Vouch request: [Apelogic]

[ApeTuzik]

What do you want to work on?

TLDR:

Possibility for OpenShell policy gateway to work well with agent gateway for MCP tool routing and LiteLLM for LLM routing:

• LiteLLM as the model router
• improve observability
• tenant policy improvements
• add an OpenShell external-gateway policy resource
• change OpenShell’s inference routing model

Components

Gateway Layer

• Workspace / tenant policy — per-tenant configuration and access rules
• External gateway adapters — integration points for external systems
• Reconciliation status — drift detection and policy sync state

Sandbox Supervisor

Receives a scoped policy bundle from the gateway and enforces it across two planes:

Inference Plane

Component	Role
inference.local	Model allowlist, credential isolation
LiteLLM	Model routing, budgets, rate limits, spend tracking
Model providers	Upstream LLM APIs

Tool Plane

Component	Role
L7 MCP policy proxy	Process-level and tool-level enforcement
agentgateway	MCP federation and RBAC
MCP backends	Downstream tool servers

Observability

Both planes emit OCSF + OpenTelemetry + LiteLLM spend events into a unified audit backend.

Why this change?

• We need such things for enterprise clients, where an admin can set policy per tenant centrally. I.e. tenant X can have $100/month access to GPT 5.5 and tool access to drive and slack.
• This policy should be set centrally in OpenShell gateway and propagate out to sandbox (if present), agent gateway and LiteLLM. If policy updates, OpenShell should be able to push the policy out immediately...
• This way you don't go to 3 different places to configure access to a single agent tenant and there is no policy drift.

Checklist

• I wrote this request myself (not AI-generated)
• I have read CONTRIBUTING.md