Skip to main content

OpenClaw with General Agents: 90-Minute Crash Course

6 Scenarios, Zero se Personal AI Employee tak

OpenClaw aap ka Personal AI Employee hai: aik open-source assistant jo aap ke apne laptop par chalta hai aur unhi messaging apps ke through reply karta hai jo aap already use karte hain, jaise WhatsApp, Telegram, Discord, Slack, iMessage, aur more.

Yeh woh project hai jis ne prove kiya ke AI Employees real hain, kaam karte hain, aur log unhein chahte hain. OpenClaw 2026 ka fastest-growing open-source project ban gaya; pehle months mein is ne hundreds of thousands of GitHub stars liye. Jensen Huang ne GTC 2026 par isay "the next ChatGPT" kaha; NVIDIA ne is ke upar NemoClaw build kiya.

In ninety minutes ke end tak aap ke paas aik AI Employee hoga: aap ke phone par messages ka jawab dene wala, tools aur external services use karne wala, aap ke mutabiq customize hone wala, apne schedule par chalne wala, aur phir bhi aap ke laptop par rehne wala. Yeh woh chatbot nahin jise aap visit karte hain; yeh woh worker hai jise aap delegate karte hain.

Yeh crash course kaise kaam karta hai. Aap aik tiny folder download karte hain, apne general agent (Claude Code ya OpenCode) ko dete hain, aur six scenarios se guzarte hain. Agent folder parhta hai, OpenClaw install aur run karta hai, aap ka phone connect karta hai, new skills pick karta hai, apna brain customize karta hai, aur aik task schedule karta hai jo aap ke baghair run hota hai. Aap steer karte hain; agent kaam karta hai; OpenClaw aap ka Personal AI Employee ban jata hai.

Reading path | prereqs | deep version (expand karne ke liye click karein)

Reading path (six scenarios + aik monthly habit):

  1. Install & chat local dashboard mein. ~15 min.
  2. Channel pair karein apne phone se (WhatsApp / Telegram / Discord). ~15 min.
  3. Real work delegate karein aur agent loop dekhein. ~10 min.
  4. Aap jaisa bolna & aap ko yaad rakhna + identity GitHub par back up karna. ~15 min.
  5. Extend karein aik skill + aik external tool ke saath. ~15 min.
  6. Isay khud se act karwayein aik cron job (ya heartbeat) ke saath jo aap ke liye run hoti hai. ~15 min.
  7. (Mahine mein aik dafa, aaj nahin) Audit run karein. Time aane par ~10 min.

Har scenario runnable success par khatam hota hai. Agar ninety minutes aik sitting mein zyada lag rahe hon to inhein alag sittings mein karein; state persist rehti hai. Aik optional appendix Google Workspace cover karta hai; voice, multi-agent safety, aur ACP-spawn dev finale ke liye Chapter 56 ki taraf point kiya gaya hai.

Prerequisites (teen cheezein; page inhein assume karta hai):

  1. Claude Code ya OpenCode installed ho. Dono mein se koi bhi chalega. Agar dono nahin hain to pehle Agentic Coding Crash Course karein.
  2. Aap Agentic Coding Crash Course kar chuke hon. Aap tool calls approve kar sakte hain, agent output parh sakte hain, aur pehchan sakte hain ke agent kahan stuck hai. Yahan hum un moves par lean karte hain; unhein dobara explain nahin karte.
  3. Node.js 22.16 ya later (Node 24 recommended). Terminal mein node --version chalayein. Agar v22.16 se neeche ho to nodejs.org/en/download se current release install karein (agar aap kahen to aap ka general agent aap ko step-by-step guide karega).

Patient version chahiye? Chapter 56: Meet Your Personal AI Employee isi material par seventeen lessons hai, plus voice, multi-agent, security, aur deployment. Agar yeh page kahin fast lage to matching Ch56 lesson par jump karein aur phir wapas aa jayein.

Collaboration Pattern​

Is page mein teen actors share karte hain. Diagram relationship ko concrete banata hai:

Teen actors is page ko share karte hain: aap, aap ka general agent, aur OpenClaw yani AI Employee. Aap prompts paste karte hain aur actions approve karte hain; aap ka general agent OpenClaw install aur configure karta hai; OpenClaw aap ke phone par reply karta hai aur scheduled tasks chalata hai.

Har scenario phir wahi five-step rhythm use karta hai:

  1. Aap apne general agent mein aik sentence paste karte hain. Yeh brief hai, script nahin. Aap batate hain ke kya chahiye; steps enumerate nahin karte.
  2. Aap ka agent AGENTS.md consult karta hai (yeh already context mein hota hai: folder mein CLAUDE.md session start par isay automatically import karta hai, isliye fetch step nahin) aur plan propose karta hai. Woh bataye ga ke kaun se commands chalane ka irada hai aur decision points flag karega (kaunsa channel, kaunsi skill, kya yaad rakhna hai). Pehle destructive command se pehle woh poochega.
  3. Aap approve karte hain aur dekhte hain. Agent install commands chalata hai, configuration set karta hai, background service restart karta hai, live log output dekhta hai, aur aap ko batata hai ke woh kya dekh raha hai. Jab known gotcha aaye to woh pattern pehchan kar documented fix apply karta hai.
  4. Aap ka agent seam par rukta hai. Kuch moves sirf aap kar sakte hain: aistudio.google.com par ja kar Gemini key lena, phone se QR scan karna, Google OAuth screens click karna, voice note play hota sunna. Agent seam name karta hai aur wait karta hai.
  5. Aap tab done hain jab aik observable cheez ho jaye. Dashboard mein real reply. Aap ke phone se message jaye aur reply wapas aaye. Disk par file appear ho. Har scenario batata hai ke kis cheez ka wait karna hai.

Har scenario wahi five-step rhythm use karta hai: aap aik sentence paste karte hain; agent plan propose karta hai; aap approve karte hain; agent execute karta hai; aap done-when verify karte hain. Agent har aise seam par rukta hai jise sirf aap cross kar sakte hain.

