Setup OpenClaw 2.0 With Gemini in Under 60 Seconds  🔗 **Source:** [Setup OpenClaw 2.0 With Gemini in Under 60 Seconds](https://twitter.com/_philschmid/status/2094426225573818511/?rw_tt_thread=True) > [!note]+ Document Notes > tags: [openclaw, gemini, google-search-grounding, personal-ai-assistant, self-hosted-agents, vendor-quickstart, model-routing, agent-gateway, chat-channels, api-key-setup, docs-drift, elestio] > author: Philipp Schmid > source: X (formerly Twitter) — @_philschmid > reference: https://twitter.com/_philschmid/status/2094426225573818511/?rw_tt_thread=True > date created: 2026-09-03 12:40 > > > [!info] Generated analysis — not the author's words. Sections below are machine-written > > from the captured source. Only text inside `> [!quote]` blocks is verbatim from the author. > > Do not quote or attribute any other line to the named author. > > > [!warning] Three things to know before trusting this note > > **1. The commands are missing from the capture.** The thread is a five-step install guide, and every one of the eleven code blocks came through Reader as "Some content could not be imported from the original document." The prose around each block survived; the commands did not. This note analyzes the prose, the linked docs, and what the thread claims the commands do. It cannot quote the commands. Inline highlights anchor only to the prose paragraphs. > > **2. Vendor voice.** The author writes "our fastest (+300 token/sec) agentic model," so he speaks for Google. Public bio places him at Google DeepMind in developer relations (general knowledge, not in the capture). The thread is a Gemini onboarding piece that uses OpenClaw as the vehicle; OpenClaw is the third party here. The speed figure links to Artificial Analysis rather than to a Google page, which is better sourcing than most vendor threads manage, but it is still the vendor picking the benchmark. > > **3. "Under 60 seconds" is a headline, not a measurement.** No timing is shown. A global npm install of OpenClaw plus a gateway start is the same procedure Courtney's Elestio host runs via Docker, and nothing in the thread says how long each step took. 540 captured words, published 2026-08-31, 25 outbound links (Google blog, Artificial Analysis, openclaw.ai, docs.openclaw.ai ×9, aistudio.google.com, plus eleven self-links back to the tweet for the missing code). > > # 🏷️ Title: Setup OpenClaw 2.0 with Gemini in Under 60 Seconds > <!-- deep-read:sections start --> ## 🧭 Table of Contents - [📝 Summary](#-summary) - [📖 Main Content](#-main-content) - [🗒️ Notes](#️-notes) - [📚 References](#-references) - [🗣️ Paraphrase](#️-paraphrase) - [📰 Newsletter](#-newsletter) - [✅ Action Items](#-action-items) - [🧪 Hypotheses](#-hypotheses) - [📊 Methodology](#-methodology) - [❓ Questions](#-questions) - [🔮 Future Research](#-future-research) - [💡 Implications](#-implications) - [🧠 Learning Styles Assessment](#-learning-styles-assessment) - [📊 Accessibility Assessment](#-accessibility-assessment) - [🌍 DEIB Assessment](#-deib-assessment) - [✅ Logical Validity](#-logical-validity) - [🛡️ DARVO Analysis](#️-darvo-analysis) - [🎯 Conclusion](#-conclusion) ### 🔗 References - Source: https://twitter.com/_philschmid/status/2094426225573818511/?rw_tt_thread=True - Linked from the thread: Introducing Gemini 3.7 Flash (blog.google) · Artificial Analysis speed leaderboard · openclaw.ai · OpenClaw 2.0 announcement ("openclaw-2-accidentally") · Release notes v2026.8.1 · Google AI Studio API key page · docs.openclaw.ai: gemini-search, image-generation, video-generation, music-generation, providers/google (TTS, realtime voice), channels - Cross-reference: memory `elestio-openclaw-host` (Courtney's production OpenClaw on the same version, 2026.8.1, upgraded 2026-08-31) ## 📝 Summary **What this is.** A Google DevRel quickstart: install OpenClaw 2.0 (v2026.8.1) with npm, export a Gemini API key from AI Studio, set Gemini 3.7 Flash as the default agent model, run one CLI agent turn to prove inference, start the gateway in local mode and open the Control UI at `http://127.0.0.1:18789/`. Then a tour of what else a Gemini key unlocks inside OpenClaw: Google Search grounding on by default, image generation (`gemini-3.1-flash-image`), video (`veo-3.1-fast-generate-preview`), music (`lyria-3-clip-preview`), batch TTS, realtime voice via the Live API, and channel connectors for Telegram, WhatsApp, Discord. **The pitch.