openclaw packet factory

turn messy business inputs into small reviewable packets of work.

this repo is a packet-first starter kit for serious beginners and operators who want a real business win first, and for expert builders who want a clean schema-first pattern they can extend later.

verified against:

• openclaw docs for lobster, llm-task, cron, security, and setup
• openclaw package version 2026.3.28

---

what this repo does

ships three working packet lines:

1. transcript to follow-up packet
2. inbox batch to action packet
3. csv or crm export to exception packet

---

what a packet is

a packet is one small unit of work with:

• source
• intent
• key facts
• recommended next step
• draft output
• review status
• destination

a packet is not a side effect. it is the object you review before a side effect.

---

prerequisites

before running anything, verify all of these:

• openclaw is installed and working
• you have one working model route in openclaw
• python3 is on the gateway host
• lobster is installed on the same host and available on PATH
• openclaw.invoke is available on PATH

check the last two:

command -v lobster
command -v openclaw.invoke

---

first win path

1. run openclaw onboard
2. if your current build does not include onboarding yet, use openclaw setup as the fallback.
3. open one of the example user files under workspace/examples/ and copy the closest one into workspace/USER.md before you start editing.
4. edit workspace/USER.md with the real details about your business, your pain, and your authority boundaries. this shapes everything the agent produces.
5. copy the five workspace files into your active openclaw workspace:
   • workspace/AGENTS.md
   • workspace/SOUL.md
   • workspace/TOOLS.md
   • workspace/IDENTITY.md
   • workspace/USER.md
6. copy config/openclaw.json5.example to ~/.openclaw/openclaw.json and set your model/auth details
7. run one packet line only, starting with the transcript packet
8. review both the sample packets already in packets/examples/ and the packet your run generates
9. only add cron or live integrations after one line works cleanly

---

exact run commands

from the repo root:

bash run/transcript_packet.sh
bash run/inbox_packet.sh
bash run/csv_packet.sh

pass a custom input file when you are ready:

bash run/transcript_packet.sh /absolute/path/to/transcript.txt
bash run/inbox_packet.sh /absolute/path/to/inbox_batch.json
bash run/csv_packet.sh /absolute/path/to/rows.csv

---

how the approval gate works

every workflow pauses for review before saving the packet. when it pauses, lobster returns a token.

to approve and save the packet:

openclaw.invoke --tool lobster --action resume --args-json '{"token": "PASTE_TOKEN_HERE"}'

to reject and discard without saving:

openclaw.invoke --tool lobster --action reject --args-json '{"token": "PASTE_TOKEN_HERE"}'

replace PASTE_TOKEN_HERE with the token lobster printed when the workflow paused.

---

what success looks like

• the command returns a lobster json envelope
• the workflow pauses for approval at the review step
• after approval, a packet file is saved under packets/examples/*
• the saved packet matches the schema for that line
• you can read the packet in under two minutes and decide what to approve

---

thinking token cost

this repo sets "thinking": "low" in bin/run_llm_task.py for every llm-task call. on models that charge for thinking tokens, each run incurs thinking token cost. to disable it, open bin/run_llm_task.py and change "thinking": "low" to "thinking": "none". to increase reasoning depth for complex inputs, change it to "thinking": "high".

---

why this repo is not gmail-first

gmail pubsub is real and documented. it is also a sharper setup path. this repo starts with file and batch inputs so you can get one clean win first. gmail lives under advanced/gmail_README.md once the core packet shape works.

---

who this is for

• founders buried in meeting follow-up and admin leakage
• service operators drowning in transcripts, forms, and spreadsheets
• serious beginners who are willing to edit one workspace file and follow a real setup path
• expert builders who want a clean packet pattern instead of another giant agent demo

---

what stays manual at first

• sending external email
• writing directly into critical systems
• destructive actions
• anything legal, medical, or financial with real downside
• shared-team or client-facing operations on one sloppy gateway

---

known limits

• this repo assumes openclaw already has one working model route
• lobster must be installed on the same host as the gateway and available on PATH
• openclaw.invoke must be available on PATH (separate from the openclaw cli)
• llm-task must be enabled and allowlisted for the agent
• this repo is not a supported multi-tenant or mixed-trust team deployment model
• "thinking": "low" is set in bin/run_llm_task.py — see above for cost implications

---

repo map

file or folder	what it is
config/openclaw.json5.example	current-schema example config
workspace/	packet-first workspace files. edit workspace/USER.md first.
workspace/examples/	ready-made USER.md starting points for common operator types.
inputs/	sample inputs you can run immediately.
run/	one-command starter scripts.
lobster/	workflow files.
schemas/	json schemas for packet outputs.
prompts/	plain-language prompt files.
packets/examples/*	sample outputs plus generated packets from your own runs
bin/	helper scripts used inside lobster workflows.
docs/	setup notes, success checks, troubleshooting, known issues.
advanced/gmail_README.md	optional gmail path once core repo works
QUICKSTART.md	step-by-step setup guide.
run/README.md	how the run scripts work

---

if you want one sentence to remember the whole repo:

the packet is the unit of business leverage.