mrkeyoor.com_
Sat 12 Sept 06:30 UTC
LLM Toolsevaluationupdated 12 Sept 2026

knowledge-work-plugins review

Knowledge Work Plugins is Anthropic's collection of 11 role-based plugin packs for Claude Cowork and Claude Code. Each pack combines instructions, slash commands, sub-agents, and MCP connectors so Claude can follow a repeatable workflow for jobs such as sales, finance, support, data analysis, or bio-research.

Verdict

Our run of the nested Allotrope skill installed 35 packages in 18 seconds and built in 4 seconds, but the repository offered no test target. Use this collection as editable Claude scaffolding when your team can inspect prompts, authorize each MCP connection, and review the output. Do not mistake 11 polished role labels for finished company process: the README itself calls them generic starting points, and current connector reports show real setup gaps.

We ran it

Lab card: what happened when we ran knowledge-work-pluginsScreenshot of knowledge-work-plugins (github.com/anthropics/knowledge-work-plugins)
Install✓ · 18s35 packages · 37 MB
Build✓ · 4s
Testsn/ano test script
Known vulns0(pip-audit)
Repo1137 files~9,921 lines of source · 6.9 MB · 6 CI workflows

Answers from our run

Does knowledge-work-plugins build from source?

Dependencies installed in 18 seconds (35 packages), and the build succeeded in 4 seconds. We cloned commit a6d8653 into a clean Debian container with 3 CPUs and no project-specific setup.

Does knowledge-work-plugins have tests you can run?

Not through a standard command: the project exposes no test script or target that our harness could run.

Does knowledge-work-plugins have known vulnerabilities in its dependencies?

pip-audit found none in the dependency tree at the time of our run.

Who should not use knowledge-work-plugins?

Users outside Claude Cowork or Claude Code: the repository documents those two hosts rather than a portable runtime for any assistant.

What are the alternatives to knowledge-work-plugins?

Anthropic Skills, Agents, MCP Servers. Our run of the nested Allotrope skill installed 35 packages in 18 seconds and built in 4 seconds, but the repository offered no test target.

Setup4/5Two CLI commands; useful connectors still need accounts and review
Docs4/5Clear structure and customization advice, with connector gaps
Community4/523,981 stars and a same-day push; 68 open issues
Maturity3/5No release feed or test target, and blank connectors remain

Who it’s for

Claude Cowork or Claude Code teams that want 11 editable starting points for common knowledge-work roles.
Administrators prepared to review skills, connect approved MCP services, and replace generic language with company rules.
Claude Code users who prefer installing a marketplace plugin over copying individual prompt files.
Teams comfortable treating generated legal, finance, data, or research work as material that still needs qualified review.

Who it’s NOT for

Users outside Claude Cowork or Claude Code: the repository documents those two hosts rather than a portable runtime for any assistant.
Teams expecting every advertised connector to work after install: current sales/.mcp.json leaves Gmail and Google Calendar URLs blank, matching open issue 933.
Organizations that cannot grant an assistant access to Slack, CRMs, calendars, document stores, or data warehouses required by their chosen pack.
Buyers who need versioned release artifacts: GitHub returned no latest release for this repository on September 12, 2026.
Anyone taking the README's no-code description literally: our measured bio-research skill installed 35 Python packages and contains executable conversion scripts.

Setup reality

Our sandbox installed the nested bio-research/skills/instrument-data-to-allotrope project in 18 seconds, adding 35 packages and 37 MB. Its build succeeded in 4 seconds. No test script or target existed, so tests were skipped; pip-audit found 0 known vulnerabilities.

The main collection installs through Claude Cowork or two Claude Code commands. Useful workflows need the relevant external accounts and MCP authorization, then edits to connector definitions, terminology, and process instructions. Some roles touch sensitive systems such as email, CRM records, and data warehouses.

The measured checkout had 1,137 files, about 9,921 source lines, and used 6.9 MB before installation. It had 6 CI workflow files, no Dockerfile, and no tests directory. The lab had no secrets, so it did not exercise Claude activation, connectors, or the quality of generated work.

Eleven role packs turn Claude instructions into installable units

The root README lists 11 plugins for productivity, sales, customer support, product management, marketing, legal, finance, data, enterprise search, bio-research, and plugin management. A pack can contain a manifest, an MCP configuration, slash commands, reusable skills, and sub-agents. This is more useful than a folder of loose prompts because Claude Cowork and Claude Code know when to activate skills and how to expose named commands such as call preparation or query writing.

The repository's best idea is packaging a whole job rather than one clever instruction. The sales pack can pair research and outreach workflows with CRM connectors. The finance pack covers journal entries, reconciliations, statements, close work, and audit support. Bio-research reaches literature and laboratory systems. Those descriptions are broad, but the README is candid about their status: all 11 are generic starting points that become useful after a company adds its terminology, tools, and process rules.

