THE CONTROL PLANE FOR AI AGENTS

Autonomy, with
accountability.

Give your agents room to work.
Keep the authority to decide.

One control plane for model calls, tool execution, and resource access. Know what happened, why it happened, and who approved it.

Self-hosted OpenAI-compatible Human in control
SYSTEM OVERVIEW01 — 03
Your agentsN
desclesPOLICY · IDENTITY · AUDIT
Policy enforced
ModelGATEWAY
ActionGATEWAY
ResourceGATEWAY
Every action. A clear chain of trust.+
FITS YOUR STACK.
KEEPS YOU IN CONTROL.
OpenAI-compatibleMCPGitHubYour toolsSelf-hosted
01 / THE PLATFORM

Three gateways.
One source of control.

Agents cross boundaries all the time.
Descles puts a decision point at each one.

TOOL I/O

Permission before execution.

Evaluate tool calls against policy. Allow routine work, deny forbidden actions, and send sensitive operations to a human for approval.

Runtime enforcementHuman approvalMCP tool calls
policy.evaluateDESCLES / 02
# Every request has a responsible identity agent research-agent gateway action request github.create_issue policy require_approval
Human review before execution
02 / FOLLOW THE KNOWLEDGE

An agent is only as trusted
as what it acts on.

A sandbox controls what an agent can touch. Descles also tracks why it believes something—and whether that knowledge should authorize an action.

Read the product thesis
External sourceUnreviewed knowledge enters context
LOW TRUST
Trust follows the knowledge
Generated skillProvenance survives summarization
LOW TRUST
Privileged action requested
Human review requiredKnowledge cannot grant itself permission
03 / THE MOMENT THAT MATTERS

Let agents act.
Decide where they pause.

Try an illustrative approval flow.
You make the call. The trail stays clear.

descles / control room INTERACTIVE PREVIEW
ACTION REQUEST / 0042

Publish a GitHub issue

Awaiting review
REQUESTED BYresearch-agent
CAPABILITYgithub.create_issue
RESOURCEacme / knowledge-base
Low-trust knowledge requires a human decision.

This skill was generated from an external source. Its provenance stays attached when it requests a write action.

Illustrative data. No real action is executed.YOUR AGENTS. YOUR AUTHORITY.

An identity behind every action.

Scoped credentials and delegation chains connect agents to the people responsible for them.

Skills with a history.

Track source, version, declared permissions, and publish approval through the skill lifecycle.

A trail you can verify.

Durable, tamper-evident audit records with Ed25519 signatures preserve the decision trail.

04 / BUILT FOR BUILDERS

Your stack.
A clearer boundary.

Start the self-hosted demo with built-in scenarios. Explore the three gateways without a model API key.

Read the setup guide
Quick start GO
# Clone and start the local demo
git clone https://github.com/chiatzenw-cur/agent-control-plane.git
cd agent-control-plane
go run ./cmd/demo

# Open the console, then use the printed operator token
http://127.0.0.1:8088/demo/demo.html
BUILD WITH AUTONOMY. OPERATE WITH CONFIDENCE.

The next step for your agents?
A clear chain of trust.

Explore the demo

Available as a self-hosted prototype. Production hardening is ongoing.