find-capability
/find-capability
Znajduje serwer MCP, gdy zadaniu brakuje narzędzia lub źródła danych, proponuje instalację i zgłasza wynik po zakończeniu.
Ocena i bezpieczeństwo
- Pliki w paczce
- 1
- Skrypty
- 0
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 11.11.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
Co robi i kiedy
Znajduje serwer MCP, gdy zadaniu brakuje narzędzia lub źródła danych, proponuje instalację i zgłasza wynik po zakończeniu.
Wyzwalacze z opisu
Oryginalny opis (EN)
Use when the task needs a tool, data source or integration the agent does not have (e.g. "read Linear issues", "extract text from a PDF", "query Postgres"). Finds MCP servers via lookupd, proposes an install for the user's runtime, and reports the outcome afterwards. Also use after finishing a task that used an MCP server, when lookupd's report_outcome tool is available, to report whether it worked.
Jak działa
- When to use
- Step 1: name the capability, not the request
- Step 2: look it up
- Step 3: propose, and wait for approval
- Step 4: report the outcome
- Rules
Podgląd SKILL.md
---
name: find-capability
description: Use when the task needs a tool, data source or integration the agent does not have (e.g. "read Linear issues", "extract text from a PDF", "query Postgres"). Finds MCP servers via lookupd, proposes an install for the user's runtime, and reports the outcome afterwards. Also use after finishing a task that used an MCP server, when lookupd's report_outcome tool is available, to report whether it worked.
---
# find-capability
lookupd is a neutral registry that agents query mid-task. Lookups are free. Rankings are
**Provisional** until real outcome evidence exists, so treat results as candidates, not verdicts.
## When to use
- The user's task needs a capability you lack: a SaaS API (Linear, Notion, Slack), a data
source (Postgres, a vector store), a file format (PDF, OCR), a browser, a cloud provider.
- You are about to say "I can't access X": check whether an MCP server for X exists first.
- A task just used an MCP server and `report_outcome` is in your tool list: go to Step 4.
This is the usual case, because most clients (Claude Code included) only load a new server
after a restart, so the server gets used in a later conversation than the one that found it.
Do **not** use it to find things you can already do with installed tools, or for general web
search. Do not call it more than three times for one need; rephrase once, then ask the user.
## Step 1: name the capability, not the request
Turn the need into a short capability phrase, 2–6 words: `read Linear issues`,
`PDF text extraction`, `send Slack messages`. Never pass the user's full prompt, file paths,
repository names, customer names or secrets. The phrase is recorded as demand data.
## Step 2: look it up
Call `find_capability` with:
```json
{
"task": "read Linear issues",
"runtime": "claude-code",
"max_results": 3,
"session": "k7Qx2mPz9Lw4",
"domain": "software-engineering"
}
```
`runtime` is one of `claude-code`, `codex`, `claude-desktop`, `cursor`, `vscode`, `generic`;
it selects the install snippet format. `session` is a random id you make up once per
conversation (8–64 of `A-Z a-z 0-9 _ -`, never derived from the user, the project, the
machine or the time) and pass on every lookupd call in it, `report_outcome` included;
`domain` is what the conversation is about, from the list in Step 4. Together they let
lookupd learn which searches lead to which outcomes. The response carries `lookup_id`, the
id of this search: keep it, and quote it in the outcome report for whichever result you end
up using. Each result has `name`, `description`, `transports`, `install[]` (paste-ready),
`status`, and `lookupd.verification` (`unverified` means nobody has proven they own the
listing).
If the lookupd tools are missing but you have a shell, use the CLI instead:
`npx -y lookupd@latest find "<phrase>" --runtime <runtime> --json` (same results). If you
have neither, tell the user how to add the tools and stop:
`/plugin marketplace add lookupd/lookupd` then `/plugin install lookupd@lookupd`, or
`claude mcp add lookupd -- npx -y lookupd@latest mcp`.
If `results` is empty, say so plainly. Try one rephrase (broader noun, no brand name). If
`note` says no searchable terms, use plain words. Never invent an entry.
## Step 3: propose, and wait for approval
Show at most three candidates. For each: `name`, one line of description, the transports,
the label **Provisional · unverified** (or the verification shown), and the exact install
snippet. Recommend one and say why in one sentence (remote beats local install when both
exist; prefer entries whose `status` is `active`).
Then ask the user to approve the install. **Never run an install command, edit MCP config,
or add a server without explicit approval.** Installing runs third-party code.
After approval, run or apply the snippet, confirm the server's tools appear, and continue
the original task. If the server fails to connect, try the next candidate once, then report.
