After publishing the previous OpenClaw hands-on review, the most common questions from comments and group chats fell into five buckets: token costs, what it can actually do, local models, privacy and security, and day-to-day experience. Today I am picking the five most frequently asked questions and answering each one in detail.
Historical and review note (October 2026): Personal experience and usage figures describe March 2026; the original includes no itemized bills or execution logs. The old price tables are obsolete and are not purchasing guidance. Billing boundaries below use current official documentation, while OpenClaw configuration is checked against v2026.10.3. Illustrations, including the token trend and status card, are conceptual rather than measured dashboard screenshots.
1. How Much Do Tokens Actually Cost? Is It Expensive?
Short answer: not as expensive as you think, but not free either. I have run two different setups—one based on a premium cloud provider, and one on a cost-effective domestic (Chinese) model plan. The difference in experience is noticeable.
Phase 1: ChatGPT Pro (Historical Choice)
I initially used the ChatGPT Pro plan described at the time as $200/month, sharing some usage with everyday coding. Not hitting a limit was my historical experience, not a guarantee of unlimited work or an OpenAI API balance included in that price.
OpenAI’s pricing documentation separates subscription usage from API token pricing. Sign in with ChatGPT also distinguishes sign-in from permission to use a plan. OpenClaw integrations must distinguish subscription sign-in and API-key routes. The original does not record complete authentication/billing details, so it cannot establish that arbitrary API calls were included or that its historical five-hour/weekly descriptions remain current.
Phase 2: Zhipu GLM Coding Plan (Historical Choice)
After Pro expired, I switched to Zhipu Coding Plan; the original records a promotional purchase at ¥980/year. That personal purchase history is retained. The former ¥1608/year figure, “roughly 30%” increase, and quarterly-price-times-four calculation are different pricing bases and cannot verify each other without matching orders or announcements. Historical notes and current plan rules are separated below.
March 2026 Plan Notes (Obsolete)
The old price table is obsolete and was not a verified annual checkout quote. Multiplying a quarterly price by four gives an arithmetic total for four quarters, not the provider’s annual, renewal, or promotional price. The old quotes and recommendation labels have been removed.
At the October 2026 review, Zhipu’s official overview lists GLM-5.3/GLM-5.3-Flash with weighted consumption for input, cached input, output, and tools. It supports OpenClaw with lower-priority, best-effort scheduling that can queue or rate-limit requests under load.
| Tier | Listed five-hour points | Weekly points |
|---|---|---|
| Lite | 2,000 | 10,000 |
| Pro | 12,000 | 60,000 |
| Max | 28,000 | 140,000 |
These are rules observed at review time, not token bundles or unlimited API access. Check the actual order for price, region, and renewal terms; the plan applies only through supported tools and the appropriate integration route.
GLM-5 API Pay-As-You-Go (If You Skip the Plan)
The original GLM-5/GLM-5-Code rate table has no reproducible historical quote and is not current pricing. Calculate pay-as-you-go cost using the actual model, endpoint, and billing rates, separating at least:
| Component | Required data |
|---|---|
| Uncached input | Uncached input tokens × applicable rate |
| Cached input | Cache-hit tokens × cache rate |
| Output | Billed output tokens × output rate |
| Other usage | Applicable tool, search, audio, or other charges |
For rates per million tokens, divide each corresponding count by one million before multiplying. Do not count cache hits again as uncached input. A total of 240 million tokens without the input/output/cache split does not establish a ¥2400 bill. Subscription limits and API charges are not directly interchangeable allowances.
Domestic Plan Comparison (Reviewed October 2026)
The old prices, model list, and support ratings are obsolete. No common benchmark supported claims of “best” integration or equivalent performance. By the October 2026 review, the offerings had changed:
| Provider | Change reflected in official documentation | Check before integration |
|---|---|---|
| Zhipu | GLM-5.3 family, points, lower-priority OpenClaw scheduling | Supported tools/endpoints and five-hour/weekly points |
| MiniMax | The old entry now describes M Plan Go/Explore/Build and supports OpenClaw | Purchased benefits, models, and M Plan key; old Starter/Plus/Max prices do not apply automatically |
| Kimi | Old/new memberships coexist; new plans remove the weekly window but retain five-hour limits and a monthly allocation | Your account’s plan; Kimi Code and the public API have distinct billing routes |
Do not choose an annual plan from the obsolete price difference. Confirm that the workload is supported, try representative tasks, then compare completion, latency, limits, and actual cost.
