The Operating System
for AI Agents
Per-agent identity, policy checks in front of agent actions, and an audit trail you can replay.
The Control Layer Every AI Stack Needs
ACL is not another chatbot wrapper. It is a full governance layer — identity, policy, audit, orchestration, and observability — for autonomous AI agents running in production.
Agent Identity
Every agent is its own principal, with its own role, tenant, and permissions — not a shared API key.
Policy-Gated Execution
The Gatekeeper reviews tool calls before they execute. The Sentry evaluates outputs before they are delivered. Side-effecting tools sit behind a permission gate and human approval.
Audit Trail
Write-once audit log tables record agent decisions, attributed to the agent that made them. Replay reconstructs a workflow run from its records for forensic investigation.
Multi-Agent Orchestration
Sequential and parallel workflow execution with typed data contracts between agents. Human-in-the-loop gates, shadow mode, and auto-retry built in.
Multi-Tenant SaaS
Complete tenant isolation enforced at the Postgres data layer with 170+ Row-Level Security policies. Hierarchical MSP structure with white-label branding support.
Mission Control Observability
Real-time health score, execution traces, cost breakdown, security event feed, and SLA-tracked escalation queue. Visibility into what your agents are doing.
Agents Are Not API Calls
- ✕Agent identity = an API key (static, long-lived, shared across instances)
- ✕No capability enforcement — agent can attempt any tool call it knows about
- ✕Audit log = application log (mutable, no chain, no attribution)
- ✕Prompt injection can cause agents to take any action in their configuration
- ✓Agent identity = its own principal, with its own role, tenant, and permissions
- ✓Gatekeeper reviews tool calls; side-effecting tools sit behind a permission gate
- ✓Audit log = write-once audit tables, attributed to the agent, with run replay
- ✓3-layer prompt injection defense: Bouncer + Sentry block unauthorized actions
Capability Enforcement
Each agent's permissions define what it may do. Side-effecting tools sit behind a permission gate, and the Gatekeeper reviews tool calls before they run.
Authentication Paths
Four distinct principal types: Human (Supabase session + MFA), Agent (per-agent identity), Service Account (machine JWT with purpose binding), API Key (HMAC-hashed with IP allowlist).
Per-Action Attribution
Tool calls and LLM invocations are attributed to a specific agent and recorded in write-once audit log tables.
Five Layers Between the Internet and Your Data
Defense-in-depth: a single boundary defeat does not produce tenant data exfiltration. Multiple layers must fail simultaneously for data to be at risk.
Pre-Built Security Guards
Reviews tool calls before execution and approves or denies them against the agent's permissions.
Policy evaluation on agent outputs before delivery. Returns structured verdicts — approved or blocked with reason and severity.
Real-time prompt injection firewall. Layer 1 regex scoring (>50 = BLOCK), Layer 2 LLM analysis for edge cases. Two-layer defense against prompt injection.
5-agent coordinated adversarial security testing using OWASP vectors. On-demand or scheduled adversarial testing; findings logged to audit trail.
Audit Log — Write-Once by Design
Audit log tables are write-once at the Postgres layer — a database constraint, not a soft “do not delete” convention.
Audit records are tied to the agent that acted, so you can see which agent did what.
Replay reconstructs a workflow run from its records, so you can review what an agent did and why.
RBAC with Grant Guards
5 roles: Owner, Admin, Editor, Viewer, Guest. Fine-grained per-entity permissions. canGrantPermissions check prevents privilege escalation — you cannot grant permissions you don't hold.
Zero-Downtime Key Rotation
API keys HMAC-hashed at rest. IP allowlisting per key. Granular permission scoping and per-key rate limits. Rotation completes without any service interruption.
Circuit Breaker
Automatic failure isolation for external service calls. Prevents cascading failures when third-party integrations degrade. Status monitored in real time via Mission Control.
Mission Control — Full Visibility into Every Agent
Real-time health scoring, execution traces, cost breakdown, and security event feed. Operators know exactly what is happening, what it costs, and whether anything needs attention.
Workflow Metrics
Total runs, success rate, average duration, cost, and state breakdown (running / paused / completed / failed). Trend lines over configurable time windows.
Agent Leaderboard
Execution count, success rate, cost, and top-performing agents ranked. Identify which agents deliver the most value and which need attention.
Security Event Feed
Real-time feed of all security events: severity breakdown, active alerts, cross-tenant access events, and permission denials.
