pwc-cli

Install:

pwc_repo=https://raw.githubusercontent.com/huggingface/pwc-cli/main
curl -LsSf "$pwc_repo/standalone_cli/install.py" | python3

The pwc CLI is anonymous and read-only. It queries the public Papers With Code catalog and requires no token. Run pwc --help or a nested --help command when the live parser and this skill disagree; the parser is authoritative.

Install or refresh this Skill from the locally installed CLI with pwc skills add --force. Use --global for ~/.agents/skills, --claude to also link it for Claude Code, or --dest PATH for another harness.

Use compact output for reading and discovery. Add --json when programmatic filtering, joining, or schema-dependent processing is useful. Use only flags documented for the exact subcommand; do not infer that sibling commands share flags. For pwc paper list, pwc task list, and pwc method list, use --page-size, never --limit.

PAPER accepts a modern or legacy arXiv ID, a numeric external-paper ID, or an exact paper title. Papers with Code supports external papers which are not on arXiv, hence those papers will have a numeric canonical ID. Quote titles containing spaces. Title matching is case-insensitive but exact; ambiguous titles fail with their matching IDs rather than silently selecting a paper.

Commands

• pwc search QUERY — Search papers by title, topic, author, or arXiv ID. This is powered by hybrid/keyword/semantic search using pgvector. Defaults to hybrid. [--limit 1-100 --page 1-100 --mode hybrid|keyword|semantic --start-date YYYY-MM-DD --end-date YYYY-MM-DD --json]
• pwc paper info PAPER — Show metadata including tagged organizations, the abstract, predecessors, and successors. --include-resources adds Markdown sections for GitHub repositories, project pages, and Hugging Face artifacts, with official links marked explicitly. --include-evals fetches every evaluation for the resolved paper and adds an Evaluations Markdown table (or structured evaluations in JSON output). [--include-resources --include-evals --json]
• pwc paper read PAPER — Print stored paper Markdown. The resolved paper must have a modern arXiv record. [--json]
• pwc paper list — List and filter the paper catalog in a paginated manner. [--page 1-100 --page-size 1-100 --search TEXT --start-date YYYY-MM-DD --end-date YYYY-MM-DD --task NAME --method NAME --conference NAME --framework NAME --organization NAME --all-versions --order-by trending|date_published|citation_count --order-dir asc|desc --include-resources --json]

Date ranges are inclusive. When both bounds are supplied, the start date must not be after the end date.

• pwc paper recent — List recently published papers. [--limit 1-100 --json]
• pwc paper trending — List papers with recent repository activity. [--limit 1-100 --max-age-days 1-365 --min-velocity FLOAT --json]
• pwc paper related PAPER — Find embedding- and taxonomy-ranked related work. [--limit 1-20 --json]
• pwc paper lineage list PAPER — Render linked Markdown sections for a paper, its predecessors, and its successors. [--json]
• pwc task --name NAME — Show a task page in CLI form, including its area, description, hierarchy, research trends, recommended frameworks, sister tasks, subtasks, common methods, benchmarks, and trending papers. NAME must exactly match a task name, slug, or ID. [--json]
• pwc task list — List research tasks, optionally filtering by a case-insensitive exact area name such as Vision, Audio, or General. Rows include direct paper, benchmark, and evaluation counts for each task. Interactive output groups visible top-level tasks into Markdown area sections; --group-by-area forces that view when output is captured, while --flat forces the paginated table. [--page 1-100 --page-size 1-100 --area NAME_OR_ID --level INTEGER --visible-only --group-by-area|--flat --order-by name|created_at|level|paper_count --order-dir asc|desc --json]
• pwc method list — List research methods, optionally filtering by area or introduction year. This subcommand has no --search or --limit flag; do not infer flags from sibling list commands. To find methods by name or topic, filter the bounded list with pwc method list --page-size 500 | rg -i -- 'QUERY'. [--page 1-100 --page-size 1-500 --area NAME_OR_ID --introduced-year YEAR --order-by name|full_name|introduced_year|created_at|paper_count --order-dir asc|desc --json]
• pwc method --name NAME — Show method metadata, its research area, description, introduction year, source paper, and paper count. NAME must exactly match a method name, full name, slug, or ID. [--json]
• pwc conference list — List conferences with imported papers. [--year YEAR --json]
• pwc conference --name NAME — Show conference dates, venue, description, links, tier, and paper count. NAME must exactly match a conference name, slug, or ID. [--json]
• pwc organization list — List research organizations and their paper and trending metadata. [--featured-only --json]
• pwc organization --name NAME — Show organization metadata and public links. NAME must exactly match an organization name, slug, or ID. [--json]
• pwc framework list — Flatten and list the framework catalog. [--domain NAME_OR_SLUG --category NAME_OR_SLUG --platform NAME --json]
• pwc framework --name NAME — Show framework guidance, platforms, links, and introducing paper. NAME must exactly match a framework name, slug, or ID. [--json]
• pwc benchmark list — Find and rank benchmarks. When --task is supplied, the default order is task-specific trending activity. Without filters, interactive output groups top benchmarks under visible tasks by area; --group-by-area forces Markdown and --flat forces the dataset table. Task, search, pagination, ordering, and open-model filters take precedence over --group-by-area and select task-scoped flat output. Rows place the benchmark description and optional Hugging Face dataset URL immediately after the benchmark name, followed by the slug and ID (plus the full name when distinct in grouped output); pass the slug or ID to pwc benchmark --name IDENTIFIER for leaderboard details. [--page 1-100 --page-size 1-100 --search TEXT --task TASK --area NAME_OR_ID --benchmarks-per-task 1-10 --group-by-area|--flat --include-descendants --min-eval-count INTEGER --is-open true|false --order-by trending|name|full_name|created_at|paper_count --order-dir asc|desc --json]
• pwc benchmark --name NAME — Show a benchmark's top models, scores, source papers, publication dates, parameter counts, and open/closed status. NAME must exactly match a benchmark name, full name, slug, or ID. For multi-metric tradeoffs, require numeric metrics, apply repeatable thresholds, sort by one metric, or select a Pareto frontier. Metric names are case-insensitive. --max-parameters SIZE keeps models at or below an inclusive parameter limit; SIZE accepts values such as 500M, 1.5B, 3B, and raw integers. Models without one consistent parameter count are excluded. [--limit 1-100 --is-open true|false --max-parameters SIZE --require-metrics METRIC[,METRIC] --min METRIC=VALUE --max METRIC=VALUE --sort METRIC[:asc|desc] --pareto METRIC:higher,METRIC:lower --json]
• pwc version — Show the CLI and API contract versions.
• pwc skills add — Install the version-matched CLI Skill. [--global --claude --dest SKILLS_DIRECTORY --force]

