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

Keystroke is an AI agent and workflow platform where you write agents as TypeScript in your own repo and deploy them to a shared team workspace. It is the most credible n8n alternative we have seen for teams that already live in a code editor. Best for: engineering-led ops teams. Price: free forever on Hobby, $20/mo on Pro. Rating: 8/10.

What is Keystroke?

Keystroke is a platform for building internal AI agents and automations. You describe what you want, or you write it yourself in TypeScript, and Keystroke builds the agent, wires up the tools, tests it, and ships it into a workspace your whole team can use. Agents get memory, a persistent filesystem, code execution, web search, browser use, and access to more than 1,000 integrations, any REST API, or any MCP server.

The company is a Y Combinator W24 alum, and Keystroke is a pivot from their original idea, Buster. Blake Rouse is co-founder and CEO. The framework is source-available on GitHub under the Elastic License 2.0, and the launch on 5 August 2026 leaned hard on that. We think the pivot is the interesting part. Buster was a BI tool. What the team learned building agents to run their own company turned into the product, and it shows in the shape of the thing: this is not a drag-and-drop canvas with a code node bolted on. It is a TypeScript SDK with a managed runtime and a web app on top.

What are Keystroke’s key features?

Agents you keep in version control

This is the whole pitch. Your agents and workflows are TypeScript files in your own repository. You build them locally, review them in a pull request, test them, and deploy them like any other code. That means a coding agent like Claude Code or Cursor can write your automation for you, and the output is a diff you can read rather than 40 boxes on a canvas someone dragged around six months ago. Every automation platform eventually hits the wall where the visual builder cannot express what you need. Keystroke starts on the other side of that wall.

Durable workflows with replay and approval hooks

Workflows in Keystroke are durable. Steps replay after a failure instead of restarting the whole run, inputs and outputs are typed with Zod, and you can insert durable sleeps that pause a workflow for hours or days without burning compute. There are approval hooks too, so a workflow can stop and wait for a human to sign off before it does something expensive or irreversible. That last part matters more than it sounds. Most agent platforms will happily let an LLM send the email.

Triggers, channels, and instant apps

Agents run on schedules, webhooks, or events from connected apps. A Stripe charge lands, a pull request merges, Granola finishes a call note, and the agent starts. Once it has done the work you can reach it from Slack, Microsoft Teams, Telegram, Discord, Linear, WhatsApp, or an API call. Agents can also build lightweight dashboards and internal tools, so the output of a nightly sales agent can be a live page rather than a message nobody reads.

Bring your own model keys

Keystroke’s AI gateway gives you any supported model with a 1.1x markup on the provider’s published price. Or you skip the gateway, plug in your own API keys, and pay your provider directly at their rate. Having both is the right call. The 10 percent markup is honest and small enough that most teams will just pay it for the convenience, and the escape hatch is there when your model spend gets big enough to care.

Self-hosting and a standalone server

You can run the Keystroke server on your own infrastructure. That is not a footnote for anyone whose agents touch customer data, internal databases or anything under a compliance regime, because the usual answer from this category is a SOC 2 report and a shrug. It also means the platform is not a hostage situation: if the company’s pricing changes or the company disappears, your TypeScript still runs. The Elastic License 2.0 permits this. What it does not permit is you turning around and selling Keystroke-as-a-service to other people, which is a restriction almost nobody reading this will ever bump into.

How much does Keystroke cost?

Keystroke starts at $0. The Hobby plan is free forever and includes $1 per month of usage credit, which buys you 100 agent runs. You can build agents and workflows, deploy from your own repo, use triggers and schedules and webhooks, connect apps and MCP servers, bring your own model keys, and read logs and run history. The only real ceilings are concurrency and a 10 minute timeout on any single agent or workflow run.

Pro is $20 per month and comes with $20 of included usage credit, 50GB of file storage, higher concurrency, a 30 minute timeout, and priority early-access support. Organization pricing is custom and adds role-based access control, agent guardrails, advanced spend limits, unlimited concurrency, and enterprise security.

