Workflow engine for AI agents

Your agent’s “Do NOT” list is a list of scars.

Every rule in it exists because the agent invented a failure once. Realm replaces the rules with structure — wrong behaviour becomes impossible, not prohibited.

The agent calls Realm. Every other tool calls the agent.

Get startedView on GitHub

The inversion

Most platforms call the model. Realm gets called by it.

01Your agentCLAUDE · GPT · ANY MCPCLIENT02RealmSTATE MACHINEState guardValidates every submissionagainst the step’s schemaexecute_step()task + json schemavalid03State advancesEVIDENCE HASH RECORDEDinvalid · unapprovedRefusedSTATE UNCHANGED — RETRY

Structure, not instruction

A prompt rule is a request. A state machine is a constraint. Only one of them survives contact with a model having a bad day.

What that buys you

Order is enforced

A step cannot run before its dependencies. Not by instruction — by the state machine.

Output is schema-checked

Malformed output is rejected at the boundary. It never enters the run.

Humans hold the gates

Critical steps pause until a person approves. The agent cannot proceed past one.

Evidence, not logs

Every step leaves a tamper-evident record. For client work, that record is the deliverable.

See it run

Step through a real run, decision by decision

Every step records what it was asked, the schema it had to satisfy, what it submitted, and a hash of the result. Not a log that something ran — a record that it ran correctly.

Interactive replay

An executed workflow, stepped through in the browser: inputs, schemas, submissions, evidence hashes, and a human gate that refuses to advance.

Try it

Nothing to sign up for

# install the CLInpm install -g @sensigo/realm-cli# register a workflow and drive it autonomouslyrealm workflow register ./workflow.yaml realm agent --workflow my-workflow

Runs locally. Apache‑2.0, published on npm; evidence is written to your own machine.