Agents and project tooling
Three separate projects, each with its own repository. None is required to use PIASO. They are listed here because they address the same practical problem from different angles: a single-cell analysis runs for weeks, and increasingly some of it is driven by a coding agent that has no memory of last week.
PIASO-for-agents
https://github.com/genecell/PIASO-for-agents
The PIASO ecosystem in a form a coding agent can consume: Claude skills, Cursor
rules, an llms.txt, and an MCP server.
The reason this exists is narrow and worth stating. An agent that has not read the current API writes plausible code against the API it remembers, and plausible-but-wrong single-cell code does not crash — it produces a UMAP. Three API mismatches in this documentation reached published tutorial pages before anyone noticed, all of the same kind: a function that had gained a keyword, a return value that had become a tuple, a default that had flipped. Every one would have been caught by an agent reading the current signature.
Install it alongside PIASO and the agent has the signatures, the defaults and the worked examples in front of it rather than in its weights.
stato
Structured expertise management for long computational projects. Skills, parameters and lessons learned live in validated Python modules rather than in a chat history, so a project’s accumulated knowledge — which parameter mattered, which approach failed and why — survives a context window ending, a break of several weeks, or a handover to someone else.
The failure mode it addresses is specific to long analyses: the reason a threshold is 0.05 and not 0.01 was established once, in a conversation, and by the time it matters again nobody can reconstruct it. Writing it down as structured, validated state means the next session starts from what was learned rather than from the code alone.
PlanDrop
Plan, review and run Claude Code on remote machines. Single-cell work runs where the data is — a cluster, a workstation, a compute node — and Claude Code runs there too, in a terminal you have to be sitting at. PlanDrop is the way to drive those sessions from somewhere else — a Chrome browser, a Mac laptop, an iPhone: write the plan, review what it intends to do, send it to the machine, and read the result back.
The reviewable-plan part is the point. An agent with a shell on the machine holding your data can do a great deal between the moment you ask and the moment you look, and a plan approved beforehand is a better checkpoint than a transcript read afterwards.
These are independent projects with independent release cycles. Follow the links for installation, documentation and current status.