Coding OS
Sign in
Browse docs▾

MCP Tool Inventory

Purpose: Route AI agents to the correct tools for research, validation, and automation tasks. Read when: Starting a task that requires external research, documentation lookup, package checks, or browser verification. Skip when: The task is purely local code manipulation (use built-in tools directly) or the playbook already specifies tool routing. Read next: The domain playbook matching your task type.

Nav: Docs Index | AGENTS

Runtime Snapshot

  • Verified on 2026-07-09
  • Standard runtime must match this doc before it is treated as available truth
  • Agent checks available tools via its session's deferred tool list

Coding-OS Provided

coding-os MCP server

Cognitive OS — memory, graph, board, and cognition tools. SQLite backend at .coding-os/coding-os.db. All tools use cos_* prefix.

  • Health (1): cos_health — DB stats, schema version, FTS5 status
  • Memory (5): cos_search (5-signal ranked), cos_timeline (chronological), cos_details (full record), cos_promote (pattern → rule), cos_observation_record
  • Metrics (3): cos_metric_record, cos_metric_query, cos_metric_trend
  • Learning (4): cos_learn_extract, cos_learn_suggest, cos_learn_validate, cos_learn_narrative
  • Retrieval Quality (4): cos_retrieval_cite, cos_retrieval_learn, cos_retrieval_quality, cos_retrieval_enrichment_check
  • Routing & Roles (6): cos_route_model, cos_route_skill, cos_compose_chain, cos_role_info, cos_situation_detect, cos_classify_prompt
  • Docs RAG (3): cos_doc_search, cos_doc_header, cos_doc_headers_by
  • Graph (22): cos_graph_query, cos_graph_resolve, cos_graph_context, cos_graph_communities, cos_graph_path, cos_graph_impact, cos_graph_references, cos_graph_rename_plan, cos_graph_similar, cos_graph_search, cos_graph_detect_changes, cos_graph_entrypoints, cos_graph_trace, cos_graph_contracts, cos_graph_export, cos_graph_centrality, cos_graph_ranking, cos_graph_dead_code, cos_graph_cycles, cos_graph_test_gap, cos_graph_diff, cos_graph_doctor — cos_graph_search(query) is free-text hybrid (semantic ANN + FTS5 + centrality); cos_graph_similar(uid) is node-to-node.
    • cos_graph_dead_code = in-repo symbols with zero non-test inbound references (dead-code candidates; inverse of centrality)
    • cos_graph_cycles = circular dependencies as strongly-connected components (scope=imports module-level design smell | calls)
    • cos_graph_test_gap = prod function/method/class with zero inbound edge from a test source (untested symbols)
    • cos_graph_diff = graph blast-radius of a git revision range (base..head → changed files → affected symbols + downstream)
  • Board / Tasks (22): cos_task_board, cos_task_show, cos_task_create, cos_task_move, cos_task_ready, cos_task_reclaim, cos_task_reconcile, cos_task_pick, cos_task_claim_next (atomic select+claim — N racing sessions each get a distinct runnable task or claimed:null), cos_task_search, cos_task_by_filter, cos_task_dependencies, cos_task_dependents, cos_task_wip_check, cos_task_daily, cos_task_retro, cos_task_reposition, cos_task_edit (attributed field/body edit), cos_task_history (merged actor-attributed timeline), cos_work_log_append, cos_task_link (set the optional external_ref forge issue/PR link — metadata, not the id), cos_presence_query (active agent sessions from presence files)
  • Cognition (5): cos_supervise, cos_supervise_record_output, cos_dispatch_formula, cos_dispatch_formula_run, cos_dispatch_parallel_run
  • Analysis (5): cos_analyze_task, cos_ambiguity_check, cos_backtrack_log, cos_discovery, cos_failure_pattern_query (aggregated backtrack root-cause anatomy)
  • Logs (1): cos_log_query — durable WARN+ error store query (level floor / scope glob / since / search / session / fingerprint); the agent's "what is broken now"
  • Trajectory (2): cos_trajectory_snapshot, cos_trajectory_read — per-session project phase/focus/decisions persistence feeding SessionStart enrichment
  • Misc (3): cos_traceability, cos_takeover, cos_digest_regenerate

Recommended External MCPs

These are commonly useful but not required. Install via Claude Code MCP settings.

  • context7 — Official framework and library documentation. Use first for framework docs.
  • ref — URL-targeted documentation lookup and direct page reads.
  • playwright — Browser automation and route-state verification.

Built-in Tools

  • File system → Read, Write, Edit, Glob, Grep
  • Command execution → Bash (make targets, git, scripts)
  • Web search → WebSearch (current facts, fallback when MCP unavailable)
  • Web fetch → WebFetch (scrape single URL to markdown)
  • Subagents → Agent tool (read-only research/inventory/verification only; write work runs single-agent — never use isolation: "worktree")

Task-to-Tool Selection Matrix

  • Framework/library docs → context7 first, ref as fallback, WebSearch as last resort
  • Memory search / past patterns → cos_search, cos_timeline, cos_learn_suggest
  • Task board / status → cos_task_board, cos_task_search, cos_task_by_filter
  • Single task lookup (full body) → cos_task_show (in-session; never raw ls/grep/Read on docs/tasks)
  • Task create/move/complete → cos_task_create, cos_task_move, cos_work_log_append
  • File/concept relationships → cos_graph_context, cos_graph_impact, cos_graph_references
  • Rename planning → cos_graph_rename_plan (callers + impact before any rename)
  • Doc search (semantic) → cos_doc_search; single-file frontmatter → cos_doc_header
  • Model/skill routing → cos_route_model, cos_route_skill, cos_compose_chain
  • Role chain for complex task → cos_compose_chain → writes .coding-os/<agent>/.roles
  • Formula dispatch → cos_dispatch_formula, cos_dispatch_formula_run
  • Agent performance metrics → cos_metric_record, cos_metric_query, cos_metric_trend
  • Breakthrough narrative → cos_learn_narrative
  • Code pattern in repo → Grep (pattern) or Glob (filename)
  • Visual UI verification → playwright (if installed)
  • Lints/tests/builds → Bash (make targets per playbook)

Usage Rules

  1. Local SSOT first. Only use external tools when local docs are insufficient or recency matters.
  2. Prefer official sources: context7 or ref for framework docs. WebSearch for current facts.
  3. Check runtime first: verify tool availability via session's deferred tool list before depending on any non-built-in MCP.
  4. Escalation path: built-in tools → coding-os MCP → optional MCPs → WebSearch/WebFetch fallback.
  5. Record findings when research changes architecture: log date, versions, and source URL in task file.
  6. Do not document aspirational tooling as available unless runtime verification confirms it.

Update Policy

Review this inventory when a new MCP is added or an existing one is removed. Update the Verified on date and the relevant section.