Bas itna. Agent woh kaam karta hai jo agent achhi tarah karta hai: install, configure, debug, restart, verify, recover. Aap woh karte hain jo sirf aap kar sakte hain: decide, approve, aur phone ya accounts se judi cheezon par act. Yeh rhythm, goal describe karna, plan lena, approve karna, aur har step par verification ke saath execute karna, wahi prompting pattern hai jo AI Prompting in 2026 crash course mein sikhaya gaya hai. Neeche har scenario do short paste prompts use karta hai, aik instructions ki wall nahin, taake aap rhythm ko feel karein.

Pore crash course ke liye aik recovery move

Agar kisi bhi point par kuch ulta ho jaye to CLI commands ya error codes yaad rakhne ki zaroorat nahin. Yeh apne agent ko paste karein:

Kuch kaam nahin kar raha. Gateway log parhein, seedhi zabaan mein batayein kya nazar aa raha hai, aur aisa fix propose karein jise main approve kar sakun.

Aap ka agent log parhta hai, jo dekhta hai usay name karta hai, aur fix propose karta hai. Aap approve karte hain. Isi crash course ke har scenario ka recovery loop yahi hai.

Agar scenario zyada lamba ho jaye

Har scenario ka budgeted time H2 mein diya gaya hai. Agar aap us budget se 2x aage nikal jayein, maslan 15-minute scenario par 30 minutes se zyada ho jayein, to agent ko wapas kheench kar paste karein: "What's blocking us, in one sentence? Let's re-plan from there." Budget se bahut aage spin karna aam tor par is baat ki nishani hai ke agent improvise kar raha hai; plan par dobara anchor karna isay fix karta hai.

Jo folder aap download karenge us mein exactly do files hain: AGENTS.md (OpenClaw work karne wale kisi bhi general agent ke liye ~600-line operational reference) aur CLAUDE.md (aik line: @AGENTS.md, jo Claude Code ko brief automatically import karne ko kehti hai). Yehi poora environment hai. Aik file plus aik one-line index, yahi woh poori "skill" hai jo aap apne agent ko dete hain.

Download openclaw-with-general-agents.zip

Isay kahin bhi unzip karein. Unzipped folder mein terminal open karein. Apna general agent launch karein:

cd openclaw-with-general-agents
claude

Aap ke agent ke paas ab brief loaded hai. Hum six scenarios aik aik kar ke karenge; har scenario next se pehle runnable success par khatam hota hai. Yeh brief capable general agent assume karta hai (Claude Code, ya OpenCode running Claude Sonnet/Opus, GPT-5, ya Gemini 2.5 Pro). Older ya smaller models longer scenarios mein drift karenge; agar Scenario 1 mein agent ka pehla plan aap ki machine ke mutabiq specific hone ke bajaye vague ya generic lage to aage badhne se pehle stronger model par switch karna signal hai.

Scenario 1 se Pehle: confirm karein ke agent ke paas brief loaded hai (~30 sec)​

Aik paste batata hai ke CLAUDE.md ne apna kaam kiya aur AGENTS.md aap ke agent ke context mein aa gaya:

Aap OpenClaw ke liye kya kar sakte hain?

Agar reply specific OpenClaw work name karta hai, jaise install probes, channels, brain files, skills, MCP servers, schedules, monthly audit, to aap loaded hain aur Scenario 1 ke liye ready hain. Agar reply generic AI capability talk jaisa lage jisme OpenClaw-specific details nahin, to import fire nahin hua: agent close karein, confirm karein ke aap unzipped openclaw-with-general-agents/ folder ke andar hain, aur relaunch karein.

AGENTS.md mein asal mein kya hai (jo file aap ka agent ab parh raha hai)

Aap ko yeh file khud kabhi parhne ki zaroorat nahin; point hi yahi hai. Lekin is ki shape janna behtar questions poochne mein madad karta hai ("walk me through the gotchas section" kaam karta hai kyun ke section exist karta hai). Brief order mein yeh cover karta hai:

PART 1 :: PRINCIPLES (apply everywhere)
  Versions checked against
  Source of truth, in order        -> live docs > this file > the gateway log
  Critical: discover before you act -> table of 17 doc-URL pointers
  Working pattern (every task)      -> read -> propose -> ask -> execute -> verify
  Safety rails (non-negotiable)
  Secrets discipline

PART 2 :: OPERATIONS (by task type)
  Install & onboard          -> the probe + onboard + paid-default gotcha
  Configure                  -> config CLI + human-path vs agent-path table
  Diagnose & recover         -> the 5 most common failures and their fixes
  Channels (WhatsApp / Telegram / Discord + the TTY constraint)
  Memory & brain             -> 3 layers, brain files, cross-channel proof
  Skills (via ClawHub) + Plugins + MCP servers
  The activation dance       -> exists -> disabled -> enabled -> configured
  Automation (heartbeats + cron + 3 hook flavors)
  Multi-agent + ACP + Safety & security
  When you don't know what to do  -> three-layer fallback
  Tone                       -> how to talk to you

Agar baad mein AGENTS.md ka koi section relevant lage to action se pehle agent se kah sakte hain ke us section ko walk through kare, maslan "walk me through the Channels section of AGENTS.md before we pair WhatsApp". Brief is tarah likha gaya hai ke agent us se self-direct kar sake.

Scenario 1: Employee install karein aur chatting shuru karein (~15 min)​

Goal: OpenClaw aap ke laptop par running ho, Gemini free tier set ho, aur dashboard mein "hi" kehne par reply wapas aaye. Teen short paste prompts: plan maangein, approve aur execute karein, phir verify karein.

1a. Install aur configure​

Pehla prompt: batayein ke kya chahiye aur plan maangein.

Main OpenClaw apne laptop par chalana chahta hun aur Gemini free tier ke through chat reply lena chahta hun. Kuch bhi touch karne se pehle mujhe seedhi zabaan mein plan samjha dein: pehle kya check karenge, kya change karenge, aur kahan mujhe step in karna hoga.

Aap ka agent AGENTS.md parhta hai, aap ki machine dekhta hai, aur plan propose karta hai. Woh do jagah flag karega jahan usay aap chahiye: aistudio.google.com/app/api-keys se free Gemini API key lena, aur system changes se pehle confirm karna. Plan parhein. Agar reasonable lage to next step. Agar kuch off lage to push back karein. Poochhein "why are you doing that?" aur agent explain ya adjust karega.

