In this AgentProto review, here’s what we’re going to cover: what this brand-new open agentic runtime actually does, how it claims to let you run any coding agent in the background and check its work before it lands, what shipped in its 22 September 2026 release, and whether you should touch it this early.
Short answer:
- AgentProto is an open agentic runtime in TypeScript: one interface to start, message, watch and stop any coding agent, with quality gates before work lands.
- It integrates 14 agents — including Claude Code, Codex, Gemini, Hermes, opencode and Mastra — per its README.
- It’s free and Apache-2.0 licensed; you only pay for the underlying model usage.
- It is VERY early: single-digit GitHub stars, v0.20.0, and the project itself labels Tier 2 components as alpha scaffolding.
- Verdict: genuinely interesting architecture, but early-adopters-only for now.
AgentProto Review: What It Actually Is
AgentProto describes itself as an open agentic runtime that lets you “run any coding agent in the background, check its work before it lands.” I pulled the README and the latest release notes (tagged 22 September 2026) for this review, and the core idea is genuinely sharp: rather than being yet another agent, AgentProto sits above your agents. It unifies multiple agent CLIs under one interface with built-in quality gates.
Concretely, the README promises three things. First, unified control: “start, message, watch, and stop any agent with the same commands” — whether that agent is Claude Code, Codex, Gemini, Hermes, opencode or Mastra (14 integrations are listed). Second, work verification: you can attach automated checks — tests or reviewer models — that run before an agent’s commits land. Third, tool flexibility: it wraps the agent CLIs you already use rather than making you rewrite anything.
If you’ve ever had two or three coding agents running and lost track of which one was doing what, that first promise alone explains why this category exists. And the second promise — a quality gate between an agent and your repository — is the feature I think matters most for anyone running agents unattended.
Key Features Worth Rating
Beyond the headline pitch, the architecture splits into five layers per the README: the runtime and daemon, adapter families for the different agents, tool authoring, knowledge and workflow, and platform services covering auth, sandboxing, secrets and telemetry. Model access goes through gateway presets — Anthropic, OpenRouter and DeepSeek are named — so you can route different agents through different providers.
The 22 September 2026 release adds some telling detail. MCP-Apps widgets now survive daemon restarts using per-boot embed tokens (the release notes describe minting “a random 24-byte token” baked into the panel’s resource HTML, specifically to work inside restricted hosts like Claude Desktop). The transmit_message tool now honestly reports { sent: false } when a delivery is blocked, surfacing the provider’s blocked reason and suggestion instead of pretending it worked. And a new @agentproto/eval package ships deterministic and model-backed scorers — text statistics, lexicon matching, pairwise style comparison — plus diarised speech-to-text via an AssemblyAI integration.
That eval package is the sleeper feature for content people: scoring agent output for style automatically is exactly the sort of quality gate an AI content pipeline needs.
🔥 Want this set up without the guesswork? Working out whether a runtime like AgentProto belongs in your AI SEO stack — and what to run your agents on while it matures — is exactly what we cover. Inside the AI Profit Boardroom you get 3,700+ members, four live calls per week, daily tutorials, done-for-you templates and a 30-day roadmap, so you’re never stuck figuring this out alone. And if you’d rather talk it through first, book a free SEO strategy session and we’ll map out exactly where AI agents fit into your SEO workflow.
AgentProto Review: Setup and First Run
Setup, straight from the README, is a four-liner:
npm i -g @agentproto/cli
agentproto install claude-code
agentproto serve
agentproto sessions start claude-code --cwd . --prompt "refactor the payments module"So you install the CLI globally, tell it which agent to manage, run the daemon, and start background sessions against any working directory. The newest release is installed with npm install @agentproto/runtime@latest @agentproto/apps@latest @agentproto/eval@latest. Licence is Apache-2.0 — properly open source, free for commercial use.
⚠️ Caveat: the “Tier 1 — live, verified hands-on” and “Tier 2 — alpha scaffolding” status labels come from the project’s own README, i.e. maintainer-reported claims, not independent testing. This review is based on the official README and the 22 September 2026 release notes, not on a long production deployment — nobody has one yet, the project is days-fresh into public attention.