The meters underneath are published in full, which we appreciate. An agent run is $0.01. A workflow run is $0.01. An empty poll trigger is $0.005. Exa web search is $0.007 per search and web fetch is $0.001, both charged at cost. Sandbox compute is roughly $0.067 per hour, also at cost. Model tokens are the 1.1x gateway markup or your own keys.

Compare that to n8n, where the cloud Starter plan runs about 20 euros a month for 2,500 executions, or Zapier, whose Professional tier starts at $29.99 a month and climbs fast once you need multi-step zaps at volume. Keystroke’s $20 with $20 of credit included is competitive on paper. The honest caveat is that it is a usage model, and usage models are only cheap until an agent goes into a loop.

Who is Keystroke best for?

Use Keystroke if your team has engineers and you are tired of automation platforms that treat code as a second-class citizen. If your ops logic already wants to be a function, if you want your automations reviewed in pull requests, and if you would rather point Claude Code at a repo than click through a builder, this is built exactly for you. It is also a good fit if you want to self-host: you can run the standalone server on your own infrastructure.

Skip Keystroke if nobody on your team writes TypeScript. The web app is real and non-technical people can use the agents once they exist, but somebody has to build them, and that somebody needs to be comfortable in a repo. Skip it too if you need a mature platform today. Keystroke is in open alpha, and open alpha means what it usually means.

What is Keystroke actually like to build in?

Worth setting expectations, because “you can just describe your agent” and “it is TypeScript in your repo” are two different products and Keystroke claims both. The web app path works: you say what you want, it assembles the agent, connects the tools, tests it, and drops it into the workspace. That is the demo, and for a straightforward agent it holds up.

The repo path is where the platform earns its rating. You write the agent as a typed TypeScript file, define the tools it can reach, wire the trigger, and deploy. A coding agent can do the writing, which is the point: Keystroke is explicitly designed so that Claude Code or Cursor is the primary author and you are the reviewer. What you get back is a diff. You can see the prompt, the tool list, the retry behaviour and the guardrails as text, in a pull request, next to a comment from a colleague asking why the agent has write access to Stripe.

The two paths meet in the same workspace, so a non-technical colleague chats with an agent an engineer shipped from a repo. That is the correct division of labour and very few tools in this space get it right. n8n makes engineers work in a canvas. Most agent frameworks make everyone else work in a terminal. Keystroke is one of the few that does not force either group into the other’s tool.

Best Keystroke alternatives

n8n is the obvious comparison and the one Keystroke picks itself. It is fair-code licensed, self-hostable, has a far larger node library, and a visual canvas that non-developers can actually navigate. Cloud plans start around 20 euros a month. n8n is more mature and less opinionated. Keystroke is better if you want code first.

Zapier has thousands more integrations and a decade of polish, and it is the right answer for a marketing team that needs a Typeform to land in a spreadsheet. Free tier available, paid from $19.99 a month. It is the wrong answer for agents with memory and code execution.

Gumloop sits between the two: a visual builder aimed at AI workflows specifically, with a free tier and paid plans from $97 a month. Easier to start with than Keystroke, much harder to version control.

Final verdict: is Keystroke worth it?

Keystroke is the first agent platform we have looked at this year where the architecture choice is the feature. Putting agents in a repo as typed TypeScript solves the problem every automation tool eventually creates, which is a pile of undocumented business logic living inside somebody’s SaaS account. Durable replay, approval hooks, published per-run meters, and a bring-your-own-keys escape hatch are all signs of a team that has run this stuff in production and got burned in the usual places.

Two things hold it back. It is in open alpha, so expect breakage and expect the pricing meters to move. And the “open source” framing is doing some work: Elastic License 2.0 is source-available, not OSI open source, and it specifically forbids offering Keystroke as a hosted service. That is a reasonable licence for a company to pick and a fine reason to trust the code will stay readable, but it is not the same promise as MIT and the marketing does not make the distinction. If you have engineers and an appetite for alpha software, start on Hobby today, it costs nothing. If you need something stable this quarter, bookmark it and check back in six months.