Doosra prompt: approve karein aur run karne dein.

Plan theek lag raha hai. Step by step aage barhein, aur har step par batayein kya nazar aa raha hai. Jab meri Gemini key chahiye ho to pause karein aur batayein main usay safely kaise dun.

Agent pause karega aur key maangega. aistudio.google.com/app/api-keys par ja kar key create karein (free, no credit card), aur agent jo safe-handling instruction de usay follow karein. Chat mein key paste karne ke bajaye terminal mein environment variable prefer karna chahiye.

1a done when: agent report kare ke OpenClaw installed hai, configured hai, aur Gemini key place par hai.

1b. End-to-end verify karein aur dashboard open karein​

Teesra prompt: pehle end-to-end verify, phir dashboard hand off.

Ab pehle apna end-to-end check khud karein (command line se gateway ke through quick "hi", jaise brief mein likha hai), phir mere liye dashboard open karein taake main browser se bhi try kar sakun.

Scenario 1 tab done hai jab: agent ka apna CLI check real reply ke saath wapas aaye, AUR browser dashboard bhi hi type karne ke baad reply kare. Dashboard footer mein active model google/gemini-2.5-flash dikhna chahiye. Agar kuch aur dikhaye, especially pro-preview model, to agent ko batayein; charge hone se pehle woh free tier par switch karega.

Under the hood, OpenClaw ab aap ke laptop par teen pieces ki tarah chal raha hai, sab ko aik background service coordinate kar rahi hai jo login par start hoti hai:

Architecture diagram: messages aap ke phone se Channel adapters ke through Gateway mein jate hain, yani port 18789 par long-running service jo sessions hold karti hai aur tool calls dispatch karti hai, phir Agent tak jate hain, jiske brain files aur state ~/.openclaw/workspace/ par hain. Gateway always-on substrate hai.

Aage scenarios mein aap har piece se milenge. Abhi ke liye: yeh installed hai, aur reply kar raha hai.

Scenario 2: Apne phone se channel pair karein (~15 min)​

Goal: apne phone se "hi" bhejein aur apne AI Employee se reply wapas lein.

Yeh apne agent ko paste karein:

Model dashboard mein jawab de raha hai. Ab main apne phone se AI Employee se baat karna chahta hun. WhatsApp pair karne ka walkthrough dein (preferred), aur agar meri jagah WhatsApp mein zyada friction ho to Telegram ya Discord par fall back karein. Start se pehle apna plan aur meri taraf se jo setup chahiye woh explain karein.

Aap ka agent bataye ga ke kaunsa path recommend kar raha hai aur kyun. WhatsApp ke liye woh personal account ke bajaye WhatsApp Business ke saath second number suggest karega (underlying library unofficial hai aur Meta personal accounts ban kar sakta hai). Telegram ke liye woh BotFather tak le jayega. Discord ke liye Developer Portal aur teen privacy intents on karne ka path bataye ga.

Aik cheez jo aap ka agent aap ke liye nahin kar sakta: login step QR code ya token prompt ke liye small terminal-based UI use karta hai, aur jab agent isay apne shell tool se run karta hai to woh UI properly render nahin hota. Isliye kisi point par agent pause karega aur kahega ke same folder mein fresh terminal window open kar ke login command khud chalayein. Phone se QR scan karein (WhatsApp Business -> Settings -> Linked Devices -> Link a Device) ya BotFather / Developer Portal se mila hua bot token paste karein. Jab done ho jaye to agent ko "linked" bata dein.

Scenario tab done hai jab: aap apne phone se bound number ko hi bhejein aur real reply wapas aaye.

Agar aap chahte hain ke AI Employee WhatsApp group chats mein bhi kaam kare, sirf one-on-one nahin, to agent se kahen:

AI Employee ko group chats ke liye bhi open karein. Mujhe walk through karein ke kya change hoga aur main isay test group mein kaise add karun.

Scenario 3 ke liye carry-forward

Aap ka phone ab aap ke laptop par chalne wali OpenClaw service ka authenticated path hai. Yeh pairing real trust hai jo phone ne grant ki hai. Isay credential ki tarah treat karein: pairing files share na karein, public repo mein commit na karein, aur agar laptop kho jaye to phone se device revoke karein (WhatsApp Business -> Linked Devices, ya Telegram/Discord ki equivalent setting).

Scenario 3: Real work delegate karein aur loop dekhein (~10 min)​

Concept. "AI Employee" ko chatbot se alag karne wali cheez agent loop hai: real task aata hai, agent decide karta hai ke kaun se tools chahiye (web fetch, calendar, file read, ya jo bhi), unhein call karta hai, result parhta hai, aur answer banata hai. Jab tak aap ne real task par loop run hota nahin dekha, "agent" marketing jaisa lagta hai. Aik dafa dekh lein to har reply ke peeche AI Employee actually kya kar raha hai, yeh name kar sakte hain.

Yeh apne agent ko paste karein:

Channel kaam kar raha hai. Ab prove karein ke yeh chatbot se zyada hai. Main phone se aisa task bhejna chahta hun jisme agent ko waqai kuch karna pade. Gateway log ka live view set karein taake main agent loop real time mein dekh sakun, phir batayein jab aap task bhejne ke liye ready hon.

Aap ka agent aik side terminal kholta hai (ya aap se khulwata hai) jo gateway log live stream karta hai. Jab ready ho, apne phone se koi real task bhejein jo aap waqai delegate karte. Tutorial demo nahin, apni real life se pick karein. Pehle task ke liye kuch useful shapes:

  • Research lookup: "What does <a competitor or vendor I care about> charge for their entry plan, and what's included? Give me a one-paragraph summary plus the source URL."
  • Web fetch and analyze: "Read this article URL I'll paste and tell me the three claims that most affect <my role or my industry>, with one sentence on whether each is well-supported."
  • Structured task: "Look at my last five outgoing emails in <a folder or label I name>; tell me which one most needs a follow-up and what the follow-up should say."

Point yeh hai: yeh us type ka task hai jo ChatGPT refuse karega ya weak karega. Isay agent ko real data fetch karna, reason karna, aur structured output banana padta hai. Aap ka AI Employee fetch karta hai, reason karta hai, aur answer deta hai.

