By You have Claude Code in one terminal, Codex in another, and Gemini in a third. You are not writing code anymore — you are a switchboard operator. AgentProto is an open-source TypeScript runtime that flips that dynamic: one daemon manages all of your agents, checks their work before anything commits, and exposes the whole stack through a unified interface. The September 22 release added opaque-origin MCP widget support and a new evaluation package. The September 17 release shipped the AIP-57 model routing reference implementation. This is the orchestration layer the multi-agent ecosystem has been missing.
The Problem With Running Multiple Agents Today
The multi-agent era snuck up on teams. You added Claude Code for architecture, Codex for test generation, Mastra for workflow automation — each one excellent at its job, each one a completely different mental model to manage. The result is what developers call “the switchboard problem”: your job becomes context-switching between sessions rather than writing code. Throw in port collisions, Git conflicts from concurrent writes, and zero unified visibility across sessions, and the productivity gains start leaking out.
This is not a fringe problem. Gartner projects that more than 40% of enterprise AI initiatives will involve multi-agent coordination by 2028 — up from under 5% in 2024. The tooling to manage that shift barely exists yet. Builder.io’s analysis of agent orchestration puts it directly: the main failure mode is not logic bugs but missing infrastructure — no unified dashboards, no environment consistency, no automated verification before commits land.
What AgentProto Does
AgentProto sits above your agents. It does not replace Claude Code or Codex — it manages them. A single daemon starts, messages, watches, and stops any of its 14 supported agents with the same commands. Supported adapters include Claude Code, Codex, Gemini, Grok CLI, Hermes, OpenCode Go, OpenCode Zen, Mastra Code, Mastra Agent, jcode, pi, Antigravity, browser-as-agent, and OpenClaw. The runtime also exposes roughly 175 MCP tools through a built-in server, so anything that speaks Model Context Protocol can plug in immediately. Sessions persist with append-only conversation logs, meaning you can restart the daemon without losing history.
npm i -g @agentproto/cli
agentproto install claude-code
agentproto serve
agentproto sessions start claude-code --cwd . --prompt "refactor the payments module"
agentproto sessions --watchPolicy Gates: The Feature That Changes the Calculus
Every other multi-agent approach assumes a developer is watching. AgentProto assumes you are not. After kicking off an agent session, you attach a gate — either your existing test suite or a reviewer model — and that gate runs automatically every time the agent completes a turn. Nothing commits until the gate clears. If tests fail, the gate blocks and surfaces the failure with full context.
This is the real distinction from tools like Cline Desktop, Claude Code Projects, or GitHub HydraFusion. Those tools give you parallel execution or multi-model access. AgentProto gives you a verification checkpoint before any of that work lands in your repository. For teams running agents on production codebases overnight, that difference is not academic.
Model Routing: Declarative and Unit-Testable
The September 17 release introduced @agentproto/model-routing, an implementation of the AIP-57 MODEL-ROUTING spec. The core function resolves which provider and model serves a given request, given a ModelRef. It operates as a pure function — no I/O, no clock, no randomness — which means routing rules are unit-testable without mocking network calls. Resolution walks three layers in order: explicit override, environment variables, then pack defaults, and always reports which layer won. For multi-provider setups where OpenCode Go (36 models) and OpenCode Zen (102 models) might both be active, deterministic routing via FNV-1a hashing ensures the same conversation prefix always routes to the same model.
What Is Production-Ready and What Is Not
AgentProto is honest about its maturity in a way most early-stage projects are not. The project uses a tiered documentation system: Tier 1 features — the daemon, CLI, session management, and the 14 adapters — are labeled “live and verified.” Tier 2 components, covering the broader AIP specification scaffolding, are explicitly marked alpha. The project sits at v0.20.0 with single-digit GitHub stars as of late September 2026. This is early-adopter territory. If you are building internal tooling for a team that already runs multiple agents and wants policy-gated automation, it is worth a close look. If you need production-proven infrastructure for client work, the category leaders — LangChain, Mastra, CrewAI — still have more ecosystem depth.
The September 22 Release in Brief
The September 22 release addressed two practical reliability issues: MCP-Apps widgets now survive inside opaque-origin hosts like Claude Desktop (previously broken under restrictive Content Security Policy), and the transmit_message tool now correctly reports sent: false when delivery is blocked rather than silently claiming success. Five new style scorers landed in @agentproto/eval — two deterministic (no model required) and three model-backed — enabling automated writing-style evaluation in content pipelines.
The Bottom Line
If you are running more than one coding agent today — and most active AI developers are — you are doing orchestration manually. AgentProto is the first open-source attempt at a proper runtime layer for that problem: unified session management, policy-gated verification, 14 adapters, and model routing you can unit-test. It is not ready to replace your existing stack. It is worth experimenting with now if you have the tolerance for a v0.20.0 project. Check the AgentProto repository and the release notes — the pace of development suggests 1.0 is closer than the version number implies.