Real Numbers: 240 Million Tokens in 7 Days
I checked the Zhipu dashboard at the time and saw roughly 240 million tokens in seven days. That number looked alarming: about 34.29 million per day. I did not include itemized billing, input/cache/output categories, or matching local counters in the article, so it cannot establish an API invoice, local-token share, or annual-plan savings.
The image shows conceptual Q1–Q4 trends, not that seven-day usage report. Cloud/local division remains my historical workflow choice; comparing cost requires cloud charges, local running costs, and task completion rates.
Mindset: Don’t Be Too Transactional
While we are on the topic of costs, let me share a bit about mindset. After running OpenClaw for this long—my kid calls it “Shrimp Bro” (虾哥)—my biggest takeaway is: don’t start out obsessing over monetization and ROI. The AI era is just getting started, and upfront investment is necessary. The time and money you put in now may not yield immediate returns, but they gradually build into something invaluable: an intuitive understanding of what AI can and cannot do, a feel for how agent collaboration works, and an instinct for which problems are solvable with AI. These things are hard to quantify in the short term, but their long-term value is enormous.
My approach has always been: think of running OpenClaw as buying yourself a great toy. The joy is worth the price. Watching it run tasks on its own, draft articles, and organize the knowledge base—that alone is deeply satisfying. If it pays for itself down the road, great. But whether it can pay for itself right now should never be the deciding factor for getting started. Just start building. That matters more than anything else.
2. What Can OpenClaw Actually Do? Give Me Real Examples
The previous hands-on review went into detail, but I know many people do not read start to finish. Here are the scenarios I find most valuable and most illustrative.
What I Am Currently Running
| Scenario | Description | Value |
|---|---|---|
| Daily Paper Digest | Auto-fetches Android/AI papers, translates, generates reading notes, saves to Obsidian | Reduce collection work; still verify sources and conclusions |
| GitHub Repo Monitoring | Watches Issues/PRs for changes, proactively notifies on important updates | I monitor Perfetto—no more daily manual checking |
| Knowledge Base Building | Auto-archives content from WeChat Official Accounts, X, blogs; structures it | 1,760 Markdown files and growing |
| Content Review | Daily push of content worth re-reading | Solves the “bookmarked and never read again” problem |
| Android News Digest | Auto-summarizes domain news, extracts key topics | No need to scroll through feeds |
What I value most is turning the things you know are important but cannot sustain through willpower alone into background tasks that run automatically by default.
Advanced Case 1: Batch Import into NotebookLM for Deep Analysis
NotebookLM can answer questions and generate audio overviews from imported material; a two-host podcast is one presentation format. It already supports adding multiple sources: official help describes multiple web URLs and selecting several Drive files, rather than requiring every source to be added individually.
OpenClaw can collect, organize, and connect these steps when the required tools, login, and permissions are available. Google’s Enterprise source API provides sources.batchCreate within its Cloud product, project, and authorization scope. It is not a universal API automatically available to every personal NotebookLM account. Browser automation is a separate integration that also needs maintenance and verification.
In Practice
This is an example task request, not a built-in one-click command:
1 | Me: Research Android 15 performance optimization features. |
Verify collected sources, accessible content, successful/failed imports, and audio-generation status separately. Model filtering does not guarantee quality, and submitting generation does not mean playable audio is ready. Preserve missing sources, authentication failures, and quota errors in the result.
Why This Combination Works So Well
| Stage | Work automation can reduce | Still verify |
|---|---|---|
| Collection | Topic searches, deduplication, source lists | Accessible originals and source reliability |
| Import | Batch submission through a supported API/UI | Success for each source |
| Updates | Periodic discovery and summaries | Meaningful changes and obsolete conclusions |
The value is less repetitive work, not treating automated filtering as proof of accuracy.