Log stream mein roughly six lines scroll hoti dikhengi:

  1. Aap ke channel par inbound message arrive hota hai.
  2. Model call: agent loop message Gemini ko bhejta hai aur poochta hai kya karna hai.
  3. Tool call: agent task ke liye needed tool invoke karta hai (web fetch, file read, calendar lookup).
  4. Tool result: tool ne content chunk ke taur par kya return kiya.
  5. Second model call: loop result ko summary prompt ke saath Gemini ko wapas bhejta hai.
  6. Outbound message: reply aap ke channel ko wapas jata hai.

Scenario tab done hai jab: aap ne yeh six-line shape scroll hoti dekhi ho aur reply phone par aa jaye. Yahi loop hai. Baad ke scenarios mein jo bhi add hoga, new skill, external tool, scheduled task, woh isi loop ke andar zyada tools ya zyada triggers add karta hai.

Scenario 4: Isay aap jaisa bolna aur aap ko yaad rakhna sikhayein (~15 min)​

Aap ke AI Employee ka behavior us ke workspace mein markdown files ke set se aata hai: ~/.openclaw/workspace/. Fresh install un mein se kai files ship karta hai; yeh scenario day one par sab se zyada customize hone wali teen files touch karta hai (SOUL.md, IDENTITY.md, USER.md), phir aap se fourth file create karwata hai (MEMORY.md, jo agent ke pehli dafa likhne tak exist nahin karti). Baqi files (AGENTS.md, agent ke apne operating rules ke liye, jo zip wale companion AGENTS.md se alag hai; TOOLS.md, tool policy ke liye; HEARTBEAT.md, ambient routine ke liye) Ch56 Lesson 4: Customize Your Employee's Brain mein cover hoti hain.

Which File Do I Edit? ~/.openclaw/workspace/ ki seven workspace files ke liye cheat sheet. Top row: SOUL.md (voice), AGENTS.md (operations), IDENTITY.md (name), USER.md (context). Bottom row: TOOLS.md (capabilities), HEARTBEAT.md (routines), MEMORY.md (memory). Har card batata hai &quot;X change karna ho to kab edit karein&quot; aur &quot;X yahan na dalen&quot;. Sab files session start par system prompt mein inject hoti hain.

  • SOUL.md: personality aur tone (yeh kaise bolta hai)
  • IDENTITY.md: is ka apna name aur role (yeh khud ko kaise introduce karta hai)
  • USER.md: yeh aap ke baare mein kya janta hai (persistent context)
  • MEMORY.md: durable facts jo yeh channels ke across commit karta hai

Aap har file aik dafa touch karenge, har edit ke baad aik message bhejenge, aur farq feel karenge. Start se pehle do cheezein jaan lein: har file lean rakhein (har line context cost hai jo agent har single turn par pay karta hai, including har channel reply aur har scheduled job; isliye aik ya do pages kaafi hain), aur baad mein in files ko zyada churn na karein kyun ke yahi aap ke AI Employee ke har reply ko shape deti hain.

Sub-scenarios shuru hone se pehle, quick orientation ke liye apne general agent ko yeh paste karein:

Customize karne se pehle quick orientation: mera workspace ~/.openclaw/workspace/ par open karein aur SOUL.md, IDENTITY.md, aur USER.md mein abhi kya hai, har file ke liye one line mein batayein. Sirf defaults; agle step mein hum unhein change karenge, phir MEMORY.md saath banayenge.

Aap ko har file ka starting snapshot mil jaye ga. Upcoming edits abstract files par nahin, un specific files par changes lagenge jo aap dekh chuke hain.

Brain edits load karne ke liye /reset chahiye (aik dafa parhein, 4a-4d par apply hota hai)

Kisi bhi workspace file (SOUL.md, IDENTITY.md, USER.md, MEMORY.md) ko edit karne ke baad new content disk par hota hai, lekin running OpenClaw session abhi bhi system prompt ka cached snapshot use kar raha hota hai. Apne phone se (paired channel) /reset bhejein taake OpenClaw disk se system prompt rebuild kare. Agar aap ne Scenario 2 skip kiya hai aur paired channel nahin hai to dashboard chat http://127.0.0.1:18789 se /reset bhejein. Neeche har sub-scenario edit aur test message ke darmiyan is step ko assume karta hai.

4a. SOUL.md: is ki voice change karein​

Yeh apne general agent ko paste karein:

SOUL.md dekhein aur teen choti changes suggest karein jo replies ko zyada direct aur kam hedgy banayein (ya jo style mujhe miss ho raha hai). Pehle diff dikhayein; apply sirf meri approval ke baad karein.

Edit land hone ke baad phone se /reset bhejein, phir casual message bhejein jaise How are you today?

Done when: reply tone Scenario 1 ke bland "hi" reply se visibly different ho.

4b. IDENTITY.md: isay name dein​

Yeh apne general agent ko paste karein:

Isay name aur role dein. Main chahta hun yeh khud ko "Atlas, my research assistant" ke tor par introduce kare (ya aap jo name aur role behtar samjhein woh propose kar ke mujhe dikha dein). Pehle diff dikhayein.

Edit land hone ke baad /reset karein aur phone se poochhein Who are you?

Done when: yeh default ke bajaye naye name aur role ke saath introduce kare.

4c. USER.md: isay apne baare mein sikhayein​

Yeh apne general agent ko paste karein:

Isay mere baare mein sikhayein. Mera full name, role, timezone, aur woh teen topics add karein jin mein mujhe aksar help chahiye hoti hai. Jo cheez pehle se nahin jante woh mujh se poochein, aur apply karne se pehle diff dikhayein.

Jo missing hoga woh poochega. Edit land hone ke baad /reset karein aur poochhein What should I prioritize this afternoon, given what you know about me?

Done when: answer generic advice ke bajaye aap ke timezone aur top topics ko factor kare.

4d. MEMORY.md: channels ke across commit karein​

Pehli teen files voice shape karti hain. MEMORY.md different hai: yeh sirf agent ke main session mein load hoti hai, isliye jo bhi cheez aap chahte hain ke channels ke across maloom rahe usay deliberately commit karna padta hai. Neeche four-step ladder teen layers prove karti hai, session memory, channel cache, long-term commit, aik aik kar ke.

