Statewright Review (2026): Pricing, Features & Honest Verdict

TLDR

Statewright is a visual state machine builder that gives AI coding agents guardrails, restricting which tools they can use in each workflow phase. We found it genuinely improves reliability by stopping agents from running wild. Best for: developers running coding agents at scale. Price: Free, Pro $29/mo. Rating: 7.1/10.

What is Statewright?

Statewright is a tool that lets you design AI coding agent workflows visually and then enforce them at the protocol layer. Instead of handing an agent all 40-plus tools at once, you define states like planning, implementing, and testing, and you control exactly which tools unlock in each. The agent gets read-only access while it plans, edit tools while it implements, and test commands while it tests. Nothing more.

The enforcement is the part that matters. It runs on a deterministic Rust engine with no LLM in the loop, so the guardrails do not depend on a model deciding to behave. If the agent calls something outside the allowed set, the gateway rejects it with a concrete error, not a gentle suggestion it can ignore. It integrates with Claude Code, Codex, opencode, and Pi through hard enforcement via MCP and hooks, and with Cursor in advisory mode. It launched in 2025 and surfaced on Hacker News as a fix for the very real problem of agents going off the rails.

What Are Statewright’s Key Features?

Per-state tool enforcement

This is the core. Each state in your workflow has an allow-list of tools, and the agent simply cannot use anything outside it. The reasoning behind this is that constraining the tool space helps models reason better. By narrowing what is available, you cut the failure modes where an agent reads the same file ten times or deploys before it has tested anything. The narrower the options, the more focused the model stays.

Bash discernment and edit guards

Statewright inspects bash commands and blocks destructive ones like rm -rf or sed -i when write access is not permitted in the current state. Edit guards cap how large an edit can be and how many files an agent can touch per state. Together these stop the classic disaster where an agent decides to refactor your entire repo in one pass, or quietly nukes a directory it should not have touched.

Conditional transitions and approval gates

Transitions between states can be guarded by conditions, such as only proceeding if tests pass. You can also insert approval gates that pause for human review before a sensitive action runs. This turns a freewheeling agent loop into something closer to a reviewed pipeline, where the risky steps require a green light from you or your tests before they execute.

Per-state model routing

You can route cheap states to smaller models and reserve frontier models for the expensive reasoning steps, and adjust the thinking level per state on compatible clients. That means you are not paying top-model prices for trivial states. For teams running agents all day, routing planning to a smaller model and only escalating to a frontier model for the hard parts can cut real cost.

Visual workflow editor and run history

You design the state machine in a visual editor rather than hand-writing every transition, and the managed cloud keeps a run history so you can see how an agent moved through the states on past tasks. That history is useful for debugging: when an agent stalled or got blocked, you can replay the path it took and tighten the workflow. The workflows themselves are defined in JSON under the hood, so anything you build in the editor is portable and version-controllable rather than locked into a UI. Teams can share a workflow definition the same way they share any config file.

How Much Does Statewright Cost?

Statewright starts free. The free tier gives you 3 workflows and 200 transitions per month at $0, which is enough to evaluate whether the guardrails help your setup. Pro is $29/mo for 10 workflows and 2,500 transitions, and Team is $99/mo for 30 workflows and 10,000 transitions. Enterprise is unlimited and custom-priced through sales.

That pricing is reasonable for what it is. The free and Pro tiers cover most solo developers, and $29/mo is cheap next to the cost of a single botched agent run that deletes files or burns a stack of API tokens on a runaway loop. The transition-based metering is the thing to watch: a busy multi-state workflow consumes transitions quickly, so heavy users may climb to Team faster than the workflow count alone suggests.

There is also a self-hosting path. The engine and agent layer are Apache 2.0, and the MCP gateway is under FSL-1.1 that converts to Apache 2.0 in 2029, with single-developer and single-team self-hosting permitted. If you would rather run the whole stack yourself via Docker Compose, you can, which makes the effective price $0 for solo and single-team use. The self-host bundle includes PocketBase for storage, the MCP gateway, and the workflow editor, so you get the full managed experience without the subscription. The project also ships with a patent pledge covering independent implementations, which is a reassuring signal if you are wary of building on top of something that could later be locked down.

Who is Statewright Best For?

Use Statewright if you run AI coding agents like Claude Code, Codex, or opencode regularly and you have been bitten by them. If you have watched an agent deploy untested code, blow away files, or loop endlessly re-reading the same file, the per-phase enforcement is exactly the seatbelt you want. Its published research showed two local models jump from 2 out of 10 to 10 out of 10 passing attempts on a SWE-bench subset once constraints were enabled, which is a striking result for the same tasks on the same hardware.

Skip Statewright if you only use coding agents occasionally or for throwaway scripts. Defining JSON state machines with states, transitions, guards, and allow-lists is real upfront work, and for light use the overhead outweighs the benefit. Also note that Cursor only gets advisory enforcement, so if Cursor is your main client and you specifically want hard blocking, you will not get the same guarantees the other clients enjoy.

It is also a strong fit for teams working in regulated or sensitive codebases. When an agent literally cannot run a destructive command or deploy without passing a gate, you have an auditable, deterministic control rather than trusting the model to follow instructions. That kind of provable constraint is exactly what a security or compliance reviewer wants to see before letting an AI agent near production code, and it is hard to get from prompt-based guidance alone.

Best Statewright Alternatives

Raw MCP and hooks

You can build tool restrictions yourself using the Model Context Protocol and your client’s hook system, which is what Statewright wraps. This costs nothing but takes real engineering time, and you lose the visual editor, the run history, and the deterministic Rust gateway. Choose it if you want full control and have the time, or pick Statewright to skip the plumbing.

LangGraph

LangGraph gives you explicit state machines for agents in a code-first framework. It is broader than Statewright and aimed at building agents rather than guarding existing coding assistants. The open-source library is free and MIT licensed, with a usage-based hosted platform. Pick LangGraph if you are building your own agent from scratch; pick Statewright if you want to constrain an off-the-shelf assistant like Claude Code.

Cursor rules

If you live in Cursor, its built-in rules let you steer the agent with natural-language guidance, included free with Cursor. The catch is that rules are advisory, the agent can ignore them, whereas Statewright hard-blocks disallowed calls on its supported clients. Cursor rules are the lightweight option; Statewright is the enforcement option.

Final Verdict: Is Statewright Worth It?

Statewright earns a 7.1. It targets a narrow but genuine problem, coding agents that misbehave when handed unrestricted tool access, and it solves it with deterministic, per-phase enforcement rather than hope. The benchmark jump from 2 out of 10 to 10 out of 10 is the kind of concrete result that makes the case better than any marketing line, and the free tier plus open-source engine lower the risk of trying it.

It is not for everyone, and that is why it sits in the low sevens rather than higher. The setup overhead and the niche use case mean casual agent users will not get their money’s worth, and Cursor-first developers miss out on hard enforcement. But if you run coding agents daily and you are tired of cleaning up after them, Statewright is well worth the $29/mo, and it is worth a look on the free tier first.