business processes, captured and replayed
Explore the work in chat.
Capture it as a workflow
that runs itself.
The work that runs your business — pulling a customer's full history before a renewal call, triaging an alert, sending the weekly report, approving a refund — is repeatable. It just isn't repeatable yet.
toolscaled turns those processes into deterministic, repeatable workflows. You explore the work in a chat, using the tools your business already runs on. When the chat does the job correctly, you capture it. The workflow runs on a schedule, a webhook, or on demand — exactly the same way, every time. When it needs a human, it pauses and waits.
Works in the chat you already use
toolscaled is bidirectional MCP. Build, capture, run, and observe from the chat you're already in — or use ours. Same workflows, same triggers, same runs.
Faster than hiring a developer
The person who knows the work captures it directly. No translation step, no spec, no Jira ticket.
More flexible than no-code
You discover the right shape by running it in chat — not by drawing boxes upfront on a canvas you have to redraw every time the requirement changes.
More reliable than agents
Every run executes a verified plan, not an improvisation. The parts of agents that work — conversational discovery, judgment at the right moment — are kept. The part that doesn't is replaced with deterministic execution.
the loop
Build → Capture → Run → Observe
Every feature below sits in one of those four phases — plus the substrate that makes the loop production-grade.
Build — explore the work in chat
Conversational tool use
A chat session that talks to your real systems. Connect Salesforce, GitHub, Stripe, your warehouse, your internal APIs — through MCP, OpenAPI specs, or built-in connectors. The chat picks the right tools, executes them, and shows results inline.
Agents
Reusable chat configurations: a name, a system prompt, and a scoped set of tools. Stand up "the support agent" or "the renewals agent" once and use it the same way every time.
Artifacts
Files, charts, documents, and outputs the chat produces are first-class objects — stored, previewable, attachable to the next step, and carried into the captured workflow.
Tool groups + scoped credentials
Tools are organized into groups so you can give an agent or a workflow exactly the access it needs — nothing more. Credentials are stored once, attached at run time, never handed to the model in plaintext.
OpenAPI & MCP import
Drop in an OpenAPI spec or point us at an MCP server and your tools show up. No glue code. The catalog is editable per-org without engineering work.
Sessions and history
Every chat is persisted. Pick up where you left off. Branch from any point. Inspect what was produced, when, and from which tool call.
Capture — turn a working chat into a workflow
Chat → workflow export
When a chat does the job, capture it. The workflow is a deterministic plan: fixed sequence, fixed scope, fixed tool selection. LLM steps appear only where judgment is genuinely needed.
Auditable, editable definition
The export is a workflow you can read and edit, not opaque code. The dependency graph is auto-detected from what the chat actually called — so the workflow reflects the real work, not your after-the-fact description.
Visual dependency view
A read-only graph of the workflow's data flow, derived from the captured definition. Sanity-check what got captured before you wire it to a trigger.
Workflow library
Captured workflows live in your org's library — versioned, searchable, shareable. Publish once and the rest of the team can run it.
Run — fire the workflow without you
Triggers: cron, webhook, manual
Cron for scheduled work. Webhooks for event-driven work — GitHub, Stripe, Linear, your own systems, with a public endpoint per webhook. Manual for user-initiated runs.
Trigger router
One incoming event can fan out to multiple workflows. A single GitHub webhook routes push → deploy, PR → tests, P1 alert → page on-call — without duplicating URLs or cramming branching logic into one workflow.
Pre-staged context
When a workflow runs, it assembles every relevant piece of context — the event, enriched data, prior history, policy docs — into a single grounded picture for whoever consumes the output. The receiver opens the notification and has everything.
Human checkpoints
Pause-for-approval is a first-class step type. The workflow runs, hits the checkpoint, presents state to a human, waits for the decision, and resumes. Approvals are part of the workflow definition, not bolted on after the fact.
Idempotent, retried execution
Runs are tracked with replay keys so a duplicate fire doesn't double-spend. Transient failures retry with backoff. Built to run unattended.
Observe & fix — keep the workflow honest
Per-run observability
Every run, pass or fail, has a full trace: which workflow fired, which tools it called, what data flowed through, where it succeeded, where it failed. Drill into any node, see inputs and outputs.
Aggregate metrics
Trigger health, workflow success rates, per-step latency, tool-call usage. Across runs, across workflows, across the org. Answers "is this thing actually working?" without you having to babysit.
Edit in chat, the same way you authored
Observe-and-fix is symmetric with build. When a workflow needs a change, you do it in a chat session — the same way you built it the first time. The workflow re-exports as a new version. There's no second editing surface to learn.
Substrate — production-grade by default
None of this is differentiating. All of it is necessary. We ship parity here so the loop on top is the only thing you have to evaluate us on.
Bidirectional MCP
We consume tools from your systems as an MCP client, and expose workflows, triggers, sessions, and runs as an MCP server. The full loop is reachable from Claude, ChatGPT, Cursor, or any MCP-compatible client.
Multi-tenant by construction
Every resource — every chat, tool, workflow, run, credential — is scoped to your organization. Hard isolation at the database layer.
RBAC
Authoring, running, viewing, and approving are separate permissions. Configure who in your org can do which.
Audit logs
Every run, every tool call, every approval decision is logged with who, when, and what.
Scoped credentials
Credentials are bound to tools at authoring time and used at run time — never improvised by the model, never visible in workflow definitions.
Credit-based usage
Per-org credit balances, idempotent accounting, and visibility into where the spend goes.
what you get, every run
Four properties, on every fire of every workflow.
Complete
The workflow fetches every data point the author said matters.
Correct
It runs the plan it was authored to run, not whatever the model improvised this morning.
Consistent
Every fire produces the same shape of output.
Fast
One deterministic execution — not an agent feeling its way through a tool catalog at run time.
Get early access
toolscaled is in early access. If you have a process that's repeatable but not yet repeated, we want to talk.
hello@toolscaled.com