** OpenClaw is "a personal AI assistant that actually does things" and 2.0 is "the largest update yet: simpler install, a rebuilt browser app that opens straight into a conversation, and setup you can finish by talking to your Claw." Gemini 3.7 Flash is the model you should put in it, because it is fast and because a Gemini key gives you web search for free (the thread twice tells you to ignore the Brave API key onboarding tip). **What it leaves out.** Everything that made Courtney's 2026-08-31 upgrade to the same version cost hours: config-key migrations doctor won't persist non-interactively, auth-profile migration that needs the gateway stopped, the zero-byte attestation deadlock, root-owned config files, the gateway token that moved from env to config. None of that applies to a fresh laptop install, which is the only case the thread covers. The thread is accurate for a first install and silent on day two. **Bottom line.** Same software version she runs, different everything else: npm global versus a Docker thin overlay, a Gemini API key versus a Claude subscription OAuth profile, local loopback versus a Tailscale-fronted VM with seven agents and WhatsApp. The one transferable idea is that a single provider key can replace a separate search-API dependency. She currently carries a Firecrawl key in plaintext for web search; the thread says a Gemini key makes search grounding a built-in. Whether that is worth adding Google as a provider is a routing decision, not a setup step. ## 📖 Main Content **Structure.** Prerequisites → five numbered steps → "Next Steps" (search grounding) → "Leveraging Other Gemini Capabilities" (five modalities) → "Connecting Chat Channels" (six CLI commands, four of which survived as prose because they were inline code rather than fenced blocks). Each step is one sentence of intent, one code block (lost), and in three cases an italic note pre-empting a confusing message the user will see. **The three pre-emptive notes are the most useful lines in the thread**, because they document real UX rough edges in the product being promoted: 1. Onboarding will suggest a Brave API key; ignore it, Gemini's grounding is used automatically. 2. Setting `gemini-3.7-flash` as default may warn that the model "is not in the local catalog"; ignore it, the model resolves dynamically against AI Studio on the next step. 3. The Control UI opens "authenticated with your local session token," which is how the reader learns a token exists at all. Each note is a small admission that the product's messages contradict the happy path. A quickstart that has to say "safe to ignore" twice is describing a product whose defaults still assume a different provider (Brave for search, a static model catalog). **Prerequisites carry a version constraint worth reading twice.** Node 26 recommended, or 22.22.3+ / 24.15+. Those are specific patch floors, which usually means a known incompatibility below them. Courtney's laptop runs Node 25.9.0, an odd-numbered non-LTS line the thread doesn't mention at all. Her Elestio host sidesteps this entirely because the Docker image ships its own Node. **The gateway paragraph is the only place the thread touches operations.** Local mode, then an aside: "To run as a persistent background daemon instead: `openclaw gateway install && openclaw gateway start`. To stop it: `openclaw gateway stop --force`." That is the whole of the production guidance. Nothing on where config lives, what the gateway token is for, how channels authenticate, or what happens on upgrade. **Search grounding is the actual product argument.** "Google Search grounding is enabled by default when GEMINI_API_KEY is present. No extra configuration or search keys are required." In chat, "Gemini autonomously runs web searches with citations whenever fresh context is needed." This is the sentence that differentiates a Gemini-backed OpenClaw from an Anthropic- or OpenAI-backed one at the setup layer: one key, two capabilities. It is also a lock-in: the search capability is a property of the provider, not of OpenClaw, so switching the default model later removes search unless another search provider is configured. **The capabilities tour is a model catalog.