Vessel Explore the architecture

Local-first agent infrastructure Under active development

Infrastructure for agents you can trust to act.

We're building the control plane for autonomous agents: governed access to models, tools and compute, with persistent execution and human control.

Explore the architecture View development status

Early-stage infrastructure. Built around operator control.

Conceptual architecture 01 / System overview
Operator intent · approval
approves & oversees
Vessel control boundary
Governed execution
Agent runtime acts only inside granted scope
requests
Models
  • Cloud
  • Local
  • Replaceable
Tools
  • Scoped MCP
  • Hosts
  • Deny by default

state · policy · visibility

Product direction. Unified platform in development.

The problem

The model is only one part of an agent system.

Once agents run commands, call APIs and maintain state, the interesting questions move outside the model. Who is allowed to do what? Where does the work run? What does the agent remember, and what can anyone see afterwards? When must a person step in?

Answering those takes permissions, execution, routing, memory, observability and escalation. Vessel is building that surrounding infrastructure.

  • 01Permissions
  • 02Execution
  • 03Routing
  • 04Memory
  • 05Observability
  • 06Escalation

Platform

The infrastructure around the agent.

Six connected capabilities. Prototypes establish the foundations; active development brings them into one control plane.

  1. Governed tool access

    Prototype

    MCP gateways with identity checks, scoped tool catalogs and deny-by-default access.

    Explicit capabilities at the tool boundary.

  2. Persistent execution

    In development

    Working toward long-running jobs that retain state and can resume after interruption.

    Durable state beyond a single conversation.

  3. Model routing

    Prototype

    Policy-based route recommendations informed by capability, privacy and available capacity.

    Automated routing across the platform is in development.

  4. Local-first operation

    Prototype

    Experiments with local inference and memory retrieval alongside cloud reasoning.

    Where work runs depends on configuration.

  5. Human control

    In development

    Explicit approval and escalation boundaries for consequential actions.

    Unified platform enforcement is in development.

  6. Observability

    Prototype

    Read-only usage history, plus tool and fleet snapshots.

    A foundation for understanding execution across systems.

Conceptual architecture / Product direction

A control plane. Your choice of intelligence.

Operator intent · approvals · oversight
sets intent, approves consequential actions
Vessel Control Plane
  • State
  • Policy
  • Visibility

A Execution branch

  1. delegates work
  2. Agent runtimeOpenBearClaw (OBC)
  3. requests a model
  4. Model routercapability · privacy · capacity
  5. routes to
    • Claude
    • OpenAI
    • Gemini
    • Local models

B Capabilities branch

  1. mediates access
  2. Policy + identity + approvalchecked before any call
  3. permits scoped calls
  4. MCP / tool gatewayscoped catalogs · deny by default
  5. reaches
    • Git
    • APIs
    • Infrastructure
    • Local hosts
    • Enterprise systems Planned

This shows the intended platform shape. Components exist as prototypes; the unified control plane is in development. Models are replaceable resources.

Why a control plane sits between the operator and the work

Agents that run commands and call APIs need a place where permissions, state and visibility are decided once, rather than re-implemented per tool or per model. The control plane is that place. It is the part of the system Vessel is building around.

What "governed" means on the capabilities branch

The intended design checks identity, scope and required approvals before forwarding a request. Current MCP prototypes scope the visible tool catalog and deny access that has not been granted. Unified approval enforcement is in development; enterprise integrations are planned.

What "your choice of intelligence" means on the execution branch

The router recommends where a request should run, based on capability, privacy and available capacity. Models sit behind it as replaceable resources with equal standing. Route recommendations are a prototype. Automatic, integrated routing is in development.

Local + cloud

Cloud intelligence. Local control.

Use cloud intelligence without making every workload a cloud workload. The intended design places work by capability, privacy, cost, latency, available hardware and policy.

Environment

Local

Close to your data and hardware.

  • Work configured as sensitive can stay on your own hosts.
  • Local inference and memory retrieval, where hardware allows.
  • Latency-sensitive or hardware-bound tasks.

Placement policy

  • Capability
  • Privacy
  • Cost
  • Latency
  • Available hardware
  • Policy

Intended placement. Actual behavior depends on configuration.

Environment

Cloud

Reach for it when it earns its place.

  • Frontier-model reasoning for tasks that need it.
  • Extra capacity when local hardware is busy.
  • Work that policy allows to leave the host.

Workload placement is an active design area. Data handling depends on the selected tools, model backends and operator configuration.

Development

Building the foundations.

Vessel is under active development. Current prototypes explore scoped MCP tool access, usage visibility, model-route recommendations and local inference.

We're working to bring these components together with durable execution and consistent approval boundaries. Longer-term direction includes enterprise and industrial integrations.

  1. Prototype Exists, being explored
    • Scoped MCP tool access
    • Usage visibility
    • Model-route recommendations
    • Local inference
  2. In development Being worked toward
    • Durable, resumable execution
    • Consistent approval boundaries
    • The unified control plane
  3. Planned Longer-term direction
    • Enterprise integrations
    • Industrial integrations

Founder

Built from the infrastructure up.

Vessel was founded by Robert Neal, drawing on experience across enterprise IT, operational technology, manufacturing systems, cybersecurity and AI automation.

The company is bootstrapped and focused on one thing: secure, governed infrastructure for autonomous AI agents, across local and cloud environments and independent of any one model provider.