Cost Breakdown
Total spend by workflow and agent with trend lines. LLM token costs attributed per agent execution. Budget alerts and per-tenant rate limiting.
SLA-Tracked Escalations
Human-in-the-loop queue with SLA tracking. Know which approvals are pending, overdue, or blocked. Full escalation audit trail per item.
Full Execution Traces
Per-agent traces: tool calls, inputs/outputs, LLM usage, cost, timing, and error logs with stack traces. Everything needed for debugging or review.
Orchestrate Any Business Process with AI
Multi-agent workflows with typed contracts, human oversight gates, governance charters, and shadow mode testing — all with an audit trail you can replay.
Multi-Agent Orchestration
Sequential and parallel agent execution. Asynchronous job queuing with priority support. Configurable concurrency limits per engine. Idempotency key support for replay safety.
Human-in-the-Loop Gates
Workflows pause at configurable checkpoints for human review, feedback, approve, reject, or escalate. Webhook-based resume. SLA tracking on pending items. The agent waits — it does not proceed without authorization.
Governance Charters
Define the rules every agent in a workflow must follow — permitted actions, data access scope, escalation triggers, and output constraints.
Step Audit
Workflow steps — input, output, tool call, LLM invocation, human decision — are recorded in write-once audit log tables, and replay can reconstruct the run.
Shadow Mode
Run a new agent or workflow version in parallel with production without affecting live operations. Compare outputs side-by-side before promoting the change.
WorkflowSmith
Coming SoonAI generates workflow definitions from natural language descriptions. Describe your business process; WorkflowSmith produces a deployable workflow graph with appropriate agent assignments and gates.
Typed Data Contracts
Strict schemas enforced at every agent boundary. Agent B cannot receive malformed output from Agent A — the contract layer rejects it before it propagates. Agent-to-agent event bus (publish/subscribe) for decoupled coordination.
AI BDR — 6-Agent Outbound Sales Pipeline
A production-ready, end-to-end Business Development pipeline. From campaign brief to sent email in a single workflow, with human review gates at every critical decision point.
This is not a demo. This pipeline runs in production. It brings together Gatekeeper policy checks, typed data contracts between agents, human-in-the-loop gates, and the audit trail — in a single workflow.
Translates a campaign brief into targeting criteria, qualification rubric, and outreach sequence plan. Defines the ideal customer profile that all downstream agents work from.
Searches 70M+ B2B contacts for decision-makers matching ICP criteria. Scores each prospect 0–100 against the rubric. Uses Apollo, Hunter.io, People Data Labs, ZoomInfo integrations.
Deep research per prospect: company news, pain points, tech stack, hiring signals, recent funding, personalization hooks. Grounds every outreach in specific, current intelligence.
BANT scoring (Budget, Authority, Need, Timeline). Classifies each lead as Hot, Warm, Nurture, or Disqualified with reasoning. Routes to human review before any outreach is drafted.
Writes personalized emails citing specific research from Account Intelligence. 150-word cap. Soft call-to-action. Each email references the prospect's actual situation — not a mail-merge template.
Sends approved emails via SendGrid or SMTP with domain reputation management. Tracks delivery, manages follow-up sequences, handles bounces. Stops all sequences when a reply comes in on any channel.
70M+ Contact Database
Apollo, Hunter.io, People Data Labs, Enrich.so, ZoomInfo, Clearbit — all unified behind a single search_prospects tool with consistent scoring.
Multi-Channel Follow-Up
Email, Voice (TalkMage), SMS (Twilio), LinkedIn-ready. Reply on any channel stops all sequences. Cross-channel coordination is automatic.
Full Audit Trail
Prospect scores, email drafts, approvals, and sent messages are recorded in the audit trail with attribution to the agent that made the decision.
70+ Integrations — One Unified Platform
Integrations are available to agents as tool calls from the tools registry, alongside the same policy checks and audit trail.
Provider routing with automatic failover — if one LLM provider fails, the next in the configured chain is used automatically. No code change required. No manual intervention. The circuit breaker isolates the failure and routes around it.
18 Native Tools — Available to Every Agent
Ground Every Agent in Real, Tenant-Specific Knowledge
Upload anything. Agents search it. Responses cite sources. PII detected at ingestion. Every document is tenant-isolated at the database layer.
Any Format
PDF, DOCX, PPTX, XLSX, TXT, web URLs, images, audio, video, code. Automatic extraction, chunking, and embedding at ingestion time.