** Five preview-suffixed model IDs (`veo-3.1-fast-generate-preview`, `lyria-3-clip-preview`, `gemini-3.1-flash-tts-preview`, `gemini-3.1-flash-live-preview`) plus one GA image model. Four of five are previews. The thread doesn't say so; the suffixes do. **Channels.** `openclaw channels list --all`, `channels add` (interactive), Telegram and Discord by bot token, WhatsApp by QR login, `channels status --probe` to verify. Matches what Courtney's host uses (WhatsApp + webchat), and the `--probe` verb is one she can use on the VM. ## 🗒️ Notes - **Same version, no drift.** The thread pins v2026.8.1 twice and links its release notes. Courtney's Elestio host went to 2026.8.1 on 2026-08-31, the day this thread was published. Anything the thread says about 2.0 behavior applies to her install; anything it says about setup does not, because hers is an upgrade of a customized Docker stack, not a fresh npm install. - **The commands are gone.** Eleven `rw-outer-content` placeholders where fenced code blocks were. The five install commands, the export line, the `openclaw models set` line, the verify command, the agent-turn command, its output, the gateway command, and the Control UI open command are all missing. The prose says what each does; nothing in this doc can be pasted. The inline-code fragments that survived (`openclaw gateway install && openclaw gateway start`, `openclaw capability web search --query`, `openclaw models set-image google/gemini-3.1-flash-image`, the six `openclaw channels …` commands) are the only executable text in the capture. - **Provider model on her host.** All seven agents run on `anthropic:subscription` (OAuth), migrated off the API key on 2026-08-31. The thread's path is an API key exported into the shell and persisted "into OpenClaw's local profile store," which is the same profile mechanism whose migration produced her SECRETS_DEGRADED errors. On a fresh install there is nothing to migrate. - **Search.** Her openclaw.json carries a Firecrawl `webSearch.apiKey` in plaintext (open item 3 in the host memory). The thread's claim that a Gemini key gives search grounding "with no extra configuration or search keys" is a candidate simplification: add Google as a provider, route only web-search-shaped turns to a Gemini model, drop the Firecrawl key. Not tested; the grounding is a Gemini feature and would not extend to turns served by Claude. - **Gateway auth.** The thread says the Control UI is "authenticated with your local session token." On her host, 2026.8.1 enforces `gateway.auth.token` from config and ignores the env token, which broke the Obsidian plugin until re-pasted. The thread's local mode hides this because loopback plus a session token needs no manual token handling. - **What "simpler install" means.** The 2.0 announcement is linked but not summarized. The thread's evidence for "simpler" is that the install is five commands. Her upgrade notes list six doctor gotchas for the same version. Both can be true: install got simpler, upgrade did not. - **Brave.** The onboarding still proposes a Brave API key, so OpenClaw's default search assumption is Brave, not the model provider. That is a product default the thread is routing around, not a feature it is documenting. - **The "+300 token/sec" link** goes to Artificial Analysis's speed chart, a third-party leaderboard. Better than a self-citation, still vendor-selected. The number is output speed, which is the right metric for a chat assistant and the wrong one for a coding agent, where tool-call latency and correctness dominate. ## 📚 References - Schmid, P. (@_philschmid), "Setup OpenClaw 2.0 with Gemini in Under 60 Seconds," X thread, 2026-08-31. Captured 2026-09-01, 540 words, code blocks not captured. - Google, "Introducing Gemini 3.7 Flash," blog.google (linked, not fetched). - Artificial Analysis, model speed leaderboard (linked, not fetched). - OpenClaw, "OpenClaw 2.0" announcement and v2026.8.1 release notes, openclaw.ai / docs.openclaw.ai (linked, not fetched). - OpenClaw docs: Gemini search grounding, image/video/music generation, Google provider (TTS, realtime voice), channels (linked, not fetched). - Cross-reference: memory `elestio-openclaw-host` — the 2026-08-31 upgrade record for the same version. ## 🗣️ Paraphrase Install OpenClaw from npm. Get a free Gemini key from AI Studio and export it, then save it into OpenClaw's profile. Tell OpenClaw to use Gemini 3.7 Flash by default; if it complains the model isn't in its catalog, ignore that, it checks with Google next. Run one agent turn from the terminal to prove it works. Start the gateway locally and open the browser UI on port 18789. You now have web search built in because Gemini does grounding itself, so skip the Brave key it asks about. The same key also unlocks image, video, music, speech, and live voice models, and you can wire the assistant into Telegram, WhatsApp, or Discord with one command each. ## 📰 Newsletter **Google's OpenClaw quickstart is a Gemini onboarding page wearing an open-source hat.