Descles
response level LLM message control with HITL approval flows
Last verified:
What is Descles?
Descles is a control plane and model gateway for teams deploying AI agents, letting admins connect model providers and remote MCP tools once and grant agents scoped access with budgets, expiry dates, and approval policies. It mediates model and MCP tool calls, blocks or holds actions for human review (including via Slack), and maintains hash-chained audit records. It targets founders, AI, and platform teams running agents with credentials, write access, or customer systems.
Descles pricing
Pricing model: Freemium
Free tier with 1,000 requests and no credit card; model API usage billed by your provider with no model markup. Pilot costs discussed before starting.
Descles pros
- Agents get Descles keys and never receive stored company provider or connector secrets
- Policy-based approval workflow with Slack integration and hash-chained tamper-evident audit records
- Fine-grained delegation: per-agent providers, tools, budgets, and expiry
- Runs on your own infrastructure; no model markup and supports standard clients like Claude Code and OpenAI SDK
Descles cons
- Only governs actions routed through the Descles gateway—local shell and direct SaaS/API calls are outside its enforcement boundary
- Not a local shell sandbox; requires mediated paths for coverage
- Still in an early design-partner pilot phase (seeking 3–5 partners)
Frequently asked questions about Descles
Does Descles sandbox everything an agent does?
No. It governs tool calls routed through the Descles gateway, including remote MCP actions, but local shell and direct SaaS/API calls are outside its enforcement boundary.
Do agents get access to company API keys?
No. Agents receive a Descles agent key; stored provider and connector secrets are never shared with them.
How are high-risk actions handled?
Policies can block or hold structured tool calls before execution, routing them for human approval in the console or a configured Slack channel, with decisions recorded in the audit trail.
How do I get started?
Create a workspace, connect a model provider, issue an agent key with a budget and policy, then point a supported client (e.g., Claude Code or OpenAI SDK) at the gateway endpoint.