pandaomics — скилл для Claude Code: что делает и как установить
Скилл для Claude Code обращается к платформе PandaOmics через PLSI, когда базовый анализ мишеней уже собран. Запрашивает ранжирование генов, гранты и списки экспертов по токену пользователя.
MIT ★ 93 обновлён 2026-11-15
npx skills add Insilico-org/plsi --skill pandaomics -a claude-codeЧто делает pandaomics
Скилл направляет запросы к PandaOmics и PandaClaw через утилиту plsi oracle, дополняя первичный открытый анализ биомишеней. Агент подключает его только после завершения бесплатной оценки, если в окружении задан токен доступа.
Для работы нужен персональный MCP-токен PANDAOMICS_BEARER_TOKEN либо PANDACLAW_BEARER_TOKEN для агента PandaClaw. Обычный ключ REST API не подойдёт и приведёт к ошибке авторизации.
Каждый платный вызов списывается из дневного лимита PLSI_ORACLE_DAILY_CAP, который по умолчанию равен 25 запросам. Полученные данные скилл сохраняет в секцию platform_reads без превращения оценок в цитаты.
Какие задачи решает
- получить независимый рейтинг генов-кандидатов для заболевания
- выгрузить данные по грантам и финансированию исследований мишени
- найти ключевых лидеров мнений (KOL) по гену или болезни
- ранжировать показания для выбранной биомишени
Примеры запросов агенту
- Запроси через pandaomics список топ-мишеней для болезни с disease_id 42
- Объясни ранг гена STAT3 в PandaOmics для текущего заболевания
- Проверь финансирование и гранты по гену через инструмент get_grants
Как установить pandaomics
1Через Skills CLI (Vercel)
npx skills add Insilico-org/plsi --skill pandaomics -a claude-code2Вручную: положить файлы
~/.claude/skills/pandaomics/SKILL.mdПапка скилла целиком; для одного проекта — .claude/skills/pandaomics/
1Через Skills CLI (Vercel)
npx skills add Insilico-org/plsi --skill pandaomics -a codex2Вручную: положить файлы
.agents/skills/pandaomics/SKILL.mdДля всех проектов — ~/.agents/skills/
1Через Skills CLI (Vercel)
npx skills add Insilico-org/plsi --skill pandaomics -a cursor2Вручную: положить файлы
.cursor/skills/pandaomics/SKILL.mdCursor читает и .agents/skills/, и .claude/skills/
1GitHub CLI
gh skill install Insilico-org/plsi skills/targets/pandaomics2Вручную: положить файлы
.github/skills/pandaomics/SKILL.md1Через Skills CLI (Vercel)
npx skills add Insilico-org/plsi --skill pandaomics -a gemini-cli2Командой платформы
gemini skills install https://github.com/Insilico-org/plsi --path skills/targets/pandaomicsВручную: положить файлы
.devin/skills/pandaomics/SKILL.mdСтарый путь .windsurf/skills/ тоже читается
Claude.ai
Упакуйте папку в zip и загрузите в Settings → Capabilities → Skills
По шагам, с командами для каждой платформы: Как установить скилл в Claude Code: пять способов
Вопросы о pandaomics
Как установить pandaomics в Claude Code?
Выполните команду npx skills add Insilico-org/plsi --skill pandaomics -a claude-code.
Что делает pandaomics?
Скилл позволяет вызывать инструменты PandaOmics через plsi oracle, запрашивать скоринг мишеней, гранты и списки экспертов.
Что нужно для работы pandaomics?
Нужен собственный MCP-токен PANDAOMICS_BEARER_TOKEN или PANDACLAW_BEARER_TOKEN в .env, утилита plsi и завершённый бесплатный этап оценки мишеней.
Исходный текст
SKILL.md · MIT
---
name: pandaomics
description: Use as PLSI Target Hunter on demand, only when the user has set their own PandaOmics or PandaClaw token and the free evaluation is already finished - which of the platform's tools answers what, what a platform score may and may not support, the daily cap, and where the answer is recorded.
---
# Asking a commercial platform
Craft, read when needed, not on
every run.
Nothing in target work needs PandaOmics: every question in `targets/evaluate` is answered by free sources, and a
dossier gathered without a token is complete. This page is for a user who already has an account. Read it only when
the dossier's `keyed.unlocked` names a platform; otherwise there is nothing to do here.
## Tokens
Both platforms are bring-your-own-token: the user sets their own token in their environment or `.env`, and PLSI
ships no key.
- `PANDAOMICS_BEARER_TOKEN` (and `PANDAOMICS_MCP_URL` when the account is on another host) opens the PandaOmics MCP
endpoint; every tool below goes through it. It is an MCP personal access token, not the REST API token: if
`plsi oracle pandaomics tools` works and a call fails with an authentication error, the token is the wrong kind.
- `PANDACLAW_BEARER_TOKEN` opens PandaClaw.