Research workflow

1. Use pwc benchmark list --task TASK to discover active benchmarks for a given task, then pwc benchmark --name NAME to inspect a specific leaderboard. Add --max-parameters SIZE for on-device, edge-inference, or other resource-constrained device requests. For accuracy/latency or other multi-metric questions, use --require-metrics, thresholds, or --pareto; do not infer a tradeoff from a leaderboard sorted by one metric.
2. Treat the row order from pwc benchmark list --task TASK as the primary benchmark priority for an unqualified "best models for TASK" or SOTA request. The order reflects support-weighted recent reporting activity. Inspect and lead with the highest-ranked benchmark that has usable leaderboard results, and preserve benchmark order when comparing several. Do not promote a familiar lower-ranked benchmark merely because its name resembles the task. Deviate only when the user's requested use case clearly favors a specialized benchmark, and explain that reason explicitly.
3. Use pwc paper info to inspect promising results. Add --include-resources when linked GitHub repositories, project pages, or Hugging Face artifacts matter. When the paper includes successors, always consider them more state-of-the-art.
4. Use pwc search to search more broadly for a topic or a known paper. Use the default hybrid mode first; use keyword mode for exact terminology and semantic mode for conceptual matches.
5. Go more in-depth with pwc paper read. Do not treat search snippets or titles as sufficient support for detailed claims.
6. Expand the literature with pwc paper related and use pwc paper lineage list when model or method ancestry matters.
7. Explore the catalog taxonomy with pwc task list --group-by-area, inspect a specific task with pwc task --name NAME, then use a method, conference, organization, or framework with its --name command; use each entity's list filters to narrow broad catalogs.
8. Synthesize only after gathering enough primary evidence. Preserve paper titles, identifiers, and URLs in the answer so claims remain traceable.

Command selection

• Use pwc benchmark list --task TASK for finding state-of-the-art (SOTA) for a given task.
• If the task is not supported by PwC, use pwc search for relevance-ranked discovery.
• Use pwc paper info to gather more info about a specific paper.
• Use pwc paper list for structured filters, date windows, conferences, and deterministic sorting. Use exact --task, --method, --conference, --framework, or --organization filters for tagged or catalog-associated papers; combine filters to require every association. Do not substitute a keyword search for an organization or taxonomy filter.
• Use pwc paper recent for recency and pwc paper trending for current repository activity; these are different signals.
• Use area names directly with task and method lists, for example --area Vision, --area Audio, or --area General. Area matching is case-insensitive; numeric area IDs are also accepted.
• Use pwc conference list --year YEAR for a specific conference edition year.
• Use --include-resources for linked repositories, project pages, and Hugging Face artifacts. Add --json when their structured metadata also matters.
• Use --include-evals with pwc paper info when benchmark scores reported by the paper matter.
• Use --all-versions only when individual arXiv versions are relevant.
• Use --is-open true to restrict benchmark results to open models.
• Paginate when stderr reports more results. Do not silently treat the first page as the complete catalog.

Output and limits

• Interactive list and search output uses aligned text columns. Benchmark lists remain aligned when piped or captured; other captured lists use lossless TSV for agent and script processing. Paper info uses labeled metadata and Markdown sections. Benchmark details use an aligned table interactively and Markdown when piped or captured.
• --json wraps responses as {"schema_version":"v1","data":...} for stable agent consumption.
• Stable exit codes are 0 success, 2 invalid usage, 3 network/server failure, and 4 invalid API response.
• PWC_API_URL may select another compatible v1 endpoint. The default is https://paperswithcode.co/api/v1.
• Catalog-filtered paper lists fail closed unless the server confirms every requested filter; never treat results from an older server as filtered.

The standalone parser contains no authentication, mutation, ingestion, publication, image, CRON, embedding, or infrastructure commands. Do not substitute repository-maintenance commands into a public or sandbox workflow.