** Philipp Schmid's five-step thread installs OpenClaw 2.0 (v2026.8.1, released the same week) and points it at Gemini 3.7 Flash. The setup is plausible and the pre-emptive notes are honest: onboarding will ask for a Brave search key you don't need, and the model picker will warn about a model it hasn't heard of yet. The real argument is one sentence: with a Gemini key present, "Google Search grounding is enabled by default… No extra configuration or search keys are required." That collapses two dependencies into one, and it also ties search to the provider, so swapping models later silently removes it. Missing from the thread: anything about running this past day one. The same version took an existing self-hoster six documented workarounds to upgrade. If you're starting fresh, the thread is fine. If you already run OpenClaw, the only line worth acting on is the search-grounding one, and only if you'd accept a second provider to get it. Read the Reader capture with care: all eleven code blocks failed to import, so the commands themselves aren't there. ## ✅ Action Items **DevRel** - The three "safe to ignore" notes are a documentation pattern: pre-empt the exact misleading message the user will see, in the step where they'll see it. Steal it for any quickstart. `#action/h3` - A quickstart that links a third-party benchmark instead of a first-party claim is the minimum bar for speed numbers. Hold vendor threads to it. `#action/h4` **OSPO** - Not applicable to this source. **Community** - Not applicable to this source. **WordPress teams** (Core, Core Performance, Core AI, Core Program Team, Mobile, Playground, Community, Training, Marketing, Support, Design, Documentation, Photos, TV, Polyglots, Accessibility, Meta, CLI, Hosting, Tide, Openverse, Themes, Plugins) - **Core AI:** the "search grounding is a provider feature" pattern is the same trap for any WordPress AI abstraction: a capability that exists only when one provider is selected must be surfaced as provider-conditional in the UI, or users lose it on a model switch without knowing why. `#action/h3` - **Documentation:** the pre-emptive "you will see X, ignore it" note beside the step that triggers X is a handbook pattern worth adopting for install pages. `#action/h4` - All other teams: not applicable to this source. **The WPOCC** - Not applicable to this source. **FAIR** - Not applicable to this source. **Personal (Courtney)** - **Test the search-grounding claim on the Elestio host** before changing anything: add Google as a provider with an AI Studio key on one non-critical agent (scholar or health), confirm `openclaw capability web search --query "…"` returns cited results without Firecrawl, then decide whether retiring the plaintext Firecrawl key (open item 3) is worth carrying a second provider. `#action/h2` - **Run `openclaw channels status --probe`** on the VM. It's the one command in the thread you haven't used and it answers the review-cron "no route" question (open item 1) faster than reading logs. `#action/h2` - **Don't re-read this as an upgrade guide.** Your `elestio-openclaw-host` memory is the upgrade guide for this version; the thread covers only fresh installs. `#action/h4` - **Re-save the thread or fetch the original** if you want the actual commands; Reader dropped all eleven code blocks. `#action/h4` ## 🧪 Hypotheses 1. **A Gemini key replaces the Firecrawl search dependency on her host for Gemini-served turns only.** Test: one agent on `google/gemini-3.7-flash`, Firecrawl key removed for that agent's config, ask a fresh-news question, check for citations. Falsified if the agent falls back to no search or errors on the missing key. 2. **The "not in local catalog" warning is a 2026.8.1 catalog-staleness artifact, not a Gemini-specific one.