The control plane for enterprise agent systems
Air-traffic control for your AI agents.
AgentGuide puts registered agents and gateway-routed model and tool traffic on one radar. Observe everything, govern with policy, and orchestrate fleets of agents that collaborate on real enterprise work — all through a single gateway you run yourself.
Why a gateway
Agents are easy to launch. They are hard to run.
Once agents call models, invoke tools, and spawn other agents, nobody can answer the basic questions: what did it cost, who approved that action, which agent broke production? AgentGuide sits on the path of routed calls and turns agent traffic into something you can see, control, and compose.
Gateway traffic becomes structured evidence
LLM requests, MCP tool calls, and agent turns produce persisted usage events — tokens (including cached and reasoning tokens), latency, and errors. Agent-bound traffic is attributed when ownership is unambiguous; builtin turns carry explicit attribution.
Observability →Policy at the gateway, not in every app
Virtual keys for consumers, pooled and scheduled upstream credentials, interactive permission gates for agent actions, recursion-depth limits, and fail-closed concurrency and timeout controls.
Governance →Compose agents into real workflows
Run agents inside the gateway today with declarative topologies — sequential, parallel, supervisor, plan-execute and more. Gateway-owned tasks and workflows across independent agents are the next control-plane layer.
Agents & orchestration →Not another agent framework — AgentGuide orchestrates agents from the outside, whatever they're built with. How it relates to LangGraph, AutoGen & co →
Three identity models, one control plane
Bring the agent you have. Or build it in the gateway.
Every registered agent — wherever it runs — gets one first-class identity: routes, resources, policy, usage, and health, all managed from a single control plane.
Builtin agents
Declared as configuration, materialized inside the gateway — no separate process. The deepest option: inner model calls and tool executions are traced individually, and orchestration topologies are first-class.
ACP agents
The gateway owns the lifecycle of Agent Client Protocol agents such as Codex and OpenCode: process pooling, sessions and transcripts, streaming turns, and permission decisions routed through your policy.
HTTP agents
Agents on your infrastructure can register as first-class identities with resource and route references. Direct HTTP task dispatch is part of the external-task roadmap; it is not yet an execution backend.
Product direction · multi-agent collaboration
Workspaces and workflows turn agents into a workforce.
Single agents answer questions. Fleets of agents deliver projects. AgentGuide is built toward the second. Today it provides the runtime, identity, topology, permission, and observability foundations; shared mission workspaces and workflows across independent agents are the next layer.
See the agent in one place
The per-agent workspace aggregates definition, runtime state, linked resources, pending permissions, usage, interactions, and shallow health.
Compose inside one governed turn
Inline sub-agents collaborate through plan-execute loops, supervisor delegation, and parallel fan-out, with model and tool calls traced beneath the turn.
Coordinate independent agents
Durable tasks, scheduling, cross-runtime hand-offs, artifact lineage, budget enforcement, and compliance-grade audit are roadmap capabilities built on the foundations shipping today.
The foundation
A serious LLM & MCP gateway underneath.
Orchestration is only as trustworthy as the layer it stands on. AgentGuide ships a production LLM gateway and MCP gateway in the same binary.
One API over every provider
- OpenAI, Anthropic, Gemini, DeepSeek, Qwen, Zhipu, OpenRouter, Ollama and more
- OpenAI- and Anthropic-compatible endpoints, plus a Claude Code-ready profile
- Logical model routing: one model name, many candidate providers, automatic fallback
- Credential pooling with scheduling — including CLI subscription credentials (Codex, Claude Code)
Tools as governed infrastructure
- Register MCP services once; expose them through routes with auth policy
- Tool discovery, execution, and in-flight request inspection
- Fail-closed tool selection — a missing tool is an error, never a silent skip
- Persisted tool calls share one event model, with agent attribution when ownership is clear
Get started
One binary. Your infrastructure. Full control.
AgentGuide deploys as a single self-hosted binary — no required control-plane SaaS. Outbound traffic goes only to providers and services you configure.