With Keystroke, anyone can build specialized agents in minutes.

Agents are model-agnostic and extremely configurable. They come with native functionality for things like:

• Tools, MCP servers, and skills
• Triggers, schedules, and heartbeats
• A persistent file system
• A memory system
• Sandbox environments
• A SQLite database
• Web search, web fetch, and browser automation
• Workflows & subagents as tools
• Letting agents set their own triggers and schedules
• Per-employee deployments; deploy a unique instance for each user
• A chat interface in our web app
• Native messaging for Slack, Telegram, Discord, Linear, GitHub, and more
• Granular credential scopes
• And lots more

With Keystroke, anyone can build robust workflows in minutes.

Workflows are real TypeScript that runs on our managed infrastructure. They come with native functionality for things like:

• Webhook, cron, polling, and event-driven triggers
• A canvas to build & visualize complex, multi-agent workflows
• A growing catalog with hundreds of integrations (Slack, Gmail, Google Sheets, Notion, PostHog, Supabase, Stripe, GitHub, Linear, Notion, Attio, Perplexity, and more)
• Durable execution with retries, sleeps, resumes, and replay
• Type-safe data flow with Zod schemas in and out
• Agents as workflow steps
• Other workflows as workflow steps
• Local testing with the SDK before you deploy
• Git-backed deploys, versioning, and rollbacks
• Built-in run history and observability
• And lots more

Every agent and workflow gets its own page in our web app where users can collaborate, make edits, share it, chat with it, inspect previous versions, etc. Its code under the hood, but inherently collaborative in the app.

Users that are less familiar with coding agents can use Keystroke's in-app agent (a coding agent, built with Pi) to refactor a workflow, build an open-claw style agent, or wire up a custom integration.

In the web app, users can:

• Edit agents & workflows with an in-app coding agent
• Edit agents & workflows with the UI & canvas editor
• Share resources with a teammate, team, company, or via a public link
• Publish reusable agents, workflows, skills, and MCP servers to your organization's registry
• Roll out maintained versions across an organization
• Expose agents and workflows as MCP tools for Cursor, Claude, ChatGPT, and more

Keystroke gives teams the controls they need before agents touch credentials, sensitive data, production systems, and real budgets.

Permissions:

• Role-based access control — reusable roles for who can create, run, publish, deploy, or administer agents and workflows
• Scoped credentials and secrets — share reusable connections across teams without exposing the underlying API keys
• AI model restrictions — choose which providers and models each team can use, with spend policies and quota controls
• Deployment boundaries — run on Keystroke-managed infrastructure, or deploy inside your own cloud or VPC for regulated environments
• AI proxy and bring-your-own-key support — route model calls through org-owned keys, private proxies, and existing approval paths
• Sandboxed execution — every run executes in an isolated environment with predictable CPU, memory, filesystem, and egress boundaries

Observability & trust:

• Usage monitoring — track organization-wide runs, failures, credit usage, model spend, budgets, and quotas in one place
• Trace every MCP client and server — follow tool calls across integrations, MCP servers, custom operations, and agent tools through one logging layer
• Per-step lineage — inputs, outputs, logs, errors, retries, and downstream tool calls tied to the exact step that produced them
• Detailed audit logging — capture who changed what, which credential was used, where data moved, and which agent or workflow took the action
• Versioned deployments — every change rides on git-backed history with branch deploys, teammate drafts, and one-click rollback
• Native OpenTelemetry — export logs, traces, and spans to Grafana, Loki, Tempo, or your observability backbone

Right now, teams use tools like Zapier and n8n to build internal AI agents and automations.

Under the hood, these tools aren't really code. They're a translation layer — a visual canvas that compiles to a JSON config, which their runtime interprets to call APIs. That's three layers of abstraction between what you build and what actually runs.

This works for simple workflows, but it breaks down as systems get more complex or AI-heavy. You end up with JSON spaghetti and an abstraction that coding agents struggle to work with.

We felt this pain ourselves. Building internal agents and automations in these existing tools was painful. We kept asking: “under the hood, why isn’t this just code?”

That’s the core insight behind Keystroke. The builder is no longer just a human dragging boxes around a canvas. Increasingly, the builder is an agent — and agents are far better at writing, testing, deploying, and maintaining code than working with brittle JSON configs.

Keystroke is built for both sides of that shift: code-first primitives that make coding agents dramatically better at building AI systems, and a user-friendly platform on top so every role in an organization can create, use, and share AI agents and automations.

We're onboarding new users & teams every week. Request early access and we'll get back to you as soon as possible. Want to talk first? Email us.