Integrations
Every PaloNexus integration follows one thesis: keep the existing runtime, add accountable
authorization. There is no new agent framework to adopt, no agent rewrite, and no move
into a PaloNexus-hosted environment. The framework keeps the agent loop; the sandbox
keeps the workspace; PaloNexus adds the missing layer — an accountable owner, a validated
delegation, task-scoped short-lived access, and a verifiable authority trail — behind a
single deny-by-default /authz decision.
Integrations attach that decision in one of three enforcement modes: in governed tool mode, PaloNexus wraps or hosts the tool so credentials never reach the agent (this is how the LangChain, LangGraph, and Deep Agents adapters work); in token exchange mode, PaloNexus issues an ephemeral, scoped credential to a trusted runtime component (the Security Token Service (STS) in Connect agents to enterprise authority); and in egress gateway mode, the agent’s outbound request traverses a PaloNexus gateway that authorizes it and injects credentials after the untrusted boundary (today’s network-layer egress enforcement on Kubernetes is one implementation of this mode).
Ecosystems
Section titled “Ecosystems”| Ecosystem | How it integrates | Status |
|---|---|---|
| LangChain | guarded_tool + middleware(pn) gate every tool (and optionally model) call through /authz | Available — SDK adapter |
| LangGraph | governed_node gates a graph node; deny → interrupt → human approval → resume | Available — SDK adapter |
| Deep Agents | tool_guard + governance_middleware + the shipped palonexus-governance skill | Available — SDK adapter |
| Agent-to-agent (A2A) delegation | The A2A hop is itself gated at /authz and carries the original on-behalf-of human subject | Available — recipe |
| kagent | Register kagent-deployed agents, resolve owners, gate tool and Model Context Protocol (MCP) calls, inject short-lived credentials | Planned — design preview |
| Kubernetes Agent Sandbox | No standing credentials in the sandbox; egress restricted to a PaloNexus gateway | Planned — design preview |
| OpenAI Agents SDK | Harness hooks call PaloNexus before sensitive operations; secrets stay outside the sandbox | Planned — design preview |
| MCP gateway | A governed gateway between agents and MCP servers, authorizing and credentialing each tool call | Planned — design preview |
Pages marked Planned carry a design-preview banner and describe intended behavior only — nothing on them is installable today. That is deliberate: every capability claim in these docs maps to shipped code or says so.
Where to start
Section titled “Where to start”- Using LangChain, LangGraph, or Deep Agents today? Start with Keep secrets outside Deep Agents sandboxes and the SDK adapter pages it links.
- Composing multi-agent systems? Read Govern agent-to-agent delegation.
- Wondering what PaloNexus deliberately does not do? See What PaloNexus is not.