{"generated_at":"2026-10-14T20:07:22+00:00","item":{"added_at":"2026-09-09","analysis_method":"independent_manual_review_2026-10-10","assessment_status":"provisional","auth_posture":"auth_required","category":"Web Connector","confidence":"high","description":"Ask Agentwork questions across your connected company knowledge. Receive source-linked answers or run tasks using the organization's tools and context. Existing access rules and configured human-approval requirements remain in effect.","did":"c82a9d29-7f3a-49ec-aec8-bac38c829ee1","documentation_url":"https://agentwork.com/docs/agent-guide","editorial_description":{"extended_description":"Connected sources can include Slack, Google Workspace, Notion, Linear, GitHub, email and CRM systems. Tasks appear in the team's Agentwork inbox, and unresolved knowledge gaps can be routed to a person for clarification. You need an Agentwork account with at least one source connected; approval requirements may differ where your organization has explicitly opted tools out.","schema_version":1,"summary":"Ask Agentwork questions across your connected company knowledge.\nReceive source-linked answers or run tasks using the organization's tools and context.\nExisting access rules and configured human-approval requirements remain in effect."},"homepage_url":"","icon":"","id":"c82a9d29-7f3a-49ec-aec8-bac38c829ee1","installs":0,"last_scanned":"","license":"","long_description":"Connected sources can include Slack, Google Workspace, Notion, Linear, GitHub, email and CRM systems. Tasks appear in the team's Agentwork inbox, and unresolved knowledge gaps can be routed to a person for clarification. You need an Agentwork account with at least one source connected; approval requirements may differ where your organization has explicitly opted tools out.","manual_assessment":"high","manual_review":{"all_findings":[{"evidence":"Official docs say the tool runs a full non-interactive agent session, including software-engineering tasks, and returns finished text and files.","reasoning":"Delegating arbitrary software-engineering work entails remote code/tool execution and file generation, meeting the High code-execution threshold.","risk_type":"code_execution","severity":"high","title":"A remote software-engineering agent can execute a full task","triggering_tools_or_skills":["run_agentwork_task"]},{"evidence":"The listing names Slack, Google Workspace, Notion, Linear, GitHub, email and CRM sources; official docs say tasks can check a connected system while skipping actions requiring human approval.","reasoning":"The exact reachable action set depends on hidden Agentwork connectors, so a generic agent task can have broader side effects than the four MCP names reveal.","risk_type":"generic_gateway","severity":"high","title":"Task execution can reach organization-connected systems","triggering_tools_or_skills":["run_agentwork_task","ask_agentwork"]},{"evidence":"The connector runs at https://mcp.agentwork.com/ and no pinned server implementation was available in the review packet.","reasoning":"The hosted implementation and action semantics can change independently of this catalog snapshot, so runtime verification and change monitoring remain necessary.","risk_type":"provenance_uncertainty","severity":"low","title":"Agentwork is a mutable remote service","triggering_tools_or_skills":[]}],"assessment_status":"provisional","auth_and_scopes":"The MCP server supports OAuth or a persistent API key and operates in a selected Agentwork organization; connected-source permissions are enforced by Agentwork but exact downstream scopes vary.","blocked_missing_evidence":null,"confidence":"high","displayed_risk":"high","evidence_completeness":"partial","manual_overall_risk":"high","official_sources":[{"accessed_at":"2026-10-10T20:24:37Z","url_or_path":"plutosecurity/plutonium@b749c8dc43536c794409be9805edd48063f045e2:app/data/conn/claudesec/c82a9d29-7f3a-49ec-aec8-bac38c829ee1.json","what_it_proves":"Candidate PR #125 record identity (name \"Agentwork\", publisher, type web_connector) and 4 stored tool names; no tool descriptions stored"},{"accessed_at":"2026-09-22T10:46:07Z","url_or_path":"https://api.anthropic.com/api/directory/servers?verified_tier=anthropic,partner,community (snapshot claude-directory-all-tiers-2026-09-22.json)","what_it_proves":"Official Claude Directory listing https://claude.ai/directory/c82a9d29-7f3a-49ec-aec8-bac38c829ee1: complete tool_names inventory (4 tools), remote streamable-http endpoint https://mcp.agentwork.com/, auth_posture auth_required, directory tier community"},{"accessed_at":"2026-10-10T21:04:12Z","url_or_path":"https://agentwork.com/docs/agent-guide","what_it_proves":"Agentwork documents run_agentwork_task as a non-interactive full agent session that performs software-engineering work and returns generated files; actions needing human approval are skipped."