Neeche test fact temporary aur aap ke week ke liye specific hona chahiye, stable identity fact nahin. Aap ka name jaise stable facts 4c se USER.md mein already hain, isliye agar woh use karein to wall fire nahin hogi. Koi real in-flight cheez pick karein: "I'm trying to finish [a real project] by Friday" ya "I'm preparing a pitch for [a real client] on Wednesday" works.

Memory layers diagram: teen stacked horizontal layers. Session memory RAM mein rehti hai aur sirf reset tak survive karti hai. Channel memory per-channel disk par rehti hai aur gateway restarts survive karti hai. Long-term memory (MEMORY.md at ~/.openclaw/workspace/) woh single layer hai jise channels ke across load hone ke liye deliberate commit chahiye.

Four steps. (Aap sirf teen real messages bhejte hain; baqi short queries hain.)

  1. Apne paired channel se: Quick context: I'm trying to finish [your real in-flight thing] by Friday. Hold onto this. Phir foran: What am I trying to finish by Friday? Yeh answer karega (session + channel memory, dono automatic).
  2. Dashboard chat se (http://127.0.0.1:18789, aik different session): What am I trying to finish by Friday? Isay nahin pata hoga. Yahi wall hai: channel memory per-channel hai, shared nahin.
  3. Wapas apne paired channel mein: Commit my Friday goal to your long-term memory. Aap ka agent MEMORY.md create karta hai (yeh pehli commit tak exist nahin karti) aur confirm karta hai.
  4. Dashboard chat se dobara (nayi committed MEMORY.md load karne ke liye pehle /reset bhejein): What am I trying to finish by Friday? Ab isay pata hai. Deliberate commit ne wall cross kar li.

Full memory model ke liye, edge cases, /reset har layer ke saath kaise interact karta hai, gateway restarts ke dauran kya hota hai, Ch56 Lesson 5: Memory and Commands dekhein.

Voice and memory ladder tab done hai jab: Step 4 succeed ho. Aap ka AI Employee ab aap jaisa bolta hai, aap ke chahne ke mutabiq introduce hota hai, aap ke baare mein context janta hai, aur channels ke across yaad rakhta hai kyun ke kuch deliberately committed hai, sirf cached nahin. Scenario 4 fully done hone se pehle aik aur step (4e) hai.

4e. Jo identity abhi banayi hai us ka backup lein​

~/.openclaw/workspace/ wala workspace hi aap ka AI Employee hai: abhi customized brain files, plus baqi workspace markdown (operating rules, tool policy, heartbeat routine), aur jo bhi aap baad mein add karenge (Scenario 6 ke schedules, installed skills, etc.). Agar aaj raat laptop mar jaye to yeh sab kho jaye ga jab tak kahin aur saved na ho. Pore workspace ko dotfiles ki tarah treat karein.

Yeh apne general agent ko paste karein:

Mere agent ka workspace ~/.openclaw/workspace/ private GitHub repo mein back up karein taake laptop kharab ho jaye to yeh lost na ho. Sab workspace files include karein (SOUL/IDENTITY/USER/MEMORY brain files plus AGENTS.md, TOOLS.md, HEARTBEAT.md, aur future additions jaise schedule files), aur secrets aur session caches exclude karein. Jo Git tools mere paas already hain un ke hisaab se easiest tareeqe se set up karein, aur end par mujhe one-liner dein jise main safe jagah save kar sakun aur fresh laptop par OpenClaw install karne ke baad workspace dobara clone kar sakun.

Scenario 4 tab done hai jab: private repo GitHub par exist kare, aap ka workspace pushed ho (brain files plus baqi workspace markdown), aur aap ke paas recovery one-liner saved ho (note app ya password manager mein paste kar lein jahan baad mein mil sake). Aap ke AI Employee ki identity ab laptop wipe survive karti hai.

Scenario 5: Aik skill aur aik tool ke saath extend karein (~15 min)​

Concept. Aap ke AI Employee mein capabilities add karne ke do different ways hain, dono ki shape alag hai:

  • A skill aik folder hai jisme SKILL.md file hoti hai: expertise jo task match hone par agent auto-invoke karta hai. Skills cross-runtime spec (agentskills.io) follow karti hain, isliye wahi folder OpenClaw, Claude Code, OpenCode, aur 50+ others mein kaam karta hai. Do registries spec ke against distribute karti hain: skills.sh (broad, cross-runtime) aur ClawHub (OpenClaw-curated, zyada vetted).
  • A MCP tool aik capability hai jise agent call kar sakta hai: external service jo Model Context Protocol ke through functions expose karti hai (kisi bhi zone ka current time lena, database query karna, calendar invite bhejna, etc.). Configure, restart, verify; agent ko bina code ke new tools mil jate hain.

Skills know-how inject karti hain; tools reach add karte hain. Dono same shape follow karte hain: install (ya configure), gateway restart taake OpenClaw unhein pick up kare, loaded verify karein, phir phone se test karein.

Neeche har prompt agent ko Ch56 lesson URL plus aap ka USER.md deta hai. Lesson exact commands rakhta hai; aap natural language mein rehte hain jab agent read, plan, execute, aur verify karta hai.

5a. Aisi skill add karein jo aap ke real work se fit ho​

Heads up: installed skill ka fire na hona almost hamesha description mismatch hota hai. Install kaam kar gaya; aap ka message skill ki trigger description se match nahin hua. Yeh broken install nahin, description ke baare mein data hai: gateway log skill-load event dikhata hai jab woh fire hoti hai.

Pehla prompt: lesson parhe, discovery skill le, propose kare.

https://agentfactory.panaversity.org/docs/Building-OpenClaw-Apps/meet-your-personal-ai-employee/install-skills-discover-ecosystem parhein taake aap samajh lein OpenClaw skills kaise install karta hai (cross-runtime spec, scopes, gateway restart). Phir check karein ke find-skills skill pehle se installed hai ya nahin. Agar nahin hai to sirf wohi skill skills.sh se Global scope ke saath install karein (taake Claude Code aur OpenClaw dono ke skill directories mein land ho) aur gateway restart karein. Jab find-skills available ho jaye to isay skills.sh par mere USER.md ke against search ke liye use karein aur do ya teen real skills propose karein jo mere kaam se fit hoti hon. Har skill ke liye batayein us ki description kis par trigger hoti hai (sharp description sahi waqt par fire karti hai; vague description kabhi fire nahin karti), main kaise verify karun ke skill fire hui hai vanilla reply nahin, aur aap pehle kaunsi choose karte. Chosen skill abhi install na karein; pehle main pick karna chahta hun.

