# Keystroke

> A platform for building internal AI agents and systems using TypeScript, with 1,000+ integrations, durable workflows, and team collaboration tools.

Keystroke is a platform built by Sprint Labs, Inc. (backed by Y Combinator) that lets teams build, deploy, and operate internal AI agents and systems. Every agent, workflow, action, and trigger is defined as plain TypeScript — inspectable, testable, git-versioned, and portable — so teams avoid platform lock-in while still getting a fully managed cloud runtime.

## What It Is

Keystroke positions itself as an end-to-end platform for internal agent development: it handles models, memory, workspaces, tools, workflows, triggers, integrations, deployment, and run history so teams can go from idea to production agent without assembling the infrastructure themselves. It targets both technical builders (who work in Cursor, Claude Code, or Codex) and non-technical teammates (who build through the browser-based chat interface or Keystroke's own built-in agent).

## Core Primitives

The framework is organized around four building blocks:

- **Actions** — typed, reusable units of work defined with Zod schemas
- **Agents** — open-ended LLM-driven workers with memory, persistent file systems, web search, and tool access
- **Workflows** — deterministic, durable step sequences that can call agents when judgment is needed
- **Triggers** — schedules, webhooks, polls, and app events that start work automatically

Agents can use workflows as tools, and workflows can call agents as steps, enabling hybrid systems that mix deterministic and AI-driven logic.

## Build Paths and Deployment Model

Keystroke supports three build surfaces that all target the same platform:

- **Browser-based agent** — build without local setup using Keystroke's own chat interface
- **MCP client** — build from Claude, ChatGPT, or another MCP-compatible assistant
- **Coding agent / CLI** — full repository context, git, tests, and `keystroke deploy` for CI/CD-style workflows

The CLI (`@keystrokehq/cli`) handles lint, typecheck, build, and upload in a single command. Projects live under `src/` with subdirectories for actions, agents, triggers, and workflows. Node.js `^22.18.0 || >=24.0.0` is required (Node 23 is explicitly unsupported).

The repository also includes a source-available standalone runtime for teams that want to operate the runtime themselves, licensed under the Elastic License 2.0 — which permits inspection, modification, and self-hosting but prohibits offering the software as a managed service to third parties.

## Integrations and Model Access

Keystroke advertises 1,000+ native integration definitions covering tools like Slack, GitHub, Notion, Linear, Gmail, Stripe, HubSpot, Salesforce, Zendesk, Intercom, Figma, and more. Teams can also connect any HTTP API via OpenAPI spec, or attach any MCP server. Credentials are stored in a vault, scoped by organization, project, or user, and never exposed directly to agents.

For LLM access, Keystroke routes through an AI gateway supporting 300+ models (including models from Anthropic, OpenAI, and xAI) at a 1.1× markup on the provider's listed price, or teams can bring their own API keys at no markup.

## Collaboration and Governance

The platform includes an organization registry where agents, workflows, skills, and shared context can be published and reused across teams. A "Company Brain" feature aggregates internal data from connected sources so every agent works from shared organizational context. Governance features include role-based access controls, usage limits and guardrails, model controls, tool permissions, safety policies (with code-defined approval rules), and full audit trails with per-run traces showing every tool call, agent handoff, and human decision.

## Update: Active Development (2026)

The GitHub repository was created in July 2026 and last pushed in August 2026, indicating active early-stage development. The project has 56 stars and 11 forks as of the last crawl. Keystroke is listed as backed by Y Combinator. The README references models like GPT-5.6 Luna, Opus 5, and Grok 4.5, reflecting a product direction aligned with frontier model releases. The standalone runtime source is available for inspection under the Elastic License 2.0, while Keystroke Cloud contains separately licensed enterprise code.

## Features
- Build AI agents with memory, persistent file system, and code execution
- Durable workflow orchestration with typed inputs/outputs (Zod)
- 1,000+ native integrations plus any HTTP API or MCP server
- 300+ LLM models or bring your own API keys
- Event triggers: schedules, webhooks, polls, and app events
- Human-in-the-loop approvals and durable sleeps
- Organization registry for sharing agents, skills, and context
- Company Brain for shared organizational context
- Role-based access controls and credential vault
- Usage limits, guardrails, and audit trails
- Per-run traces with full tool call and agent handoff history
- CLI for deploy, prompt, test, and credential management
- Build via browser, MCP client, or coding agent (Cursor, Claude Code, Codex)
- Persistent browser for agents that stay signed in across runs
- Web search and web fetch via Exa
- Self-scheduling triggers set by agents
- Instant app generation (dashboards, internal tools)
- Test suites for agents and workflows before deployment
- Source-available standalone runtime (Elastic License 2.0)

## Integrations
Slack, Microsoft Teams, Telegram, Discord, WhatsApp, Gmail, GitHub, Notion, Linear, Stripe, HubSpot, Salesforce, Zendesk, Intercom, Figma, PostHog, Sentry, Apollo, Perplexity, Exa, Superhuman, Google Calendar, Granola, Attio, Cursor, Claude Code, Codex, ChatGPT (MCP), Any HTTP API via OpenAPI spec, Any MCP server

## Platforms
WEB, CLI, API

## Pricing
Freemium — Free tier available with paid upgrades

## Links
- Website: https://keystroke.ai
- Documentation: https://keystroke.ai/docs
- Repository: https://github.com/keystrokehq/keystroke
- EveryDev.ai: https://www.everydev.ai/tools/keystroke