Two Claude Code commands install a pack, then configuration begins

The documented Claude Code path uses 2 commands: one adds anthropics/knowledge-work-plugins as a marketplace, and the other installs a named plugin. Cowork users can install from Claude's plugin directory. Once present, relevant skills activate automatically and slash commands become available. That is a short installation path for the file-based portion of the product, and the Apache-2.0 repository license leaves room for internal edits and redistribution under its terms.

Installation is only the first boundary. MCP connectors point Claude at services such as Slack, Notion, HubSpot, Jira, BigQuery, and Microsoft 365. Each chosen service brings its own account, authorization, and data-access decision. The README tells teams to edit .mcp.json, add company context, and adjust workflows. A legal or finance pack needs more than valid JSON: someone qualified must decide which sources Claude may read, which actions it may take, and who reviews the result.

What happened when we ran it

Our sandbox targeted bio-research/skills/instrument-data-to-allotrope at commit a6d8653 because that directory contains an installable Python project. Installation succeeded in 18 seconds, adding 35 packages and taking 37 MB on disk. The build completed in 4 seconds. Pip-audit reported 0 known vulnerabilities in the installed set. The environment was an unprivileged Debian container with 3 CPUs, 8 GB of RAM, Python 3.12, and no secrets.

There was no test script or target, so the test step was skipped. The 6.9 MB checkout held 1,137 files and about 9,921 lines of source. Our scan found 6 CI workflow files, no Dockerfile, and no tests directory. These results verify that one nested skill's dependencies installed and its build completed. They do not verify all 11 role packs, Claude's skill activation, any remote MCP connection, or the correctness of work produced through those instructions.

One Python skill makes the no-code description too broad

The README says every component is Markdown and JSON with no code or build steps. Our 4-second build found a meaningful exception inside the bio-research plugin. Its instrument-data-to-allotrope skill includes Python scripts for converting laboratory files to Allotrope Simple Model JSON, flattening output to CSV, exporting parser code, and checking the result. Its requirements pin allotropy, pandas, openpyxl, and pdfplumber, which account for the 35 packages installed in our run.

That skill also labels itself an example and tells adopters to customize schemas, ontology mappings, and proprietary instrument formats. Its validator uses a December 2024 Allotrope specification and can treat unknown values as warnings unless strict mode is selected. Those details are sensible, yet they raise the review burden. A scientist should confirm instrument support and field mappings, while a data engineer should inspect generated parser code before it enters a production pipeline. The missing test target leaves those checks to the adopter.

Open issue 933 matches blank URLs in the current sales pack

Issue 933 reports 24 blank Gmail and Google Calendar connector entries across 12 plugin manifests at commit be9c0da. The current sales/.mcp.json still gives both services empty URLs. According to the report, Claude Code registers those entries and then shows connection failures with no endpoint to reach. The issue describes cache editing or disabling the whole plugin as the available workarounds. That is specific friction in a collection whose value depends heavily on connected services.

A separate open report, issue 1031, says Gmail, Google Calendar, and Slack are redeclared per plugin, which can force repeated authorization when several packs use the same account. We did not run authenticated connectors in our no-secret sandbox, so these are issue findings rather than lab reproductions. Administrators should install one pack first, inspect claude mcp list, and settle connector ownership before handing several role plugins to a department.

A September 12 push shows activity despite no release feed

GitHub recorded the last push on September 12, 2026, along with 23,981 stars and 77 combined issues and pull requests. GitHub issue search returned 68 actual open issues. The latest-release endpoint returned 404, so there is no current GitHub release tag to cite. That absence complicates version pinning, but it does not signal abandonment when code moved on the same day and users were filing concrete reports during the preceding week.

Anthropic Skills is the closer choice when you want individual skills without an 11-role package. The MCP Servers repository suits teams that mainly need reference tool connections and will author their own operating instructions. Wshobson's Agents is aimed more at coding work and supports several agent hosts. Knowledge Work Plugins is strongest inside a Claude-only organization that has owners for permissions, prompt review, and output review. The 18-second nested install is easy; making each role reflect the company is the actual project.

Alternatives

ProjectWhat it isPick it when
Anthropic Skills gh↗Anthropic's public collection of individual Agent Skills and examples.pick this instead when you want smaller task instructions without adopting an entire role pack and its connector set.
Agents gh↗A plugin marketplace for several coding-agent hosts, including Claude Code and Codex.pick this instead when software engineering agents and support for several hosts matter more than office-role workflows.
MCP Servers gh↗The Model Context Protocol project's collection of reference servers.pick this instead when direct tool connections matter and you want to write the role instructions yourself.

What people are saying

  1. [velocity-scout] anthropics/knowledge-work-plugins

Sources

  1. Knowledge Work Plugins repository and README
  2. Instrument Data to Allotrope skill
  3. Sales plugin MCP configuration
  4. Blank Gmail and Google Calendar connector issue 933
  5. Duplicate connector authorization issue 1031

More llm tools reviews

llamafile · OpenContext · hyperresearch · prompt-optimizer · agent-router · llm-for-zotero · the whole board →