Tool families

Liminal ships 500+ tools in the family catalog (additional tools when Google/Microsoft MCP sidecars attach). With AGENT_TOOL_LAZY=1 (default) only a baseline subset is registered at startup; the model activates more families on demand. This keeps the tool list the model reasons over small and relevant.

• Discover: list_tool_families (pass a task_hint to get suggestions).
• Activate: activate_tool_family("<family>") loads a whole family into the live registry.
• Pre-declare: breakdown can declare the families a plan will need so they pre-activate.

Baseline set is chosen by AGENT_ALWAYS_TOOLS_PROFILE (balanced default · knowledge_first · max_autonomy). Authoritative catalog: packages/tools/src/tool_catalog.ts; registration: packages/tools/src/index.ts. Every registered tool belongs to a family or the always-set (enforced by tool_catalog.test.ts).

See Lazy tool loading below if a tool you expect isn't available, and Configuration for the gating env keys.

Always loaded (baseline)

These are registered up front in every session — code is the agent's substrate, so reading, editing, navigating, and reasoning never wait on activation.

Group	Tools
Reasoning	think, reason, plan, hypothesize, breakdown
Read & navigate	read_file, read_file_chunked, read_file_with_imports, file_metadata, list_dir, grep_file, find_files, repo_map, workspace_snapshot
Write	write_file (create/overwrite/append), edit_file (replacements or fuzzy diff)
Meta (always)	list_tool_families, activate_tool_family, api_connect/api_list, mcp_attach/mcp_detach

Family catalog

Family	What it adds	Notable gate
files_edit	move_file, copy_file, copy_tree, mkdir_p, delete_file (approval-gated), multi_file_apply (atomic, rollback), path_guard	—
shell	run_shell (destructive, approval-gated), run_background, kill_process, list_processes, read_process_output, run_command_with_pty, execute_code	—
code_intel	ast_grep, symbol_index, find_references, rename_symbol (project-wide TS/JS rename), run_tests, run_lint	—
git	git_status, git_diff, git_log, git_branch, git_commit, git_checkpoint, git_rollback, git_worktree	—
web	web_search, web_fetch, http_request, weather_lookup	—
markets	markets_quote	AGENT_MARKETS_ENABLE
browser	browser_open/navigate/snapshot/act/extract/wait_for/cookies/close, browser_serve_file	AGENT_BROWSER
captcha	captcha_solve	AGENT_CAPTCHA_KEY
vision	vision_analyze, upload_image	AGENT_VISION_*
audio	speak (TTS), transcribe_audio	AGENT_TTS_ENABLED / AGENT_TRANSCRIBE_ENABLED
memory_advanced	recall_relevant, memory_query, memory_graph, memory_consolidate, curate_memory, restore_memory, memory_promote, memory_neighbors, consolidate_chat, read_artifact, failure_review	—
vault	vault_write/read/search/list/links/graph/delete	—
document	doc_plan → doc_* → doc_render_pptx/docx/pdf → doc_export	AGENT_DOC_ENGINE
tasks	task_checkpoint, resume_task, feature_checklist, pattern_record	—
agenda_scheduler	agenda_set/get/clear, schedule_create/list/delete/run	—
meta	suggest_improvement, view_insights, self_telemetry, dynamic-tool + connector tools	—
dynamic_tools	create_tool, edit_tool, remove_tool, list_dynamic_tools	—
external_api	api_<conn>_<op> (from api_connect), mcp_<conn>_<tool> (from mcp_attach)	—
connectors	connect_provider, disconnect_provider, list_connectors, integrations_snapshot, revoke_integration_account, mail_search_inboxes	OAuth in Integrations hub
google_workspace	Google MCP + Calendar/Gmail REST + email_style_infer	Google OAuth
microsoft_365	Microsoft MCP + Outlook REST + email_style_infer	Microsoft OAuth
github	GitHub MCP (dynamic)	GitHub OAuth / GITHUB_TOKEN
slack	Slack REST (channels, search, threads, reactions, files)	Slack OAuth
linear	Linear REST issues	Linear OAuth
notion	Notion REST workspace	Notion OAuth
xero	Xero accounting REST	Xero OAuth
youtube	YouTube Data + Analytics REST	YouTube OAuth
reasoning_advanced	decompose_goal, verify_contract, dispatch_graph, query_tool_outputs, research_state	harness-scoped
synthesis	synthesis_run (cross-domain synthesis sub-agent)	harness-scoped
orchestration	spawn_agent, wait_for_agents, cancel_agent, list_agents, verify_result, evidence_critic, path_critic, policy_critic, reflect_debate, branch_explore, branch_evaluate	harness-scoped
workflow	plan_workflow, run_workflow, workflow_status, query_workflow	root-only, AGENT_WORKFLOWS
navigation	extended repo/file navigation helpers	—
independence	breakout_start, independence_status	—
harness_ui	check_context, compress_context, get/set_runtime_settings, set_persona, refresh_world_context	harness-scoped

Harness-scoped families are recreated per child agent via onChildCreated (never copied parent → child). Root-only families (workflow) are excluded from children so workflows can't nest. See Sub-agents & orchestration and Dynamic workflows.

Feature deep-dives

Area	Guide
Web research (web_search + parallel web_fetch)	Web research
Browser automation (Playwright)	Browser automation
Memory, recall, vault	Memory & recall
Documents (PPTX/DOCX/PDF)	Document engine
Voice (TTS + dictation)	Voice
Parallel sub-agents	Sub-agents & orchestration
Multi-phase fan-out	Dynamic workflows
External APIs / MCP	api_connect (OpenAPI 3.x) · mcp_attach (Streamable-HTTP MCP)
Integrations (OAuth)	Connectors · per-provider guides
Multi-account mail	Multi-account mail
Compose dock & attachments	Compose dock & attachments
Inbox automation	Inbox watcher

Troubleshooting inactive tools

If the model reports a tool isn't available under lazy loading:

1. list_tool_families (pass a task_hint) to see families and what they hold.
2. activate_tool_family("<best-fit family>") — one family, not many.
3. Retry the exact tool with corrected args.

Never activate the same family twice on one registry — tool_catalog.ts tracks activated families. To keep a family always-on (e.g. browser), set its always-active env flag or use the max_autonomy profile.

Contributors

Full per-file map: repository root CLAUDE.md (packages/tools section). Eval scenarios: npm run eval -w packages/eval.