** Test: set any model released after the OpenClaw build date and look for the same warning. If it appears for a Claude model too, the warning is about the catalog, and the thread's "safe to ignore" generalizes. 3. **"Under 60 seconds" excludes npm install time.** Test: time the five steps on a clean machine with Node 26. If `npm install -g openclaw` alone exceeds a minute on a normal connection, the headline counts only the commands typed, not the wall clock. ### 📊 Methodology A vendor quickstart with no measurements: no timing, no benchmark run, one third-party speed link. The evidence for "simpler install" is the step count. The evidence for search grounding is the docs page it links. Nothing in the thread is reproducible from the Reader capture because the commands were lost; reproduction requires the original tweet or the linked docs. ### ❓ Questions - What does `openclaw models set` do when the model isn't in the local catalog: store the string and defer validation, or fail on the next non-Gemini provider? The thread says validation happens "in the next step," which implies deferral. - Does Google Search grounding apply to every Gemini model in OpenClaw or only to the ones AI Studio marks grounding-capable? - Where does the "local profile store" live on disk, and is it the same `auth-profiles` store whose migration produced SECRETS_DEGRADED on the Elestio host? - Why does onboarding still propose Brave when a Gemini key is present? Ordering bug, or does onboarding run before the key is read? ### 🔮 Future Research - Compare the day-one experience across providers on a fresh 2026.8.1 install: Gemini (this thread), Claude subscription (her host), OpenAI. Which provider needs the fewest "ignore this warning" notes? - Trace whether the 2.0 "setup you can finish by talking to your Claw" path persists config keys that non-interactive doctor won't (host memory, doctor gotcha 1). If the conversational setup writes them, it's a workaround for the upgrade pain too. - Check the four preview model IDs against Google's model lifecycle page in 60 days; preview IDs in a quickstart go stale fast. ### 💡 Implications - **Capability bundled with provider is lock-in by another name.** Free search grounding is a real convenience and a real switching cost. Anyone building on OpenClaw with Gemini should configure a provider-independent search tool before they need it, not after. - **Quickstart truth and operations truth diverge on the same version.** Both this thread and her upgrade record describe 2026.8.1 accurately. The gap between them is the gap between "install" and "run for a year." Documentation that only covers the first is not wrong; it's incomplete in a way that costs the reader hours later. - **Preview-suffixed model IDs in onboarding content are a maintenance liability.** Four of the five capability examples will need editing when the previews graduate or die. ## 🧠 Learning Styles Assessment - **Visual:** four screenshots (install output, CLI agent turn, Control UI, search result) carried as images; the images survived the capture, the code didn't. Visual learners get more from this doc than readers do. - **Spatial:** the five-step sequence is linear; no architecture diagram of gateway, node, channels. - **Diagrammatic:** none. - **Auditory:** not applicable; text thread. - **Reading/writing:** the prose is instruction-shaped and clear; without the code blocks it reads as a description of a tutorial rather than a tutorial. - **Kinesthetic:** the whole thread is do-this-then-that; the only learning style it serves fully, and only from the original tweet. - **Social:** none; no community, forum, or issue tracker links. - **Solitary:** fits a single developer at a terminal. - **Logical:** the pre-emptive notes reward readers who want to know why a warning appears; the rest is procedural. ## 📊 Accessibility Assessment #### Alternative text and non-text content Four images with alt text "Image" and one with the thread title. The screenshots carry the command outputs, so a screen-reader user gets neither the commands (lost) nor the outputs (unlabeled). Fails. #### Keyboard navigation Not applicable; static thread. #### Contrast and readability Not assessable from the capture. #### Text resizing and