Aap ko aap ke actual work se grounded short list milegi, real install URLs ke saath. Aik pick karein.

Doosra prompt: dono runtimes mein install kare, phir verify kare.

[your pick] ko Global scope ke saath install karein taake woh Claude Code aur OpenClaw dono ke skills directories mein aik saath land ho, phir gateway restart karein. Mujhe batayein kis directories mein write hua taake main dekh sakun. SKILL.md description mujhe wapas list karein taake mujhe exactly pata ho paired channel se kya bhejna hai usay trigger karne ke liye, aur reply mein kya dekhna hai jo prove kare skill fire hui hai, vanilla model response nahin.

Apne paired channel se woh test input bhejein jo agent ne suggest kiya ho (meeting transcript, draft email, code snippet, ya skill ke hisaab se jo bhi).

5a tab done hai jab: agent confirm kare ke skill installed hai (aur location dikhaye), AUR test input skill ke specific format ya framing ke saath reply produce kare, generic answer nahin. Agar skill fire nahin hoti to aam tor par description mismatch hai (aap ka message skill description trigger nahin kar raha) ya missed restart; universal recovery prompt paste karein.

5b. Aik external tool connect karein (credentials nahin chahiye)​

Canonical hello-world MCP mcp-server-time hai: no API key, do tools (get_current_time, convert_time). Yeh standard proof hai ke "aap ne external tool connect kar liya." Heads up: MCP silently fail kar sakta hai. Misconfigured server chat mein error nahin deta; agent ko bas tool nahin milta. Gateway log hi diagnostic hai.

Pehla prompt: lesson parhe, configure kare, verify kare.

https://agentfactory.panaversity.org/docs/Building-OpenClaw-Apps/meet-your-personal-ai-employee/connect-external-tools parhein taake configure-then-restart shape aur Silent Failure pattern samajh aa jaye. Phir lesson se mcp-server-time example set up karein (API key ki zaroorat nahin). Pehle plan dikhayein, phir execute karein. Gateway restart ke baad prove karein ke time 2 tools ke saath registered hai. Agar missing ho ya 0 tools dikhaye to yeh Silent Failure hai: gateway log parhein, seedhi zabaan mein batayein kya nazar aa raha hai, aur fix propose karein.

Agent lesson follow karta hai, commands chalata hai, aur registration list dikhata hai. Jo line dekhni hai: time with 2 tools. Agar nahin ho to agent diagnose karta hai; aap fix approve karte hain.

Doosra prompt: phone se tool trigger karein, dashboard badge dekhein.

Time MCP connected hai. Main paired channel se real timezone question poochun ga. Gateway log live tail karein taake hum real time mein get_current_time invoked hota dekh saken, aur mujhe batayein dashboard http://127.0.0.1:18789 mein kya watch karna hai: tool badge dikhna chahiye jo prove kare ke agent ne training data se guess karne ke bajaye time MCP use kiya.

Apne phone se real time question poochhein jo aap ke liye matter karta ho. Examples:

  • "If I send this proposal to my client in <their city> right now, what's their local time? Is that a reasonable hour to email?"
  • "My team in <another timezone> ends their workday in how many hours? Should I wait until tomorrow morning my time?"
  • "What's the deadline in <the timezone the deadline is set in> if it's currently 3pm my time?"

5b tab done hai jab: agent aap ko time server registered with 2 tools dikhaye, AUR phone se real time question specific live time produce kare (generic timezone rule nahin), AUR dashboard reply par get_current_time tool badge dikhaye. Badge proof hai ke agent ne tool call kiya, hallucinate nahin kiya.

Scenario 5 tab done hai jab: 5a aur 5b dono done conditions hold karein.

Raaste mein aap ka agent activation dance explicitly name karega: har OpenClaw extension (skills, plugins, MCP servers, channels, hooks) wahi four steps se guzarti hai: exists -> disabled by default -> enabled -> configured (restart). Aik dafa pattern dekh lein to har new feature first try par broken lagne ke bajaye familiar lagta hai.

Activation dance diagram: four-step cycle (Exists, Disabled by default, Enabled, Configured) arrows ke saath order dikhata hai. Har OpenClaw extension yahi four steps follow karti hai. Jab new feature pehli try par broken lage to in four steps se walk through karein.

Scenario 6 ke liye carry-forward

Is scenario ki additions apne USER.md mein add karein taake scheduled jobs (coming next) jante hon ke yeh available hain. Yeh apne agent ko paste karein:

Jo skill aur MCP tool hum ne abhi set up kiye hain unhein mere USER.md mein add karein taake scheduled jobs chalne par jante hon kya available hai. Phir updated USER.md ko 4e wale backup repo mein commit aur push karein.

Aap ke AI Employee ki capabilities, sirf identity nahin, ab laptop wipe survive karti hain.

Scenario 6: Isay khud se act karne dein (~15 min)​

Concept. Ab tak aap ne AI Employee ko message kiya aur us ne reply diya. Schedules isay flip karte hain: agent clock ya interval par act karta hai, aap ke message kiye baghair. OpenClaw mein proactivity ki teen flavors hain:

  • Cron precise times ke liye ("every morning at 7am", "every Monday at 9am", "at end of day"). Aap sab se zyada yahi use karenge. Aap ki real life clock times ke around organized hai.
  • Heartbeat fixed cadence par ambient checks ke liye ("every 30 minutes scan for urgent unread", "every 4 hours look at calendar for prep notes"). Isay tab use karein jab trigger "periodically check" ho, "exactly X o'clock par yeh karo" nahin.
  • Hooks event triggers ke liye (webhook fire ho, session reset ho). Yahan out of scope; agar chahiye to Ch56 dekhein.

