GenomeOps
user7401/GenomeOps
Executa pipelines nf-core via IA com aprovação humana em cada etapa
Ver no GitHubMCP server que integra pipelines de bioinformática nf-core/Nextflow com agentes de IA, permitindo descoberta, configuração e execução de análises genômicas com controle humano em cada etapa. Converte o ecossistema nf-core em ferramentas estruturadas consumíveis por Claude e outros agentes, eliminando alucinações sobre conhecimento específico de bioinformática.
Instalação
npx convez-ai user7401/GenomeOps1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Descoberta automática de pipelines nf-core adequados para análises genômicas
- Geração e validação de samplesheets sem conhecimento prévio de bioinformática
- Execução de análises Nextflow com interface amigável via agentes de IA
- Automação de workflows bioinformáticos complexos com segurança
Veredito

Ferramenta especializada e bem-arquitetada para democratizar pipelines nf-core via IA, com modelo HITL robusto que mantém controle humano sobre execução real. Ainda em desenvolvimento pré-release, mas resolve um nicho crítico: análises genômicas reproducíveis sem expertise técnica prévia.
README
A tradução para o português deste README ainda está na fila.
GenomeOps
⚠️ Status: developmental / pre-release. This project is under active development. Tool names, schemas, and behavior may change without notice. Execution tools run real subprocesses (Nextflow, container engines) and can consume real compute — review every gated output carefully before approving anything, and expect rough edges.
Unofficial, community-built project. Not affiliated with, endorsed by, or maintained by the nf-core community or Seqera/Nextflow.
GenomeOps is an MCP (Model Context Protocol) server that exposes the nf-core / Nextflow bioinformatics pipeline ecosystem as structured, agent-consumable tools. Any AI agent — Claude, GPT, Cursor — can go from "I have these files and this goal" all the way through pipeline discovery, samplesheet generation, parameter configuration, and launching and monitoring an actual analysis run, without hallucinating nf-core-specific knowledge.
The server is built around a human-in-the-loop (HITL) model: every step that involves a judgment call, a cost, or an irreversible action returns a structured "review required" signal so a human stays in control of what actually runs.
Quickstart
# Run directly with uvx (no install required)
uvx genomeops-mcp
# Or install with uv
uv pip install genomeops-mcp
genomeops-mcp
Add to Claude Desktop
Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):
{
"mcpServers": {
"genomeops": {
"command": "uvx",
"args": ["genomeops-mcp"]
}
}
}
Add to Claude Code (CLI)
claude mcp add genomeops -- uvx genomeops-mcp
Or add to .claude/mcp.json in your project:
{
"mcpServers": {
"genomeops": {
"command": "uvx",
"args": ["genomeops-mcp"]
}
}
}
Enable AI parameter suggestions
export ANTHROPIC_API_KEY=sk-ant-...
uvx genomeops-mcp
Tool Reference
Discovery & matching
| Tool | Inputs | Outputs |
|---|---|---|
check_feasibility | goal: str, file_paths: list[str] | Reads actual file content (magic bytes, gzip, first lines) via an LLM to identify formats — even with scrambled filenames — and matches them + the stated goal against candidate pipelines, surfacing gaps |
list_pipelines | topic?: str | List of pipelines with name, description, version |
get_pipeline_info | pipeline_name: str, version?: str | Full metadata, profiles, schema URLs |
get_latest_version | pipeline_name: str | Latest stable release tag; flags dev/pre-releases (is_dev) so agents stop and ask before using one |
Samplesheet
| Tool | Inputs | Outputs |
|---|---|---|
get_samplesheet_schema | pipeline_name: str, version?: str | Column definitions, types, allowed values |
generate_samplesheet | pipeline_name: str, file_paths: list[str] | CSV string + review warnings |
validate_samplesheet | pipeline_name: str, samplesheet_content: str | valid: bool, plain-English errors with row numbers |
Parameters
| Tool | Inputs | Outputs |
|---|---|---|
analyze_pipeline_schema | pipeline_name: str, version: str | Maps the pipeline's full decision tree — every branch point and parameter classified by an LLM into tiers (provided_input → expert_required), with extracted expert guidance and conditional dependencies. Cached 24h per version. |
configure_parameters | pipeline_name, version, data_summary?, experiment_description?, user_choices?, mode? | Resolves non-obvious parameters: auto-derives what it can, lets the user set any subset themselves via user_choices, and (in mode="expert") uses an LLM to propose context-dependent values — while always escalating expert_required parameters back to the human rather than guessing |
get_parameters | pipeline_name: str, version?: str, group?: str | Grouped parameter definitions |
suggest_parameters | pipeline_name: str, experiment_description: str | AI-suggested params + justifications |
Launch & execution
| Tool | Inputs | Outputs |
|---|---|---|
generate_launch_command | pipeline_name, version, profile, samplesheet_path, outdir, params? | Validated nextflow run command string |
check_execution_environment | profile?: str | Probes for Nextflow, Java, and container/conda engines (Docker, Singularity, Apptainer, Podman, Conda, Mamba), checks the Docker daemon, recommends a viable -profile, and reports what's missing with install hints |
setup_environment | pipeline_name, version, profile | Runs nextflow pull + nextflow config to cache the pipeline and resolve its configuration as a dry run, before any compute is spent |