If the client needs a restart before the new tools appear (Claude Code does), say so; the
outcome is reported in the conversation that uses the server (Step 4).
## Step 4: report the outcome
After a task that used an MCP server finishes (success or not), call `report_outcome` once
per server, once per conversation. This covers servers found in this conversation and servers
that were already installed.
`entry` is the registry name. For a server found in this conversation it is the result's
`name`. For an installed server, look up its launch details in the MCP config (in Claude Code:
`claude mcp get <server>`, or `.mcp.json`), then call `get_entry` with `url` (remote servers)
or `package` (the npm/PyPI package after `npx -y` or `uvx`). Use the returned `entry.name`; if
`get_entry` finds nothing, skip the report. Never guess a name.
```json
{
"entry": "io.github.acme/linear-mcp",
"success": true,
"latency_bucket": "lt5s",
"runtime": "claude-code",
"capability_tag": "issue-tracking",
"session": "k7Qx2mPz9Lw4",
"step": 1,
"tool": "list_issues",
"domain": "software-engineering",
"found_via": "lookupd",
"lookup_id": "7b1f0c2e-5d4a-4b8e-9f1a-2c3d4e5f6a7b"
}
```
| field | values |
| ---------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| `entry` | the registry name (above) |
| `success` | required: `true` if the server did what was needed, `false` if not |
| `latency_bucket` | `lt1s`, `lt5s`, `lt30s`, `gte30s` (time for the server's calls) |
| `runtime` | your runtime, as above |
| `capability_tag` | one tag from lookupd's taxonomy, e.g. `issue-tracking`, `pdf-extract`, `database-sql`, `web-search`, `email`, `other` |
| `error_class` | on failure: `auth`, `timeout`, `schema`, `tool_error`, `other` |
| `session` | a random id for this conversation, 8–64 of `A-Z a-z 0-9 _ -` (see below) |
| `step` | 1, 2, 3…: the order in which this server was first used in the conversation |
| `tool` | the name of the tool you called most on the server, e.g. `list_issues` |
| `domain` | what the conversation is about: `software-engineering`, `devops-cloud`, `data-analytics`, `finance`, `legal-compliance`, `sales-crm`, `marketing`, `customer-support`, `hr-recruiting`, `operations`, `research`, `education`, `healthcare`, `ecommerce`, `design-media`, `personal`, `other` |
| `found_via` | `lookupd` if this conversation found the server through lookupd, else `installed` |
| `lookup_id` | the `lookup_id` from the `find_capability` response that found this server (omit for an installed server) |
`session` groups the lookups and reports of one conversation so lookupd can learn which
searches lead to which outcomes and which servers get used together. Before the first lookupd
call in a conversation, make up one random id; never derive it from the user, the project, the
machine or the time. Reuse it on every lookupd call and report in the same conversation, and
nowhere else.
Only these fields exist. Never include task text, prompts, paths or output. If
`report_outcome` is not in your tool list, the user turned telemetry off (`lookupd telemetry off`
or `DO_NOT_TRACK`): skip the report, and never change that setting yourself.
## Rules
- A `429` means wait the given seconds; lookups are rate limited per IP.
- `verified_only: true` keeps only verified publishers (claimed listings); few exist yet, so
leave it off unless the user asks for verified servers only.
- lookupd mirrors the Official MCP Registry (CC0) and is not affiliated with it. Credit it
if the user asks where the data comes from.
Wersje
W bazie nie ma innych skilli o nazwie find-capability.
Migawka bazy z 13.11.2026 · wersja paczki e51423bb7314 (skrót SHA-256 całej paczki)
Pobrania w ostatnich 30 dniach: 0 · ranking
Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.
Podobne skille
- 100find-agent-tools
Wyszukuje w katalogu serwer MCP lub skill agenta do zadania, od najwyżej ocenianych, z gotową konfiguracją dla klienta. - 100mcp-builder
Buduje, rejestruje i wykorzystuje własne serwery MCP w środowisku, gdy trzeba dodać nowe narzędzie, możliwość lub integrację jako serwer MCP. - 81mcp-setup
Instaluje, konfiguruje i diagnozuje serwer MCP dla wybranego narzędzia, gdy jego narzędzia nie działają lub brakuje. - 95add-tool
Tworzy szkielet nowej definicji narzędzia MCP dla serwera, gdy trzeba dodać nową funkcję. - 100capability
Wyszukuje, porównuje i weryfikuje możliwości agentów, takie jak skille, CLI, serwery MCP i biblioteki, także pod kątem bezpieczeństwa przed instalacją. - 91add-tool
Tworzy szkielet nowej definicji narzędzia MCP, gdy użytkownik chce dodać kolejną możliwość do serwera.