Is scenario ke do parts hain. Part 6a aik fast heartbeat demo hai jo prove karta hai ke proactive mechanism wired hai. Part 6b keeper hai: aik real schedule (usually cron job) jo kal waqai aap ke kaam aaye. 6a ke baad na rukain; demo jise aap disable kar dete hain proactive dimension nahin. Real schedule jo daily run hota hai, woh hai.

6a. Aik demo heartbeat fire hota dekhein (phir off karein)​

Yeh apne agent ko paste karein:

Low-cost task ke saath five-minute demo heartbeat schedule karein: har five minutes gateway log errors ke liye check kare aur one-line summary post kare. Jab main log mein ek fire hota dekh lun, sirf is demo ko disable kar dein taake mera Gemini quota burn na ho. Real schedule next add karenge.

Done when: log aik heartbeat-driven tool call dikhaye AUR demo disabled ho. Log dekhne ke liye five-minute window fair hai.

6b. Aik aisi cheez schedule karein jo aap waqai rakhenge (cron ya heartbeat)​

Demo jise aap disable kar dein yeh prove nahin karta ke aap ka AI Employee woh tool hai jo aap kal use karenge. Aik real schedule karta hai. Zyada tar first-time keepers ke liye cron right choice hai: aap ke real workdays clock times ke around organized hote hain, check-intervals ke around nahin.

Pehla prompt: aap ke baare mein known context par grounded options suggest karwayein.

Main ek real schedule add karna chahta hun jo waqai mere kaam aaye, koi aisa demo nahin jo main bhool jaun. USER.md se mere baare mein jo jante hain usay dekh kar do ya teen options suggest karein jo main shayad rakhun. Har option ke liye batayein woh kya karega, kab fire hoga, aur cron (precise time) ya heartbeat (ambient interval) mein se kaunsa primitive right hai. Main ek pick karun ga.

Aap ka agent USER.md par grounded options dega (7am summary, Monday morning priorities list, end-of-day outstanding commitments check, interval calendar scan, aur so on). Woh pick karein jo kal sab se useful lage.

Doosra prompt: set up kare aur backup kare.

[name your choice] ke saath chalte hain. Isay set up karein, confirm karein next kab fire hoga, aur schedule file ko 4e wale backup repo mein commit karein taake laptop wipe survive kare.

Done when: chosen schedule running ho, backup repo mein committed ho, aur agent ne bata diya ho ke next fire kab hoga. Isay on chhor dein. (Agar kal regret ho to sirf woh schedule disable kar sakte hain, baqi kuch touch kiye baghair.)

Scenario 7: Aap ka monthly AI Employee audit (~10 min/month)​

Concept. Aap ka AI Employee waqt ke saath accumulate karta hai: installed skills, captured credentials, connected MCP tools, memory entries, logs mein autonomous tool calls. Har addition aik chhota decision hai jo aap ne approve kiya; chain opaque tareeqe se compound hoti hai. Defense install time par vigilance nahin (jo cheez abhi exist hi nahin karti use aap catch nahin kar sakte); defense fixed cadence par ten-minute review hai. Yeh scenario aap ke pehle ninety minutes ka hissa nahin; yeh woh move hai jo aap apne AI Employee ki baqi life ke liye mahine mein aik dafa karte hain.

Time aane par yeh apne agent ko paste karein:

Mera OpenClaw monthly audit chalayein. Last audit ke baad jo kuch install, store, schedule, ya write hua hai usay walk through karein, aur koi bhi cheez flag karein jo main ne explicitly approve nahin ki, memory mein revealing lagti ho, ya approval setting apni zaroorat se zyada loose ho. Sab ka ek short report banayein jise main approve ya trim kar sakun.

Aap ka agent running inventory se guzarta hai (skills, memory entries, approvals, MCP tools, recent tool calls) plus stored credentials, phir aik short report likhta hai jo name karta hai ke last audit ke baad kya change hua aur kahan tighten ya trim karna chahiye.

Done when: aap ne report review karne mein ten minutes lagaye hon aur kam az kam aik decision liya ho (forgotten credential delete karna, over-broad approval revoke karna, stale memory entry prune karna, unused skill uninstall karna). Next month ke liye calendar mark karein.

Yeh kyun kaam karta hai​

Do cheezein fresh rehti hain; aik cheez durable rehti hai.

Fresh #1: Is page ke scenarios book site par live rehte hain. Agent har session current version fetch karta hai (aap batate hain ke aap kis scenario par hain, aur woh relevant section parhta hai).

Fresh #2: Current OpenClaw commands docs.openclaw.ai/llms.txt par live hain, full docs ka LLM-friendly index. Aap ka agent har dafa unhein fresh parhta hai jab woh kisi command ke baare mein unsure ho. OpenClaw fast ship karta hai; isi tarah brief accurate rehta hai jab individual flags drift karein.

Durable: AGENTS.md (aap ke two-file zip se operational reference) carry karta hai ke OpenClaw kya hai, docs navigate kaise karni hain, safety rails (pooche baghair sudo nahin, paid models nahin, keys ko ~/.openclaw/ ke bahar write nahin karna), recovery patterns, aur activation dance. Yeh full platform cover karta hai: install, debugging, channels, memory, skills, plugins, MCP, automation, multi-agent, ACP, aur sandboxing. Yeh page se lamba hai kyun ke yeh everything cover karta hai jo general agent ko OpenClaw ke saath karne ko kaha ja sakta hai, sirf upar ke six scenarios nahin. Folder mein kuch stale nahin hota, isliye aap isay aik dafa download kar ke reuse karte hain.

Intelligence files mein nahin; intelligence aap ke general agent ke unhein parhne aur aap ke next ask par apply karne mein hai. Aap ne six disconnected demos nahin kiye; aap ne woh tool assemble kiya hai jise aap kal touch karenge.