estimate_resources | pipeline_name, version, samplesheet_path?, file_paths?, data_summary?, profile? | Sizes the job against this machine before launch: derives the workload (sample count + input size) and, with the pipeline's resource profile, proposes CPU/memory/disk/walltime, returns a verdict (sufficient/marginal/insufficient), and suggests --max_* overrides. LLM-refined when ANTHROPIC_API_KEY is set; otherwise heuristic + raw signals returned for the host agent to reason over |
run_pipeline | pipeline_name, version, profile, samplesheet_path, outdir, params?, run_name?, confirm: bool | Gated. With confirm=false (default) it only previews the validated command and changes nothing. It launches a detached background nextflow run only when called with confirm=true — which must follow explicit human approval of the preview. Parameters are written to a -params-file (more robust than inline flags), and a full provenance manifest is recorded |
get_run_status | run_name: str | Polls a launched run: process liveness, log tail, and a parsed status (running / completed / failed / unknown) with suggested next steps |
list_runs | — | Lists every run launched through the server (newest first), refreshing liveness and resolving terminal status from the log |
stop_pipeline | run_name, confirm: bool | Gated like run_pipeline: previews unless confirm=true, then sends SIGTERM to the detached process group. The run can be continued later with resume=true |
Diagnostics & provenance
| Tool | Inputs | Outputs |
|---|---|---|
diagnose_run_failure | run_name: str | For a failed run, locates the failed process and its work dir, reads .command.err/.command.sh/.exitcode, classifies the root cause (out of memory, time limit, disk, container/engine, missing input, tool error) and proposes concrete fixes. LLM-refined when ANTHROPIC_API_KEY is set, heuristic otherwise |
diagnose_resume | run_name: str | Predicts whether -resume will actually reuse cached work: checks the .nextflow cache and work/ dir exist and whether any recorded input file changed (path/mtime/size) since launch — the common silent causes of a full re-run |
generate_methods_note | run_name: str | Drafts a citable Methods paragraph from the recorded provenance (pipeline, pinned version, profile, exact non-default parameters, tool versions) with canonical nf-core + Nextflow citations and a pipeline Zenodo DOI pointer. Deterministic and auditable |
Results
| Tool | Inputs | Outputs |
|---|---|---|
inventory_results | results_dir: str | Walks the output tree and classifies every file by format and semantic role (alignment, variants, gene_counts, qc_report, provenance, …) into a structured manifest grouped as primary_outputs, qc, provenance, indices, and unclassified. LLM-refined with one-sentence downstream hints when ANTHROPIC_API_KEY is set; heuristic otherwise |
parse_run_summary | results_dir: str | Per-sample QC (from MultiQC), flagged samples with reasons, and overall run status |
Typical Agent Workflow
check_feasibility(goal, file_paths) # match files+goal → candidate pipeline(s)
→ get_pipeline_info("rnaseq")
→ get_latest_version("rnaseq") # pin to a stable tag, never "master"
→ analyze_pipeline_schema("rnaseq", "3.14.0") # map every decision point & parameter
→ get_samplesheet_schema("rnaseq")
→ generate_samplesheet("rnaseq", ["/data/sample_R1.fastq.gz", ...])
→ validate_samplesheet("rnaseq", <csv>) # fix any errors
→ configure_parameters(...) # resolve params; ask human for expert_required
→ generate_launch_command("rnaseq", ...) # human reviews the command
→ check_execution_environment() # confirm Nextflow + a container engine
→ setup_environment("rnaseq", "3.14.0", "docker")
→ estimate_resources("rnaseq", "3.14.0", ...) # will this machine cope? size CPU/mem/disk/time
→ run_pipeline(..., confirm=False) # preview command + params
→ run_pipeline(..., confirm=True) # human has approved — launch for real
→ get_run_status("myrun") # poll until completed/failed
→ diagnose_run_failure("myrun") # IF failed: root cause + fixes
→ diagnose_resume("myrun") # before re-running with resume=true
→ inventory_results("./results") # what files were produced + downstream hints
→ parse_run_summary("./results") # QC once completed
→ generate_methods_note("myrun") # citable Methods paragraph
Human-in-the-Loop (HITL)
Tools that generate executable content, propose AI-derived values, or gate compute always signal that a human should be in the loop — via review_required: true, requires_confirmation: true, or an explicit confirm parameter:
check_feasibility— pipeline match is a recommendation, not a decisionget_latest_version— if the latest release is a dev/pre-release, the agent must stop and ask before using itanalyze_pipeline_schema— surfaces decision points (e.g. aligner choice) for the human to choose, or to delegate to expert modegenerate_samplesheet— strandedness and similar fields cannot be inferred reliablyconfigure_parameters—expert_requiredparameters are never silently set; they're always returned for a human decisionsuggest_parameters— AI suggestions need domain expert reviewgenerate_launch_command— pipelines consume real compute resourcesestimate_resources— a planning estimate, never a guarantee; aninsufficient/marginalverdict should be raised with the user before launchingrun_pipeline— hard-gated behindconfirm=true; withconfirm=falseit only previews and launches nothingstop_pipeline— hard-gated behindconfirm=true; previews what would be terminated otherwisediagnose_run_failure/diagnose_resume/generate_methods_note— return findings for human review, not automatic action
These are not blocking terminal prompts — they're structured signals in the JSON response. The calling agent is expected to surface them to the human, wait for a reply, and only then proceed.
Development
git clone https://github.com/user7401/nfcore-for-agents
cd nfcore-for-agents
uv sync --extra dev
uv run pytest -v
License
MIT