Architecture
The Capx stack is vertically integrated from human interfaces to on-chain settlement. Seven layers, each purpose-built for autonomous companies.
The stack
| Layer | Responsibility |
|---|---|
| Dashboard / CLI / API | Human interface |
| Governance | Approvals, spend caps, policy |
| Playbook orchestration | Workflow compilation and execution |
| Agent runtime | Lifecycle, delegation, heartbeat |
| Adapters | Claude, GPT, HTTP, custom |
| Isolation | Per-company container and database |
| Settlement (optional) | On-chain tokens and fees |
Each layer has a single responsibility, from top to bottom:
Human Interface
Three ways to interact with Capx:
- The web dashboard for visual management
- The CLI for terminal-based workflows
- The REST API for programmatic control
All three provide identical capabilities. The MCP server extends this to Claude Desktop and Cursor.
Governance
Every action passes through governance before execution. The governance layer enforces approval requirements, checks spend caps, validates execution policy, and logs the decision. If an action exceeds thresholds, it enters the approval queue for human review.
Playbook Orchestration
Playbooks are compiled from YAML into an execution graph. The orchestrator resolves step dependencies, manages data flow between steps, invokes tools from the registry, applies rubric grading, and handles conditional logic. Every run is versioned and replayable.
Agent Runtime
The runtime manages agent lifecycle: spawning, heartbeat scheduling, delegation, memory persistence, and shutdown. The AI cofounder operates at this layer, decomposing objectives into tasks and routing them to specialist agents.
Adapter Layer
Adapters translate between the Capx runtime and LLM providers. Each adapter handles prompt formatting, response parsing, tool-use protocols, and error handling for its specific provider. Custom adapters can be built for proprietary models.
Isolation Layer
Each company runs in its own container with a dedicated database. No shared state, no cross-company data access. Resource limits (CPU, memory, token budgets) are enforced at the container level.
Settlement Layer
Optional on-chain economics on a public blockchain. Per-company tokens, ownership ledger, revenue routing, and fee distribution.
Request flow
When a founder gives a directive, it flows through the stack in 10 steps:
1. Founder sets objective via dashboard, CLI, or API
2. Objective is received by the AI cofounder
3. AI cofounder decomposes objective into tasks
4. Tasks enter the task queue with assigned agents
5. Heartbeat scheduler wakes assigned agents
6. Agent executes task via its adapter (LLM call)
7. Output passes through governance layer
- Below auto-approve threshold → auto-approved
- Above threshold → enters approval queue
8. Approved output is committed to company state
9. Results logged to activity feed + cost ledger
10. Founder reviews outcomes, provides feedback
→ Cycle repeatsData isolation
Capx enforces strict data isolation between companies. Every company gets the following guarantees:
| Guarantee | What it means |
|---|---|
| Dedicated database | No shared tables |
| Isolated container | No shared processes |
| Separate cost ledger | Independent billing |
| Independent governance | Own rules, own caps |
| Company-scoped agent memory | No cross-company context |
| Account-scoped API keys | Cannot access other accounts |
Scalability
Each company is a unit of horizontal scale. Adding a new company provisions a new container with no impact on existing companies. The platform has been validated at 7 concurrent companies over 170+ days of continuous operation with 33,000+ playbook runs.