reflow Not applicable. #### Media and captions No audio or video. #### Forms and interactions None. #### Standards reference WCAG 2.2 SC 1.1.1 (non-text content) is the relevant failure: the information is in the images and the alt text is "Image." ## 🌍 DEIB Assessment #### Cultural breadth US/EU developer default: English, npm, a Google account. WhatsApp channel support is the one nod to markets where WhatsApp is the primary chat surface. #### Inclusive language Neutral throughout. "Your Claw" is product voice, not exclusionary. #### Neurodiversity support Numbered steps, one action per step, warnings placed at the point of confusion. Good structure for readers who need explicit sequencing. #### Representation Single author, vendor voice. No user stories. #### Psychological safety The "safe to ignore" notes reduce the fear that a warning means you broke something. Small but real. ## ✅ Logical Validity **Strong reasoning.** The thread's core claim, that a Gemini key gives you search grounding without a separate search provider, is specific, links its documentation, and is falsifiable in one command. The pre-emptive notes show the author actually ran the steps and hit the warnings. **Weak reasoning.** "Under 60 seconds" and "largest update yet" are asserted with no measure. "Our fastest agentic model for coding and agents" cites an output-speed leaderboard, which measures tokens per second, not agentic task completion; the link supports "fastest" and not "for coding and agents." **Fallacies.** - *Selective benchmark:* speed is the one axis where Flash-class models lead; quality and tool-use reliability are not mentioned. - *Composition:* "simpler install" is presented as evidence that 2.0 is simpler overall; upgrade complexity on the same version says otherwise. **Consistency.** Internally consistent. No contradictions between steps. ## 🛡️ DARVO Analysis Not applicable to this source. No conflict, no accusation, no reversal. The thread is promotional, not defensive. ## 🎯 Conclusion An accurate fresh-install guide for the exact OpenClaw version Courtney runs, written to sell a Gemini key. Its one durable idea is that search grounding can ride on the model provider, which is both a simplification and a coupling. Its one durable gap is that day two doesn't exist in it. For her, the action is a test, not an adoption: prove the grounding claim on one agent, then decide whether a second provider is worth retiring a plaintext search key. ## 🔄 Refactor Appendix — Setup OpenClaw 2.0 with Gemini in Under 60 Seconds ### 📚 Literature Atoms (H3) #### [H1] Provider-bundled search removes a dependency and adds a coupling > [!quote] ID:rw_01m1dgprxz2sj3waqnzvx3fzz4 > "Google Search grounding is enabled by default when GEMINI_API_KEY is present. No extra configuration or search keys are required." One key, two capabilities. The search capability belongs to the provider, so a later model switch removes it silently unless a provider-neutral search tool was configured first. Tags: `#atom/literature/h1` #### [H2] A quickstart's "safe to ignore" notes are a map of the product's stale defaults > [!quote] ID:rw_01m1dgprxz2sj3waqnzvx3fzz4 > "(Note: If OpenClaw displays a warning that the model is not in the local catalog, it is safe to ignore. The model is dynamically resolved and validated against Google AI Studio in the next step.)" Two such notes in five steps (Brave key, local catalog). Each marks a place where the product still assumes a different provider or a static catalog. Read them as a defect list the author chose to route around. Tags: `#atom/literature/h2` #### [H3] "Largest update yet" is measured in install steps, not in operations > [!quote] ID:rw_01m1dgprxz2sj3waqnzvx3fzz4 > "OpenClaw 2.0 (v2026.8.1) is the largest update yet: simpler install, a rebuilt browser app that opens straight into a conversation, and setup you can finish by talking to your Claw." The same version took six documented workarounds to upgrade on a customized host (memory `elestio-openclaw-host`). Install simplicity and upgrade complexity coexist in one release. Tags: `#atom/literature/h3` #### [H4] Speed is the vendor's chosen axis > [!quote] ID:rw_01m1dgprxz2sj3waqnzvx3fzz4 > "Five commands to install OpenClaw 2.0 (v2026.8.1), connect Gemini 3.7 Flash, our fastest (+300 token/sec) agentic model for