tools for network, filesystem and compute effects.

DNAKit MCP entry point

After installing client support, start the unified server:

dnakit-mcp

Search genomics or integrations with search_dnakit_tools, then inspect full signatures and schemas with describe_dnakit_tool. External analysis and remote MCP calls have the external_process effect and use the existing allow_side_effects=true invocation mechanism.

Full API signatures

dnakit.integrations

Reviewed external genomics analyses with optional, caller-installed backends.

call_mcp_tool(server_id, tool_name, arguments, *, transport='stdio', executable_path=None, workdir=None, environment_variables=None, timeout_seconds=60.0, max_response_bytes=4194304, max_tools=5000, verbose=True)

Validate against a server's live inputSchema, then invoke exactly one tool.

Tool effects belong to the caller-selected upstream tool: external services may read/write files, contact networks or load models. Returned content and structuredContent retain the MCP isError flag; remote errors are explicitly marked remote_error. Known forwarded credential values are redacted. Each call starts a fresh session, so upstream in-memory state is not retained.

call_mcp_workflow(server_id, calls, *, transport='stdio', executable_path=None, workdir=None, environment_variables=None, timeout_seconds=60.0, max_response_bytes=4194304, max_tools=5000, verbose=True)

Run explicit tool calls sequentially within one retained MCP session.

Each call is {"tool_name": str, "arguments": dict}; 1 to 100 calls are allowed. Schemas are retrieved and validated before every step. Server in-memory state persists between steps, supporting model loading followed by prediction. Calls use explicit arguments; output substitution and arbitrary scripts are not run. A remote isError stops the workflow and returns all executed steps, including the error. Configuration/connection errors raise; earlier effects cannot be rolled back. timeout_seconds bounds the complete session, and the response byte limit applies to accumulated decoded step results.

describe_analysis(analysis, *, verbose=True)

Describe inputs, parameters, dependencies, outputs and sources without execution.

list_analyses(query='', *, domain=None, verbose=True)

Search available executable adapters; availability does not imply installation.

list_mcp_servers(*, verbose=True)

Return source-backed connection recipes without importing SDKs or starting services.

Catalog inclusion confirms an upstream startup recipe, not local availability, server health or scientific validation. No credentials are read by this call.

list_mcp_tools(server_id, *, transport='stdio', executable_path=None, workdir=None, environment_variables=None, timeout_seconds=60.0, max_response_bytes=4194304, max_tools=5000, verbose=True)

Connect to one catalog server and retrieve its live tool names and schemas.

Stdio launches the separately installed executable; source servers require workdir. No installer or shell is used. environment_variables maps supported server variable names to existing environment variable names, never values. Only explicitly requested runtime variables (e.g. PYTHONPATH) are forwarded. HTTP is supported only for author-documented catalog endpoints. Response byte limits are checked after SDK decoding, not at the transport wire layer. Validated local description revisions apply only to this listing; upstream invocation schemas remain authoritative.

plan_analysis(analysis, inputs, output_directory, *, parameters=None, threads=1, executable_paths=None, verbose=True)

Validate an analysis and return exact argument vectors without writing files.

Input prefixes refer to complete file sets, e.g. BED/BIM/FAM. Paths are resolved before changing the working directory. Missing executables remain explicit in the plan; running requires a separately installed backend.

run_analysis(analysis, inputs, output_directory, *, parameters=None, threads=1, executable_paths=None, timeout_seconds=3600.0, max_artifact_bytes=1000000000, verbose=True)

Run an explicitly selected scientific backend in a fresh output directory.

No packages, databases, model weights or credentials are installed or fetched. Results include original sources, exact commands and persistent output paths. Process success verifies produced artifacts, not scientific validity.

search_projects(query='', *, domain=None, verbose=True)

Search the dated evidence survey, including projects not integrated into DNAKit.

Project entries distinguish skills, MCP servers, software, models and papers. A search hit is evidence of discovery, not a claim of local execution.