Ab asal mein kya running hai​

Six demos nahin: aik system. Scenario 6 ke baad jo persist karta hai us ki inventory:

ArtifactAsal mein kya haiKal yeh kyun matter karta hai
Background serviceOpenClaw, aap ke OS ke saath auto-startingAap ka AI Employee terminal close aur reboot survive karta hai
Channel pairingAap ke phone aur laptop ke darmiyan trusted linkWoh path jisse aap ka phone service tak pahunchta hai
Workspace files~/.openclaw/workspace/ mein seven markdown filesAap ke AI Employee ki identity, context, behavior, aur memory
GitHub backupWorkspace ki private repo plus recovery one-linerWorkspace laptop loss survive karta hai
One installed skillClawHub se aik expertise packAik real know-how extension jo aap ka agent auto-invoke karta hai
One external toolAik MCP server jise agent call kar sakta haiAgent ke liye aik real external service available hoti hai
One scheduled taskAik cron job ya heartbeat jo aap ke baghair fire hoti haiAik cheez jo schedule par aap ke liye chalti hai

Yahi picture hai. In mein se koi bhi woh demo nahin jo aap ne walk through kar ke disable kar diya; yeh sab us tool ke pieces hain jise aap kal touch karenge.

Is ke saath working day kuch aisa lagta hai: subah 7am par aap ka phone buzz karta hai whichever schedule aap ne choose kiya (agar 7am summary keeper hai to cron job); mid-morning aap quick question reply karte hain jo time MCP ya S5 ki skill trigger karta hai; mid-afternoon aap agent se teen emails ke replies draft karwate hain; day end par aik new fact long-term memory mein commit karte hain. Aap ne laptop khola bhi nahin.

Agar baad mein in artifacts mein se kuch missing ho jaye (laptop wipe, accidental delete, version upgrade gone wrong), 4e wali GitHub repo plus fresh OpenClaw install plus recovery one-liner aap ko isi exact picture par wapas le aata hai.

Public-facing channel connect karne se pehle

Crash course assume karta hai ke AI Employee sirf aap ke messages parhta hai. Agar aap kabhi public-facing channel connect karne ka plan karein (support inbox, contact form, kuch bhi jahan strangers likh sakte hon), yahan rukain aur pehle Chapter 56 Lesson 14: Gate Your Agent's Tools aur Lesson 16: Isolate with NemoClaw parhein. Sandboxed-reader pattern prompt injection ke khilaf aap ki structural defense hai, yani woh threat jahan email mein hidden adversarial instructions aap ke AI Employee ko aap ki taraf se actions lene par trick kar sakti hain. Pairing lock down karti hai ke kaun aap ke bot ko write kar sakta hai; sandboxing lock down karti hai ke aap ka bot jo parhta hai us ke saath kya kar sakta hai. Dono matter karte hain.

Agla kahan jayein​

Scenario 6 ke baad aap ke paas working AI Employee hai jiska workspace customized hai (voice, identity, aap ke baare mein context, committed memory), workspace GitHub par backed up hai, aik installed skill hai, aik external tool hai, aur aik scheduled task hai jo aap ke liye fire karta hai. Zyada tar logon ko isi surface ki zaroorat hoti hai.

Is page ne jis topic ko touch kiya (ya skip kiya), us ki patient walkthrough ke liye Chapter 56 mein full platform cover karne wale seventeen lessons hain. Quick map:

Aap chahte hain...Yahan jayein
Voice replies (WhatsApp / Telegram / Discord par audio)Ch56 L10: Give it a voice
Reader-agent pattern (untrusted-email safety, sandboxing)Ch56 L14: Gate Your Agent's Tools
Second specialized agent chalana (routing, separate identity)Ch56 L11: Add a second agent
AI Employee se general agents summon karwana (/acp spawn choreography finale, developers ke liye)Ch56 L13: Orchestrate other agents
Sandboxing modes aur security hardeningCh56 L14: Gate Your Agent's Tools, L16: Isolate with NemoClaw
More channels (Slack, Matrix, Signal, iMessage, Zalo)Apne general agent se poochhein: "Walk me through the <channel> setup using your brief."

Baqi har cheez ke liye aap ka AGENTS.md already platform ka zyada tar hissa cover karta hai. Apne general agent se poochhein: "What does AGENTS.md say about sandboxing?" Brief reference hai; page tour hai.

Meta-lesson: aap ke unzipped folder ki sab se valuable cheez AGENTS.md hai. Aik evening nikal kar isay end to end parhein (install steps ke liye nahin, document ki shape ke liye: discover-before-act table, human-path-vs-agent-path table, working pattern, gotcha catalog, activation dance). Phir agle tool ke liye bhi aisa aik brief likhein jise aap general agent ke saamne rakhne wale hain. Pattern portable hai: har tool jiska surface learnable hai us ke liye aik "little skill" likhne ke qabil hoti hai. OpenClaw early example tha kyun ke install agent-driven setup se actively benefit karta hai; aap ko aur examples milenge. Agla brief khud author karein.

Appendix: Google Workspace connect karein​

Frame upfront. Google Cloud Platform OAuth screens ke fifteen-plus minutes, real account par jise aap throwaway treat karein. Google consent flows time-bound hote hain (kuch links ten minutes mein expire ho jate hain) aur click-heavy hote hain. Yeh specifically Google integrate karne ki price hai; is ka OpenClaw se lena dena nahin, aur yeh kisi other integration ko easy nahin banaye ga.

Yeh apne agent ko paste karein:

Google Workspace (Gmail, Calendar, Drive) ko mere AI Employee se connect karein. Throwaway Google account use karein; GCP aur OAuth steps mein mujhe walk through karein, aur agar koi consent screen aise scopes maange jo aap ne pehle nahin bataye to explicit STOP conditions rakhein.

Aap ka agent live Workspace plugin docs fetch karta hai, plugin install karta hai (usually gog ya similar; assume karne se pehle verify karein), OAuth flow browser mein open karta hai, consent token env-var-backed reference ke through capture karta hai, aur small probe se verify karta hai (maslan "list my next three calendar events").

STOP conditions. Koi quota ya permission error jo aik fix attempt ke baad repeat ho. Koi indication ke aap se woh scopes grant karne ko kaha ja raha hai jinke baare mein agent ne nahin bataya. Koi sign ke GCP project khud misconfigured hai (yeh appendix clean throwaway account assume karta hai; existing GCP project's auth debug karna crash-course scope se bahar hai).

Pointer. Deep walkthrough Ch56 Lesson 12: Connect Google Workspace hai.

Flashcards Study Aid​