Record. Repeat. Diff. Optimize.

Record what AI does.
Repeat it for free.

Name: OSOP
Author: OSOP

OSOP captures any Claude Code session as .osop + .osoplog from real transcript evidence — then re-runs the same path with a fresh AI agent. Collect runs, diff them, optimize the workflow.

See osop log See osop replay

Open source · Apache 2.0 · 8 CLI commands · canonical osoplog v1.0 schema

Pillar 1 · Record

osop log

Capture any Claude Code session into .osop + .osoplog from the real transcript JSONL — every tool call, every timestamp, every output. No LLM self-report.

Optional SessionEnd hook auto-runs `osop log` after every session. Each capture is honest evidence: real timestamps, real tool counts, real failures. The first half of the loop.

See osop log →

Pillar 2 · Repeat

osop replay

Re-execute a captured .osop with a fresh AI agent via `claude -p`. Same prompts, same tool sequence, same path — producing a comparable .osoplog you can diff against the original.

Each replay = one more data point for `osop diff` and `osop optimize`. Run a workflow N times to see what's deterministic vs noisy, what got faster, what regressed. The second half of the loop.

See osop replay →

See it in action

OSOP uses its own format to explain itself. Click Visual to see the flow. Click .osop to see the YAML.

Stripe Payment Flow (API SOP)

stripe-payment-flow.osop.yaml

POST /v1/customersapi

Create a Stripe customer with email and metadata.

↓sequential→ POST /v1/payment_intents

POST /v1/payment_intentsapi

Create a payment intent. Amount, currency, customer ID.

↓sequential→ POST /v1/payment_intents/{id}/confirm

POST /v1/payment_intents/{id}/confirmapi

Confirm the payment with a payment method.

↓sequential→ GET /v1/payment_intents/{id}

↓fallback→ POST /v1/payment_intents

GET /v1/payment_intents/{id}api

Verify the payment succeeded.

↓conditional→ AI Generate Receipt

AI Generate Receiptagent

AI generates a personalized receipt email.

Open in Editor

The Record / Repeat Loop

the-loop.osop.yaml

osop logcli

Capture a real Claude Code session into .osop + .osoplog from the transcript JSONL.

↓sequential→ osop replay

osop replaysystem

Re-execute the captured .osop with a fresh AI agent via claude -p, producing more .osoplog files.

↓sequential→ osop diff

osop diffagent

Side-by-side compare two .osoplog runs. Per-step duration, cost, status changes.

↓sequential→ osop optimize

osop optimizecli

Feed multiple .osoplog files to AI. It writes a better .osop.

↓sequential→ Human Reviews

Human Reviewshuman

Accept, modify, or reject. The human decides what lands.

↓loop→ osop log

Open in Editor

Browse captured sessions →See osop replay →

Everything serves the two pillars

CLI, editor, SDKs, converters, MCP server, agent integrations. All open source.

SOP Doc

Visual Editor

Interactive node graph. Drag and drop .osop files.

113 Examples

API SOPs, DevOps, AI agents, business processes.

Format Converters

Import from GitHub Actions, Airflow, n8n.

VS Code Extension

Syntax highlighting, validation, preview.

The Loop

OSOP CLI (9 commands)

run, diff, synthesize, validate, render, test, report.

MCP Server (5 tools)

Any AI agent can execute and optimize workflows.

Agent Rules (18 platforms)

Drop-in .osoplog generation. Start The Loop.

OpenClaw Skill (5 skills)

Session logging, reports, review, optimization.

View all on GitHub →

“It started with wanting to know what the AI did. Then I realized every team has the same problem — processes are invisible. Every API sequence, every company workflow, every human-AI collaboration. If you can't see it, you can't improve it. OSOP makes it visible.”
— Why OSOP exists