| name | seqeron-setup |
| description | One-time install & configuration for a freshly cloned Seqeron repo, so the bioinformatics skills are ready to solve real tasks. Use when the user says "install and configure", "set up Seqeron", "get me started", "onboard me", "prepare the repo", "why aren't the bio tools working", or asks how to make the Seqeron MCP tools / skills available. Verifies the .NET 10 SDK and Python are present, builds all 11 MCP servers into the on-demand cache (so the first real tool call is instant), and runs a live smoke test — without registering MCP anywhere (tool schemas stay out of the model's context; skills pull tools in on demand). Run this once per clone/machine; after it, just describe a biology task. |
| allowed-tools | Bash, Read |
Seqeron Setup
The one command a new user runs after cloning: build everything, verify the
path, then get out of the way. After it, the user describes a biology task in
plain language and the matching domain skill (bio-qc, bio-alignment,
bio-annotation, …) loads itself and runs the real, validated algorithms.
The model this sets up
- The 20 skills live in
.claude/skills/ (Claude Code) and a mirror
in .github/skills/ (Copilot / VS Code). Claude Code auto-discovers them —
no install step. They are a thin routing + rigor layer, not the algorithms.
- The algorithms run in the 11 shipped MCP servers (
Seqeron.Mcp.* +
SuffixTree.Mcp.Core, 427 tools). Skills call them on demand — spawn the
one server a task needs, call it, tear it down — via the stdlib client
scripts/skills/seqeron_mcp_client.py.
- No MCP registration. We deliberately do not write a
.mcp.json or run
claude mcp add. Registering all 427 tools would load every schema into the
model's context and drown it. On-demand keeps context lean and lets skills pull
in only what the current task needs. The only thing that must happen up
front is the build, so the first on-demand call is instant.
So "setup" = check the toolchain, build all servers, prove one real call works.
When to use
- A freshly cloned repo, or a new machine, before the first biology task.
- The user says "install and configure", "set up Seqeron", "onboard me".
- A bio task just failed with a build/
dotnet/server-launch error → re-run this;
it's idempotent and will rebuild what's stale.
Do not run it before every task — once per clone/machine is enough.
How to run
Run the idempotent orchestrator from the repo root:
scripts/setup.sh
It (1) checks the .NET 10 SDK and Python 3 are on PATH, (2) builds all 11 MCP
servers into the same temp cache the on-demand client uses, and (3) runs a live
gc_content smoke test end-to-end. Add --full to also start each server and
list its tools (slower, but confirms every server actually launches):
scripts/setup.sh --full
First run compiles everything (a few minutes). Later runs are cached and fast.
Reading the result
- Ready. → done. Tell the user they can now just describe a biology task; the
right skill will load and compute with real numbers. Offer a first example
(e.g. "Read this FASTA and report GC% and any EcoRI sites").
dotnet not found / wrong version → point them to the .NET 10 SDK
(https://dotnet.microsoft.com/download); the repo targets .NET 10.
python3 not found → the on-demand client needs Python 3 (stdlib only,
3.9+). Any system Python 3 works.
- A server failed to build → surface the compiler output the script printed;
that's a real build error in that server's project, not a setup bug.
After setup
Nothing else to configure. Point the user at a real task and let the domain
skills route it. If they want to know which tool does something without loading
schemas, that's seqeron-discovery; for the
scientific-honesty rules every result must follow, that's
bio-rigor.