Is Keystroke worth it in 2026?

Verdict: 6.5/10 (Provisional, Open Alpha)

This is a first evaluation of a days-old product, scored on what is verifiable in July 2026 rather than on promise alone.

What Keystroke Gets Right

The thesis is sound. Coding agents genuinely work better with typed, testable code than with JSON workflow graphs: TypeScript with Zod schemas can be unit-tested with Vitest, diffed in pull requests, and versioned in git. Keystroke is the first funded product built around that observation end to end, from agent-native onboarding (pasting one docs URL into Claude Code or Cursor) to workflows living in the user's own repository.

The engineering foundations are credible. Durable step execution (successful steps recorded and replayed on retry), documented agent memory, VM-backed sandboxes, MCP support in both directions, and a transparent usage rate card ($0.01/run, $0.005/empty poll, 1.1x model markup with a free bring-your-own-keys path) all read like a team that has operated automation infrastructure before. The founders built internal automations on n8n, Gumloop, and Trigger.dev at their previous startup, Buster (YC W24), and raised $2.4 million.

Trialing is effectively free. The Hobby tier's included credit covers a few hundred runs a month, and open-alpha signups get $20 in credits.

What Holds the Score Down

It is days old. The public repository appeared on July 13, 2026; at review time it had roughly 40 GitHub stars, zero tagged releases, and npm at v0.1.98. There are no independent reviews, no community discussion threads, and no presence in any published n8n-alternatives round-up: no third party has validated reliability, support quality, or the Organization tier's RBAC, SSO, and audit-log claims.

Marketing runs ahead of documentation in places. The homepage advertises a database primitive for agents that the documentation does not yet cover, and the vendor's "open source" label understates the Elastic License 2.0's restrictions (source-available; hosted-service use prohibited).

Pricing may change. An open-alpha rate card is not a commitment; "Keystroke 1.0" is advertised but unshipped.

Editor's Note: We build client automations through coding agents daily, and Keystroke's pitch matches our experience: TypeScript an agent can test beats surgery on an exported JSON graph. But the repository was four days old when we reviewed it, and we do not put client production workloads on four-day-old platforms. It goes on the watch list, not the stack. — Rafal Fila, ShadowGen

Who Should Try It

Developer teams already working through Claude Code, Cursor, or Codex, automating internal or experimental workloads. Teams running production-critical business processes, or teams without developers, should stay on established platforms such as n8n and re-evaluate Keystroke after 1.0 ships and third-party operational evidence exists.