Advanced Case 2: Drop a Link, Auto-Fetch All Related Content and Save to Disk
This is one of my most-used scenarios—it triggers almost daily. While scrolling X or reading a WeChat Official Account, I come across a great article that references several papers, a few GitHub repos, some related links, and a couple of book recommendations. The old workflow is all too familiar: bookmark it, tell yourself you will read it later, forget, never look at it again. With OpenClaw, the process is different:
1 | Me: https://example.com/awesome-article |
Handle each item according to access and authorization: download publicly available papers, extract readable content where saving is permitted, and record book metadata with purchase/borrowing links. Login requirements, paywalls, and dead links can prevent full retrieval; do not report that everything was captured.
This example explicitly requests stars and clones; execution still needs bounded repositories and counts, and cloning is not permission to run repository code. The final index should record successes, failures, original URLs, and save locations. Personal-Knowledge/source/ is a personal directory convention, not an OpenClaw default.
For Android Developers
These are directions I have not built yet but am actively planning. See if any resonate with your work:
| Pain Point | What OpenClaw Can Do | Audience |
|---|---|---|
| High barrier to trace analysis | Upload Perfetto/Systrace traces, AI-assisted interpretation, generate plain-language reports | App perf optimization, system UX perf |
| Version compatibility checklists | Auto-fetch Android Release Notes, generate adaptation checklists, cross-reference code to flag check points | App developers (annual necessity) |
| Issue backlog overload | Auto-classify (bug/feature/question), priority scoring, key info extraction | Open source maintainers, team collaboration |
| Time-consuming code review | Detect common patterns (null pointers, resource leaks, main-thread IO), generate review checklists | All teams |
| Forgotten tech debt | Periodically scan TODO/FIXME comments, sort by priority, send reminders to “pay down the debt” | All projects |
| Benchmarks nobody reads | Run benchmarks on schedule, auto-compare against history, alert on anomalies (e.g., startup +500ms) | Performance teams |
| Internal docs are unfindable | Answer questions like “have we seen this before?” or “where is the design doc?” based on the knowledge base | All teams |
| Upstream change tracking | Monitor AOSP/vendor code changes, flag items that need merging | System developers |
None of these work out of the box—they require custom development for your specific scenarios. But OpenClaw’s architecture naturally supports extending in all of these directions. The key is to get the foundation running first, then iterate based on real needs.
One-Line Summary
Papers read daily, repos monitored daily, knowledge organized daily, reports written daily. What OpenClaw truly achieves is turning “things that require long-term consistency”—something that goes against human nature—into “tasks the system quietly handles in the background.”
3. What Can Local Models Do?
In my workflow at the time, local models mainly handled repetitive, high-frequency tasks. One line to describe the division of labor: the cloud “thinks deep,” local models “run hard.”
How Much Do Local Models Handle?
Here is the task distribution I recorded at the time; I did not include the time window, counting method, or raw logs:
| Task Type | Model | Share |
|---|---|---|
| Monitoring, inspection, batch processing | Local Qwen3.5-27B | 30% |
| Status checks, lightweight classification | Local Qwen3.5-4B | 20% |
| Ultra-light tasks | Local Qwen3.5-2B | 10% |
| Cloud failure fallback | Local Qwen3.5-9B | 5% |
| Long-form writing, polish, reviews | Cloud GLM5 | 15% |
| Daily/weekly reports, auditing | Cloud GLM5 | 10% |
| Paper digest, ClawFeed | Cloud GLM5 | 10% |
The listed shares add to 65% local tasks and 35% cloud tasks; this is table arithmetic, not a new runtime validation. Task share does not establish token share. The former “80%+ local tokens” claim lacks matching local counts and definitions and is not retained as a verified result.