coding and agents, and start chatting with Google search grounding in under a minute." Output tokens per second, sourced to a third-party leaderboard, presented as the qualification for "coding and agents." Tool-call correctness, the metric that matters for agents, is not mentioned. Tags: `#atom/literature/h4` ### 🧠 Permanent Candidates (H4) #### [P1] Install documentation and operations documentation describe the same version and different products A fresh-install quickstart and an upgrade record for one release can both be accurate and share almost no content. The quickstart's truth ends at first run. Judge a project's documentation by whether the second document exists, not by how short the first one is. *Candidate; one source plus one first-hand upgrade record.* Links: [[Elestio OpenClaw Host]] · [[Documentation Debt Shows Up on Day Two]] · [[AI Tooling MOC]] Tags: `#atom/zettel/z1` #### [P2] A capability that exists only under one provider must be labeled provider-conditional or it becomes silent lock-in Search grounding tied to a Gemini key is the instance; any abstraction layer that exposes provider-specific features without flagging the condition will lose them on a switch, and the user will not know why. Applies to WordPress AI abstractions as much as to OpenClaw. *Candidate; the mechanism is the thread's, the generalization is mine.* Links: [[Provider Lock-In by Feature]] · [[Model Routing Policy]] · [[WordPress AI Abstraction Layer]] Tags: `#atom/zettel/z2` ### ✏️ Fleeting Queue - Reader dropped all eleven fenced code blocks from an X thread but kept inline code. Check whether that's an X-capture limitation or a Reader parser rule; it affects every tutorial thread saved. `#atom/fleeting/f1` - Node floors 22.22.3 / 24.15 are specific enough to imply a known bug below them; find the OpenClaw issue that set them. `#atom/fleeting/f1` - Four of five capability examples use `-preview` model IDs. Set a 60-day reminder to check which still resolve. `#atom/fleeting/f1` ### → Things 3 (max 2 actions) 1. **Test Gemini search grounding on one Elestio agent** (AI Studio key, `openclaw capability web search --query`), then decide whether to retire the plaintext Firecrawl key. `#action/h2` 2. **Run `openclaw channels status --probe` on the VM** and use the output to re-point the four failing review crons. `#action/h2` <!-- deep-read:sections end --> 📄 **Full text:** [[Full Document Contents/Tweets/Setup OpenClaw 2.0 With Gemini in Under 60 Seconds.md|Setup OpenClaw 2.0 With Gemini in Under 60 Seconds by Philipp Schmid]] ## Highlights - [OpenClaw](https://openclaw.ai/) is a personal AI assistant that actually does things. It runs on your devices, in your chats, with your rules. ([View Highlight](https://read.readwise.io/read/01m1kr9pq4rrhzs91vt7jw7g96)) - Note: H4 — The product framing; the thread is a Gemini onboarding piece using OpenClaw as the vehicle #atom/literature/h4 - Tags: atom/literature/h4 - [OpenClaw 2.0](https://openclaw.ai/blog/openclaw-2-accidentally) (**v2026.8.1**) is the largest update yet: simpler install, a rebuilt browser app that opens straight into a conversation, and setup you can finish by talking to your Claw. Full notes: [v2026.8.1](https://docs.openclaw.ai/releases/2026.8.1). ([View Highlight](https://read.readwise.io/read/01m1kr9kfrytn0bsv99n7aqs32)) - Note: H3 — Same version as the Elestio host (2026.8.1, upgraded 2026-08-31); "simpler install" and six upgrade workarounds coexist in one release #atom/literature/h3 - Tags: atom/literature/h3 - *(Note: If OpenClaw displays a warning that the model is not in the local catalog, it is safe to ignore. The model is dynamically resolved and validated against Google AI Studio in the next step.)* ([View Highlight](https://read.readwise.io/read/01m1kr9gkre514r8q1mwtt9rjx)) - Note: H2 — "Safe to ignore" notes map the product's stale defaults (Brave search, static catalog) #atom/literature/h2 - Tags: atom/literature/h2 - Google Search grounding is **enabled by default** when **GEMINI\_API\_KEY** is present. No extra configuration or search keys are required. ([View Highlight](https://read.readwise.io/read/01m1kr9ddb197x57j6gzg9skex)) - Note: H1 — Provider-bundled search removes a dependency and adds a coupling; switching models later removes search silently #atom/literature/h1 - Tags: atom/literature/h1