Hybrid Search
Vector search + keyword search combined with query enhancement and reranking. Agents receive the most relevant chunks, not just the most similar ones.
PII Detection & Isolation
PII detected during ingestion. All documents are isolated to the uploading tenant's RLS scope.
Run ACL as Your Own Platform
Managed Service Providers can brand, manage, and bill for ACL as their own product. Complete white-label with full audit trail for every support interaction.
Custom Branding
Custom domain, logo, and color branding. Your clients see your product, not ACL. Full white-label from login screen to report exports.
Child Tenant Management
Manage all client tenants from a single admin view. Bulk creation, maintenance window scheduling with client notification, and tenant transfer between MSPs with approval workflow.
Break-Glass Access
Time-limited support access tokens with full audit trail. Every break-glass entry is logged — who accessed, when, which tenant, what they did.
Per-Client Cost Tracking
Detailed cost breakdown by client, workflow, and agent. Generate billing reports per tenant. Know exactly what each client's agents are costing and billing accordingly.
Hierarchical Structure
Parent-child tenant hierarchy. MSP parent has governance visibility into all child tenants. Child tenants operate independently with full isolation from each other.
Marketplace Access
Publish agent templates and workflow templates to the marketplace. Share across your client tenants or make available to the broader ecosystem with version control.
Complete Isolation — Not Convention, Enforcement
Tenant isolation is not enforced by application code alone. It is enforced at the Postgres row level, on every table, on every query, by the database engine itself.
170+ Row-Level Security Policies
Every table with a tenant_id column carries a RESTRICTIVE RLS policy. Queries that bypass application-layer auth still cannot read another tenant's rows — the database enforces the boundary on every access.
Permission-Tier RLS
Sensitive tables carry a second tier of RLS policies. Not only must the row's tenant match — the querying principal must hold the specific permission required for that table. Two independent RLS checks on every sensitive read.
UPDATE Smuggling Prevention
WITH CHECK clauses on all UPDATE policies prevent rows from being moved between tenants via an UPDATE statement. A tenant admin cannot reassign another tenant's row to their own namespace.
Tenant Provisioning
New tenant creation copies default templates atomically. Per-tenant configuration: LLM provider selection, rate limits, feature flags, and integration credentials — all isolated.
Operator Cross-Tenant Access
Support operators hold an is_operator claim in their JWT — but the claim is verified live on every request. Revocation is immediate; stale JWTs are rejected. Every cross-tenant access is logged in the audit trail.
System Tenant Isolation
Platform-admin privileges are confined to the System Tenant by a database trigger. No code path can escalate from tenant-admin to platform-admin without an explicit, logged operator grant.
250+ API Endpoints — Build Anything on Top of ACL
Full programmatic access to every platform capability. Service accounts, webhook triggers, and JWT-based service-to-service authentication for deep integrations.
250+ REST Endpoints
Every platform action exposed via API. Workflow management, agent configuration, audit log export, tenant provisioning, and more — all accessible programmatically.
Service Accounts
Machine identities with purpose binding and granular permission scoping. JWT-based authentication. Purpose-bound tokens prevent lateral movement if a service account is compromised.
Webhook Triggers
External systems can trigger workflow execution, inject events, and receive completion callbacks via webhooks. Every inbound webhook is validated and attributed in the audit trail.
Invite AI to Your Board Meetings
Coming SoonAI that doesn't just assist in meetings — AI that participates. Board of Directors brings structured AI agents as named participants in your strategic meetings, contributing research, analysis, and recommendations with the same audit trail as other ACL actions.
Named AI Board Members
AI agents join your board as named participants with defined roles: Chief Research Officer, Devil's Advocate, Risk Analyst, Market Intelligence. Each has a specific charter and perspective.
Pre-Meeting Intelligence Briefs
Board members prepare independently before each session. Every agenda item arrives with research, analysis, and position papers — authored by AI and attributed to the agent.
Accountable Participation
Every AI contribution is attributed to a specific agent and recorded in the audit trail. Who said what — including the AI.
Governance by Charter
Each AI board member operates under a Governance Charter defining their scope, permitted data access, and escalation triggers. Their role is fixed until you change it.
Board of Directors ships immediately after MVP launch. Design partners get early access.
Become a Design Partner →Agents without governance are a liability. Agents with governance are a competitive advantage.
Every enterprise that deploys AI agents faces the same questions: Who authorized this action? What did the agent see? Can we review it afterward? ACL is built to answer those questions with records, not convention.