My Setup
The hardware is a Mac Studio (M-series Apple Silicon) running multiple local models: Qwen3.5-2B/4B/9B using MLX 8-bit quantization for lightweight tasks and cloud-failure fallback; Qwen3.5-27B deployed via Ollama as the primary local workhorse for high-frequency structured tasks. MLX is a computation framework; Ollama is a model runtime/serving tool. Names alone do not establish a speed comparison. No matched model, quantization, context, and concurrency benchmark is provided here, so this does not prove MLX is always faster.
Division of Labor Principles
| Local Models Excel At | Cloud Models Excel At |
|---|---|
| High-frequency structured tasks (RSS, queue processing) | Long-form writing and polishing |
| Monitoring and inspection | Complex reasoning and judgment |
| Data preprocessing (configured locally; verify outbound paths) | High-quality translation and interpretation |
| Cloud failure fallback | Paper digest, deep analysis |
I reused existing hardware, but electricity, capacity, maintenance, and depreciation still count. Without energy measurements, “no per-token invoice” is not a zero-cost result.
Token Burn Rate
I used GPT-5.4, GLM5, and Qwen3.5 together at the time. This describes my setup then, not a current default or ranking. Before choosing a Coding Plan, check supported tasks/tools, quota windows, concurrency, and queuing policy. There is no general “all-you-can-eat” guarantee. Budget ongoing automation from actual input, cached input, output, and retries before choosing subscription or pay-as-you-go billing.
Backup Channel
In the early days, OpenClaw (“Shrimp Bro”) occasionally ran into stability issues. When that happened, I would use Claude Code (Remote Control) to connect remotely and troubleshoot manually. This does not happen often, but having a backup channel gives real peace of mind.
The Analogy
Think of local models as “hardworking blue-collar workers” and cloud models as “brilliant senior consultants.” What OpenClaw does is fully automate the collaboration logic between them.
4. How Do You Protect Privacy and Security?
This question comes up constantly, and rightfully so—it is something everyone needs to think through carefully before getting started.
Core Principle
Keeping sensitive data on the machine is my operating goal, not an automatic property of installing a local model. The original principles still need end-to-end enforcement:
- Restrict sensitive files to authorized local inference and local tools; prevent fallbacks, summaries, or attachments from silently reaching a cloud model.
- Telegram, web tools, NotebookLM, and iCloud can all transmit data externally. “Local-first” does not remove those paths.
- Enforce access through effective tool policy, sandbox mounts, host permissions, and credentials. Workspace paths and Markdown instructions alone are not isolation mechanisms.
Technical Safeguards
Command execution: In the v2026.10.3 safe-bin guide, tools.exec.safeBins is a constrained stdin-filter fast path in allowlist mode, not a general read-only certification. Its built-in default list is cut, uniq, head, tail, tr, and wc. Commands such as git and find can write or execute work and are not wholly read-only by name.
deniedFlags constrains arguments in a particular safe-bin profile; it is not a global switch blocking every dangerous command. Execution also depends on host, security/ask policy, path allowlists, tool authorization, and sandboxing. These controls do not prove that prompt injection is impossible.
Sender and command authorization: In the command reference, commands.ownerAllowFrom applies to owner-only command surfaces, separately from commands.allowFrom, channel pairing, and ordinary messages. For Telegram, check channels.telegram.allowFrom, group policy/groupAllowFrom, and the documented account inheritance. The old generic channels.allowFrom is not a universal channel configuration. Permission to issue commands, reach the agent through messages, and invoke a tool are separate controls.
Custom checks: Daily configuration SHA256 comparisons and secret scans were part of the operational setup I configured at the time, not an automatically verified OpenClaw guarantee. Hash changes can be legitimate; protect the baseline and alert path, and confirm the job ran. Daily checks are not real-time detection, and secret scans can miss findings. No alert is not proof of no leak.
Runtime Isolation
- OpenClaw runs on a dedicated Mac Studio, separate from my daily development machine
- Sensitive accounts (company email, internal systems) are not connected to OpenClaw
- External publishing channels like WeChat Official Accounts only have “save draft” permissions, not “publish directly”
These were the limits I set for my deployment. Draft-versus-publish restrictions need actual API, credential, and tool enforcement. If the same browser session can publish, a “draft only” instruction is not complete isolation.
