Deploy teams of autonomous AI agents governed by versioned contracts, progressive trust levels, and measurable outcome intelligence — with semantic memory, vector search, and human oversight built in.
Each capability is accessible through a unified tool system. Agents invoke tools by name, and the platform handles authorisation, execution, and result tracking automatically.
Deploy agents with unique personas powered by Claude. Run autonomously via scheduled heartbeats or on-demand wakeups, with configurable model selection and monthly budgets.
Agents have managers and direct reports. They delegate downward, escalate upward, and send structured briefings to peers — mirroring how real teams operate.
SQL Server graph tables store facts and relationships while Qdrant provides semantic vector search. Agents recall context through both keyword matching and embedding similarity.
Full task lifecycle with checkout locks, dependencies, priorities, and automatic routing. Strategic goals cascade into concrete tasks agents execute.
Multi-step workflows with agent turns, conditions, delays, loops, parallel execution, and for-each iteration. Trigger manually, by schedule, via webhook, or by platform event.
Sensitive operations require human approval. Agents request handoffs when stuck. Approvals auto-escalate if ignored — nothing falls through.
Native integrations for Twitter, Instagram, TikTok, LinkedIn, and Microsoft Graph email. Custom APIs connect through the connector framework.
Tool-level RBAC ensures agents call only authorised tools. Resource access policies, encrypted secrets, and a complete audit trail of every action.
Ingest documents that are chunked and embedded via OpenAI embeddings. Agents search business knowledge through cosine similarity in Qdrant — with keyword fallback.
Distributed caching, event pub/sub, and rate limiting via Redis. Eliminates repeated database queries and provides the backbone for multi-instance scaling.
Versioned operating agreements defining role, mission, allowed actions, escalation rules, and forbidden behaviors. Immutable once published through draft, staging, certified, and production channels.
Six progressive trust tiers from observe-only to fully autonomous. A policy matrix evaluates every side effect by risk, financial impact, and agent trust level.
Every external action creates an auditable record with evidence bundles and decision narratives capturing trigger, reasoning, policy context, alternatives, and outcome.
Persistent case objects that group tasks, handoffs, and actions into durable multi-day workflows. State machine tracks open, waiting, blocked, and resolved states.
Test suites validate agent behavior against contract versions before production promotion. Replay historical runs as regression scenarios to catch behavioral drift.
A/B test agent variants with traffic splitting. Compare models, personas, and skill configurations side by side. Declare a winner based on outcome metrics.
Agents extract learnings from conversations with confidence scores. High-confidence insights are promoted to the system prompt, making agents smarter over time.
Department-native work models: Customer Ops with cases and queues, Revenue Outreach with campaigns and prospects, Back Office with work items, and Scheduled Ops with playbooks.
Built-in site management with code changes, Docker build and push, deployment to hosts, and health check monitoring — all accessible as agent tools.
Inbound webhooks trigger rules that create tasks, start workflows, or wake agents. JSONPath conditions filter events before dispatch.
Agents wake up, gather context, reason about what to do, invoke tools to take action, and persist what they've learned for next time.
Agents wake via scheduled cron jobs, task delegation, escalation events, or direct user messages.
The platform runs semantic vector search across memory and knowledge, assembles recent episodes, unread briefings, and pending tasks into a rich context window — cached in Redis for sub-millisecond access.
Claude processes the context through the agent's persona and generates tool calls or text responses with full planning capability.
Each tool call passes through RBAC authorisation before dispatching to the appropriate handler. Domain entities update in real time.
Every side effect passes through the action policy matrix. Risk level, financial impact, and autonomy tier determine whether the action auto-executes, requires approval, or is blocked.
High-risk actions enter a review queue with full evidence bundles and decision narratives. Supervisors approve, reject, edit drafts, or request more evidence before execution proceeds.
Facts are stored in the SQL graph and simultaneously embedded as vectors in Qdrant. Conversations are episodised with key entities and importance scores — enabling both structured traversal and semantic recall.
Every tool call is authorised, executed, and logged. Agents interact with the world through a unified dispatch layer.
Agents operate under versioned contracts with progressive autonomy. Every action is evaluated by policy, every decision is narrated, and every outcome is scored — while low-risk operations run at full speed.
Versioned agreements defining role, mission, allowed systems, forbidden behaviors, evidence requirements, and budget rules. Contracts progress through draft, staging, certified, and production release channels. Immutable once published — changes require a new version.
Six trust tiers from observe-only to fully autonomous. Agents graduate as they prove reliability through certification and outcome scoring.
Rules evaluate every side effect by action type, resource type, risk level, and financial threshold. Per-agent or business-wide policies.
Every approved or rejected action includes a structured narrative: trigger event, reasoning chain, policy context, alternatives considered, and observed outcome. Full transparency into why an agent acted — or didn't.
Test suites run scenarios against contract versions. Agents must pass certification before production promotion. Historical runs replay as regression tests, catching behavioral drift before it reaches live operations.
Quantifiable business impact: resolution rate, approval rate, estimated value delivered in cents, and a 0–100 holistic health score.
Every tool call, approval, escalation, and handoff logged with timestamps, agent identity, and full context.
API keys stored encrypted via Azure Key Vault. Agents reference secrets by name — never by value.
Every agent has a monthly token budget. Cost events tracked per API call. At the limit, execution stops.
Sliding window counters in Redis enforce API rate limits across all instances. Replaces in-memory counters with distributed state, so scaling out doesn't bypass limits. Event pub/sub enables real-time notifications across the platform.
Policy-based access control evaluates every external resource request. Learning mode allows gradual policy rollout — observe and log before enforcing.
Configurable event-to-channel routing with templates. Route approvals, escalations, and alerts to the right team through the right channel automatically.
Each layer has a well-defined responsibility and communicates through interfaces. Multi-tenancy is enforced at every level.
The complete platform runs as a Docker Compose stack. Every service has health checks, persistent volumes, and graceful degradation when dependencies are unavailable.
If Qdrant is down, search falls back to SQL keyword matching. If Redis is unavailable, caching falls back to in-memory. If no embeddings API key is set, zero vectors are returned and vector search is skipped. The platform never hard-fails on infrastructure dependencies.
Bring your own Anthropic or OpenAI API key. You pay the AI providers directly at their rates. The Clovar license is purely for the orchestration platform.
Any business process that can be broken into tasks, assigned to roles, and governed by policies can be orchestrated by Clovar agents.
Clovar is the operating system for your AI workforce. Agents operate under versioned contracts with progressive autonomy, every action is policy-evaluated and narrated, and outcomes are scored for continuous improvement. Combined with vector-augmented memory, semantic knowledge search, and Redis-backed infrastructure — you get autonomous operations with full accountability.