}],"primary_reason":"Agentwork runs a hidden, non-interactive software agent that can produce files and use organization-connected context.","provisional_risk":null,"publication_recommendation":"accept","requires_runtime_verification":true,"review_key":"claude:c82a9d29-7f3a-49ec-aec8-bac38c829ee1","reviewed_at":"2026-10-10","reviewer_notes":"Reviewed all 4 declared tools and the complete official guide. Human-approval-required connector actions are documented as skipped, but filesystem/repository and deployment reach are not fully enumerated and remain unknown.","risk_domains":{"arbitrary_network_access":"false","authz_ambiguity":"false","browser_control":"false","credential_or_admin_actions":"false","deletes_data":"false","deployment_or_infrastructure":"unknown","dynamic_hidden_actions":"true","filesystem_or_repo_access":"unknown","forwards_data_to_third_party":"true","modifies_data":"false","persistent_credentials":"true","prompt_injection_exposure":"true","provenance_uncertainty":"true","reads_private_data":"true","runs_code_or_queries":"true","sends_messages_as_user":"false","spends_or_moves_money":"false"},"tools_or_skills_reviewed":4,"tools_or_skills_total":4},"name":"Agentwork","needs_runtime_verification":true,"plutonium_url":"https://plutonium.pluto.security/detail.html?planet=claudesec&did=c82a9d29-7f3a-49ec-aec8-bac38c829ee1&utm_source=plutonium_analysis_skill&utm_medium=claude_skill&utm_campaign=connector_risk_assessment","privacy_policy_url":"https://agentwork.com/privacy","provisional_risk":null,"publication_recommendation":"accept","publisher":"Agentwork","publisher_url":"https://agentwork.com/","remote_url":"https://mcp.agentwork.com/","repository_url":"","risk":"high","risk_severity":"high","security_risks":[{"description":"This connector can execute code, commands, queries, or dynamically selected actions. Review findings: Delegating arbitrary software-engineering work entails remote code/tool execution and file generation, meeting the High code-execution threshold.; The exact reachable action set depends on hidden Agentwork connectors, so a generic agent task can have broader side effects than the four MCP names reveal.","evidence":"A remote software-engineering agent can execute a full task: Official docs say the tool runs a full non-interactive agent session, including software-engineering tasks, and returns finished text and files. | Task execution can reach organization-connected systems: The listing names Slack, Google Workspace, Notion, Linear, GitHub, email and CRM sources; official docs say tasks can check a connected system while skipping actions requiring human approval.","manual_risk_type":"code_execution,generic_gateway","remediation":{"block_tool_categories":["code_execution"],"steps":["Run only in a sandbox with least-privilege credentials and strict resource limits.","Disable generic execution gateways unless their reachable action set is fully understood.","Require approval before execution against production systems or sensitive data."]},"risk_type":"runs_code","severity":"high","title":"Can execute code, queries, or dynamic actions","triggering_tools":["run_agentwork_task","ask_agentwork"]}],"signature_status":"unknown","source_code_reviewed":false,"support_url":"david@agentwork.com","tags":["network_access","oauth_required","runs_code","web_connector"],"tools":[{"description":"","name":"ask_agentwork","risk":{"category":"code_execution","level":"high","recommendation":"Follow the connector-level remediation and require approval for consequential actions.","why":"This connector can execute code, commands, queries, or dynamically selected actions. Review findings: Delegating arbitrary software-engineering work entails remote code/tool execution and file generation, meeting the High code-execution threshold.; The exact reachable action set depends on hidden Agentwork connectors, so a generic agent task can have broader side effects than the four MCP names reveal."}},{"description":"","name":"get_agentwork_result"},{"description":"","name":"ping"},{"description":"","name":"run_agentwork_task","risk":{"category":"code_execution","level":"high","recommendation":"Follow the connector-level remediation and require approval for consequential actions.","why":"This connector can execute code, commands, queries, or dynamically selected actions. Review findings: Delegating arbitrary software-engineering work entails remote code/tool execution and file generation, meeting the High code-execution threshold.; The exact reachable action set depends on hidden Agentwork connectors, so a generic agent task can have broader side effects than the four MCP names reveal."}}],"tools_count":4,"transport":"streamable-http","type":"web_connector","url":"https://claude.ai/directory/c82a9d29-7f3a-49ec-aec8-bac38c829ee1","uuid":"2997c990-93b6-5bef-a4e1-6bc4c47b856b","verification_tier":"community","version":""},"query_key":"agentwork agentwork","schema_version":1,"status":"match"}