ConnectLMT is not an open-ended chatbot. It is a directional, provider-agnostic pipeline of nine specialized agents, deployed across three surfaces so that de-escalation, mediation, and professional handoff each get the right level of intervention.
Each surface has a different job. Together they cover a conflict from the first drafted message to a professional handoff.
Where a message gets written before anyone else sees it. This is where escalation gets caught earliest.
Every draft is scored for aggression, sarcasm, and accusatory language before it can be sent.
A revised version is suggested that keeps the intent but removes the parts likely to escalate.
Revisions are suggestions. You choose what actually goes to the other party.
Drafts stay in your channel until you choose to send them into the shared space.
A layer above the conversation that watches for patterns neither party is positioned to see: repeated escalation, abuse indicators, or a matter that has outgrown what mediation alone can resolve.
Monitors the conversation over time, not just message by message.
Flags abuse or emergency indicators for a human to review, rather than continuing to mediate.
When a matter needs to move to a professional, this layer produces a structured, fact-checked case summary.
Summaries are built to be handed to HR, a therapist, a mediator, or a lawyer without re-litigating the conversation from scratch.
ConnectLMT is not an open-ended multi-agent chat where anything can happen. It is a fixed pipeline: tone analyzer, message clarifier, message revisor, mediator, legal advisor, summary generator, fact-checker, and routing components, each with one job, always in the same order.
Predictability is what makes neutrality trustworthy.
An open-ended multi-agent system can be steered, argued with, or gamed by whoever is more persistent. A fixed pipeline can't be talked out of being neutral. Every message passes through the same stages, in the same order, whichever party sent it. That constraint is the point, not a limitation.
Each of the nine agents is independently configurable across model providers, so cost and capability can be tuned per agent rather than for the system as a whole.
Each agent in the pipeline, tone analyzer, mediator, fact-checker, and the rest, can run on OpenAI, Anthropic (Claude), Google (Gemini), xAI (Grok), or a local model.
Lighter-weight tasks like tone scoring don't need the same model as fact-checking a legal summary. Provider choice is tuned per agent, not applied uniformly.
If a provider changes pricing, availability, or policy, the affected agent can be re-pointed without redesigning the pipeline.
Conversation data and vector memory are stored locally first (SQLite plus a local vector store), not routed through a third-party data lake by default.
SQLite and local vector memory keep conversation history close to the user rather than centralized by default.
Consent controls, data export, and deletion are on the product roadmap as core, not optional, features.
Legal-advisor agent output and case summaries are informational and are not a substitute for advice from a licensed lawyer.
No obligation. We'll follow up to see if ConnectLMT fits what you're dealing with.