When a token is missing, say so in the `result`; PLSI helps the user set it (`plsi keys`).
| | Returns | Recorded in | May support |
|---|---|---|---|
| PandaOmics (`plsi oracle pandaomics`) | ranked target lists, a gene's rank with its scores, indications, grants, KOLs | `platform_reads` | nothing as a citation |
| PandaClaw (`plsi oracle pandaclaw`) | prose from an agent | `platform_reads` | nothing as a citation; every identifier in it must verify |
## Four rules
1. **The free path is the floor.** Finish the evaluation first; if the assessment could not be written without the
platform, the evaluation is not done.
2. **A platform and a curated record do not outrank each other.** Where they differ (a development level, say),
report both and say which one the question needs.
3. **A score is not a citation.** It carries no PMID, NCT or accession, so it cannot carry a claim. The `llm_*`
scorings are a language model's judgment arriving as a number. Its scores have no published range or direction:
write the number with the score's own name, never a word like "high".
4. **Every billed call counts against a daily cap** per platform (`PLSI_ORACLE_DAILY_CAP`, default 25), refused before
anything is sent. Catalogue listings (`plsi oracle pandaomics tools`, `plsi oracle pandaclaw-skills`) are not runs
and are not counted. `keyed.unlocked[].today` shows what is left.
## Which tool answers what
```bash
plsi oracle pandaomics tools # the live catalogue: check names and arguments first
plsi oracle pandaomics call --tool <name> --file args.json
```
| Tool | Ask it | Why the free path cannot |
|---|---|---|
| `resolve_disease` | `{"query": "<name>"}` -> the platform's `disease_id` | the tools that take a disease need it |
| `resolve_gene_symbol` | `{"query": "<symbol or alias>"}` -> the platform's gene entity | the same, for a gene |
| `explain_target_rank` | `{"disease_id": <id>, "gene": "<SYMBOL>"}` -> why the gene ranks where it does | an independent ranking with its reasoning, to set beside Open Targets |
| `get_target_ids` | `{"disease_id": <id>, "top_n": <n>}` -> the top-ranked genes | a second shortlist beside `targets discover` |
| `get_indication_prioritization` | `{"gene": "<SYMBOL>"}` -> ranked indications | another source's ordering of the dossier's `indications` |
| `get_grants` | `{"gene": "<SYMBOL>"}` or `{"disease_id": <id>}` | current research funding; no free source here has it |
| `get_kols` | `{"gene": "<SYMBOL>"}` or `{"disease_id": <id>}` | who works on this |
The catalogue can change; trust `tools` over this table, and use the short names it lists. Resolve a disease or gene
once and reuse the id. On a shortlist, ask `get_target_ids` once for the disease and read every candidate from it;
`explain_target_rank` is the follow-up for the two or three worth understanding. Never spend a call on what the
dossier already holds (structures, potencies, residues, trials, aging curation, direction of effect).
## Reading the scores
- **Evidence-derived** (`knockout`, `knockin`, `mutation`, `pathway`, `ppi`, `transcriptomics`, `bayes_gwas`): a
second opinion on what the free sources cover.
- **Attention-derived** (`target_metric`, `funding`, `grant_size`, `high_impact_total`, `trend`, `dl_burst`): how much
the field has looked at a gene, a competition fact, not quality.
- **Model-derived** (every `llm_*`): report as a model's judgment or not at all.
A rank belongs to a disease; record the question with it.
## Recording it
`platform_reads` on the assessment is the only place a platform answer goes:
```json
"platform_reads": [
{"source": "pandaomics", "tool": "explain_target_rank",
"asked": "rank of <SYMBOL> for <disease> (disease_id <id>)",
"says": "<what came back, in the platform's own terms>"}
]
```
`source` is `pandaomics` or `pandaclaw`. `says` keeps the platform's terms; never turn a score into a verdict.
Refusal 7 rejects an entry with no billed call to that platform (for PandaOmics, that tool) in the egress
ledger in the last thirty days. In the target file these go on the
`Model output` line, `model_derived`.
## PandaClaw
`plsi oracle pandaclaw` is an agent that reads literature and pipeline pages and writes prose; use it for the
landscape question `targets/landscape` leaves to you. Name a skill from `plsi oracle pandaclaw-skills` in `ask.json`
(without one it answers fast and shallow) and keep the whole answer with
`--out <PLSI_HOME>/work/<thread>/pandaclaw_<topic>_<n>.md`. Every identifier in it must pass `plsi evidence verify`
before you cite it, and every asset it names goes through `plsi targets trials` before `harvested_assets`.
## Before you spend a call
- [ ] The free evaluation is finished.
- [ ] `keyed.unlocked` names the platform and `today.remaining` is above zero.
- [ ] The free sources cannot answer the question.
- [ ] On a shortlist, one disease-level call, not one per gene.
- [ ] The answer goes in `platform_reads` with `asked`, not into the prose.
- [ ] Where it disagrees with a curated record, both are reported.
© Insilico-org · MIT · Insilico-org/plsi