AgentProto vs Running Agent CLIs Directly
| Raw agent CLIs | AgentProto | |
|---|---|---|
| Running multiple agents | One terminal per agent, different commands each | Same start/message/watch/stop commands for all 14 supported agents |
| Checking work before it lands | Manual review, or trust and pray | Attach tests or reviewer models as automated gates |
| Cost | Free + model usage | Free (Apache-2.0) + model usage |
| Maturity | Battle-tested (Claude Code, Codex etc.) | v0.20.0, Tier 2 parts explicitly alpha |
| Community | Massive | Single-digit stars, 1,500+ commits, one lead maintainer |
Who Should (and Shouldn’t) Use It
Use it if you’re the kind of person who already runs Claude Code or Hermes on scheduled jobs, feels the pain of unchecked agent commits, and enjoys being early. The unified runtime plus eval-gates concept is where the whole agent space is heading, and getting fluent now is cheap — it’s free and wraps tools you already know. My Hermes Agent review and OpenAI Agents API review cover the two ecosystems it plays nicest with.
Skip it if you need boring reliability today. A project at v0.20.0 with self-declared alpha components and a tiny community is not where you run client deliverables. Bookmark it, star it, revisit at v1.0. If your agents live in one repo anyway, conventions like AGENTS.md support in Claude Code may get you 80% of the coordination benefit with none of the new moving parts.
The Bottom Line on This AgentProto Review
My honest verdict in this AgentProto review: the idea is ahead of the implementation’s maturity, and that’s fine — that’s what early looks like. An Apache-2.0 runtime that runs any coding agent in the background with quality gates before commits land is a genuinely good idea, the 22 September 2026 release shows real engineering care (honest failure reporting is a green flag), and the eval scorers hint at serious ambitions. But with single-digit stars and alpha-labelled components, it’s a tool for tinkerers this month, not for production. Watch this one.
Where to go from here: AgentProto is one to watch, but your agent stack needs to make money in the meantime. Join the AI Profit Boardroom for the live calls, templates and the 30-day roadmap, or book a free SEO strategy session and I’ll help you pick the agent setup that fits where your business is right now.
FAQ: agentproto review
What is AgentProto in one sentence?
AgentProto is an open agentic runtime — a TypeScript-based layer that lets you run any coding agent in the background, watch it, and check its work before it lands, all through one unified set of commands.
Is AgentProto free?
Yes. It’s fully open source under the Apache-2.0 licence. You pay nothing for AgentProto itself — only for whatever model provider your agents call underneath (via gateway presets such as Anthropic, OpenRouter or DeepSeek).
Which agents does AgentProto support?
The README lists 14 integrated agents, including Claude Code, Codex, Gemini, Hermes, opencode and Mastra. The pitch is that it works with the agent CLIs you already use, without rewrites.
How do I install AgentProto?
From the README: npm i -g @agentproto/cli, then agentproto install claude-code (or another agent), agentproto serve, and start a session with agentproto sessions start claude-code --cwd . --prompt "...".
Is AgentProto production-ready?
Partly, by its own admission. The project labels features in tiers: Tier 1 is described as live and verified hands-on, while Tier 2 components are explicitly alpha scaffolding. It’s also very early — single-digit GitHub stars as of late September 2026 — so treat it as an early adopter’s tool.
Who maintains AgentProto?
It’s maintained by Jeremy André and AgentProto contributors, with active development — over 1,500 commits on the main branch and a fresh release tagged 22 September 2026.
About the author
Julian Goldie is an SEO agency owner with 10+ years in SEO, 394K+ YouTube subscribers, a 100% Upwork job-success score, 75K+ community members across his groups, and the author of a best-selling SEO book. He tests AI tools and agents daily and publishes what actually works.
Watch the experiments on YouTube, learn AI SEO with 3,700+ members inside the AI Profit Boardroom, or book a free SEO strategy session to get a custom plan. For done-for-you work through the Goldie Agency, book a call for a custom quote.
Related reading
Last updated September 2026. This is the living guide to agentproto review — it gets updated as the tools change.