Trust Boundary
I want OpenClaw to access only data I intend this system to process. That does not establish that it is inherently safer than another AI product. Cloud requests, Telegram messages, iCloud sync, and browser sessions can create separate copies. Local logs can themselves contain sensitive information, be incomplete, or be modified.
Audit destinations, readers, retention, and unrecorded operations. Owning the host does not prove that all data and a complete operation history stay exclusively on that host.
The Nature of Security
Default safeguards still need to be checked against your deployment and permissions; they do not automatically meet every security requirement. Build the cage before you raise the shrimp—never the other way around.
5. Day-to-Day Experience: Mobile Control and Data Retention
This one is not asked as often, but I personally think it is the key to OpenClaw feeling truly seamless. It deserves its own section.
Telegram: One Main Bot + Multiple Group Chats
Here is how I use it: the main bot handles everyday conversation, commands, and questions via private chat. Different daily tasks get forwarded to different groups—the Android group gets tech briefings, paper digests, and GitHub updates; the Daily group gets daily reports, weekly reports, and system notifications; other groups are split by topic as needed. Why separate groups? Three reasons:
- No information mixing: Technical content and operational logs are managed separately. Want a specific category? Just open that group.
- Controllable notifications: Important groups get notifications turned on, unimportant ones get muted. No notification overload.
- Easy collaboration: Some groups can include colleagues, sharing the same information sources.
Data Persisted to Obsidian
In my workflow, Telegram is the interaction/notification channel and important outputs are also written to the Obsidian Vault. This helps search, organization, and linking, but persistence is not a backup or a guarantee against loss. Deletion, overwrites, disk failure, and sync conflicts remain possible. Telegram messages being pushed up the chat history also does not mean they were deleted.
Verify the saved file and completeness, and keep a separate restorable backup. iCloud or another sync service can propagate deletions/conflicts; test recovery rather than relying only on a sync indicator.
Here is how I organized my directories at the time:
OpenClaw定时任务/: Scheduled outputs论文/: Original PDF, translation, and reading notesPersonal-Knowledge/source/: Knowledge base reported as 1,760 files at the timeX 文章/: Archived X content
These are personal paths; installing OpenClaw does not ensure every task writes to them.
Mobile Access Anytime
This matters more than I initially expected. When I am out, I pull out my phone, open Telegram, and send commands to Shrimp Bro (虾哥)—no different from sitting at my desk. Want to read today’s daily report? Open Obsidian on mobile—the files are already synced via iCloud. Need to adjust a task urgently? No need to fire up a computer; it is all doable from the phone. My daily routine now starts with checking Telegram in the morning to see what tasks ran overnight and whether anything went wrong. If there is an issue, I handle it right from my phone. When I am out, a quick glance at the running status keeps me informed.
That was my experience when the workflow was available; it still depends on the host, network, channel authorization, and completed synchronization. Tasks needing host interaction or approval may not be fully manageable by phone. The status card’s GPT-4 label, task counts, and health badge are illustrative, not evidence of a measured run. A sensitive Vault required to remain on one machine must not also be assumed safe to sync to iCloud.
Architecture Summary
Telegram handles interaction; the Obsidian Vault stores content. OpenClaw Gateway and related services perform persistent execution; the Vault is not the execution backend. The combination can make daily work convenient, but availability and recovery still depend on monitoring, permissions, sync, and backups—not a promise of uninterrupted 24/7 operation or permanent retention.
Related Posts
About Me && Blog
Below is my personal intro and related links. I look forward to exchanging ideas with fellow professionals. “When three walk together, one can always be my teacher!”
- Blogger Intro
- Blog Content Navigation: A guide for my blog content.
- Curated Excellent Blog Articles - Android Performance Optimization Must-Knows
- Android Performance Optimization Knowledge Planet — historical community introduction; closed in 2026.
One walks faster alone, but a group walks further together.
