LabscoMCP early access ↗
ContoriumLabsMCP SERVER · OFFICIAL PROJECT

Contorium Runtime MCP

Keep a project's decisions, reasoning and constraints in a local layer every AI tool can read — and find out which of those decisions no longer hold.

by ContoriumLabs ↗Agent Memory✓ Official

LABSCO SUMMARY

Git remembers what changed; this is built to remember why — and to notice when the why stopped being true.

The decision lifecycle is the part that earns its keep. A decision does not sit valid forever: it moves through warning, decaying, suspected invalid and needs revalidation as the assumptions under it change, and the review queue is what surfaces the ones worth looking at. Everything lives in a directory inside the project, so switching from one AI tool to another does not reset the accumulated understanding. Be aware of the size: this is a lot of surface, and the descriptions themselves point you at the preferred read for each job — inspect_state over get_cognitive_state, get_decision_context over the legacy control one.

What it is

A project intelligence layer stored locally in a .contora/ directory beside your code. It preserves intent, decisions, reasoning, structure, history and current state so a new session can read the project's accumulated understanding instead of reconstructing it. The surface is large, and organised around three operations: capture, inspect, transfer.

What you get

  • Capture as a deliberate act: record the current focus, a note, or a decision with the reasoning behind it, and record project direction as an intent (capture_focus, capture_note, capture_decision, record_project_intent).
  • Targeted reads over the stored layers — state, intent, decisions, timeline, graph, confidence, health, why, impact, evolution and provenance — each with its own inspect tool rather than one catch-all query (inspect_state, inspect_intent, inspect_decision, inspect_timeline, inspect_graph, inspect_confidence, inspect_health, inspect_why, inspect_impact, inspect_evolution, inspect_provenance).
  • A plain question answered against the connected knowledge rather than by grepping files, with suggested questions when you do not know what to ask (ask_project, get_suggested_questions, explain_context).
  • Decision validity tracked as a lifecycle rather than assumed permanent: pending decisions surfaced for review, committed or ignored, with the evolution of a decision readable afterwards (inspect_pending_decisions, review_pending_decisions, commit_decision, ignore_pending_decision, get_decision_evolution, get_review_queue).
  • Provenance for a decision — where it came from and what it rests on — derivable, traceable, snapshot-able and exportable (derive_decision_provenance, derive_decision_trace, decision_snapshot, get_provenance_chain, export_decision_provenance).
  • A project knowledge graph connecting intent to module to file to function to dependency, with impact and blast-radius queries over it (get_project_knowledge_graph, get_project_graph, get_impact_graph, get_blast_radius, get_understanding_graph).
  • History in the shapes people actually ask for: recent changes and events, the module's own history, the project's timeline, and a narrative version of how it got here (get_recent_changes, get_recent_events, get_module_history, get_project_timeline, get_project_story, get_project_journey).
  • Transfer as a first-class operation: a compact context payload, a handoff, or the full intelligence exported into a fresh session in another tool (transfer_context, transfer_handoff, transfer_intelligence, transfer_story, synthesize_context_payload, prepare_execution_context).
  • Health and readiness signals on the layer itself — confidence and stability indices, knowledge health, state conflicts, and readiness checks before a cycle runs (get_confidence_index, get_stability_index, get_knowledge_health, get_state_conflicts, inspect_system_ready).
  • An optional cognitive overlay, off by default: mode A is pure observation, and mode B adds insight bundles, skill discovery from a local registry plus GitHub and npm, and a task-mode preset — all read-only, with links displayed rather than anything installed (get_cognitive_mode, set_cognitive_mode, get_cognitive_insights, get_skill_suggestions, get_model_preset).

Requirements

Nothing beyond the package: everything is stored locally in .contora/ inside the project, with no account and no provider tie. It ships as a CLI and an MCP server from the same package, and integrates with Cursor, Claude Code, Gemini CLI, Codex, VS Code and any MCP client. The default mode is pure observation; the cognitive overlay has to be switched on.

Setup effort

One command — npm install -g @contorium/mcp

See installation options