Hey,
Three things shipped this week that change what an AI tool is allowed to touch.
Adobe started rolling out an agentic Firefly assistant that runs multi-step jobs across Photoshop, Illustrator, and Premiere — a lighter version is being built directly into Claude. Anthropic shipped connectors letting Claude drive Blender, Ableton, Autodesk, and Splice in plain English. ChatGPT moved model selection — including thinking-effort controls — into the composer where you actually write.
The pattern: agents stopped giving advice and started touching the file. That changes the question you ask before you subscribe to anything.
Last week it was can the agent do the work? This week it's do I trust this thing inside my apps?
Five sections. Skim what you need.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
🔥 TOOL OF THE WEEK — Lindy
The cleanest "AI assistant that actually does things" we've tested. Lindy lives in your inbox, calendar, CRM, and meeting notes — and the way you steer it is by texting it like a coworker. Not by configuring a workflow tree.
Three things to know:
The interface is iMessage. You add a phone number, you text "schedule a 30-min with Sarah next week, prep me with her last 3 emails," it does it. The 60-second setup claim is real — we timed it. For people who already live in their phone, this is the lowest-friction agent on the market right now.
It's credit-based, not seat-based. Simple actions burn 1 credit. Heavy actions (web research, multi-tool chains, voice calls) burn many. The 7-day free trial gives you full Plus access — long enough to figure out whether your workflows are the kind that stay cheap or the kind that don't. Verify current pricing before you commit annual; review sites have been quoting four different numbers.
It belongs to this week's news for a reason. The tools winning right now aren't the smartest models — they're the ones with hands. Lindy has a phone number, a Gmail OAuth, a Google Calendar API, and a CRM connector. That's the trust stack we were talking about up top.
What it's NOT good for: anyone who needs deterministic, repeatable workflows with versioning and audit logs. n8n still owns that. Lindy is for the messy, judgment-heavy automation that lives between humans and tools.
→ Try Lindy: try.lindy.ai
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
⚡ THREE QUICK HITS
GAMMA — the AI deck builder that quietly became the most-used presentation tool inside YC. Free plan gives you 400 credits (around 10 full decks, one-time, no monthly refresh). Plus is $8/mo annual ($10 monthly), removes Gamma branding, unlocks PowerPoint export, and gives you ~600 credits/mo. Pro at $15/mo annual ($18 monthly) adds analytics, custom fonts, and unlimited generation. The Generate API went GA in January — you can now build presentations programmatically. The Studio Mode update (still rolling out) generates full-image cinematic cards. Right now it's the fastest path between "I have an idea" and "I have a deck I'd actually present."
→ try.gamma.app
ELEVENLABS — voice AI's center of gravity. Free plan gives you 10K credits/mo (~10 minutes of audio, no commercial rights). Starter at $5/mo unlocks instant voice cloning and commercial use. Creator at $22/mo ($11 first month) gives you 100K credits, professional voice cloning, and 192kbps quality — the tier most newsletter writers, podcasters, and agencies actually run on. Pro at $99/mo if you're shipping production. The interesting subplot: the Conversational AI tier (separate from TTS) has burst pricing now, which means voice agents can spike to 3x concurrency for double the per-minute cost. That's the shape of voice infra in 2026 — pay-as-you-spike, not flat-rate.
→ try.elevenlabs.io
FOLK — the CRM that's actually built around how solo operators work. Native integrations with Gmail, LinkedIn, and your calendar; the AI fills in contact context automatically; pipelines are flexible without being chaos. The crowd it actually serves: founders, agency owners, solo consultants, people who tried HubSpot and gave up. There's a free trial, and the entry tier is priced for individuals — not teams of 50. If you've been holding contacts in a spreadsheet because every CRM felt like overkill, this is the one designed for you.
→ try.folk.app
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
✍️ PROMPT OF THE WEEK — The Trust Audit
Given the news this week — Adobe, Claude connectors, agents writing files in your apps — the question stopped being "what can this tool do?" and became "what should I let it do?"
Steal it, fill the brackets, paste into Claude/ChatGPT/Gemini before connecting any new agent to your stack:
You are a security-minded operator. I'm about to connect [TOOL NAME] to my workflow. It will have access to: [LIST WHAT IT'S ASKING FOR — email, calendar, files, CRM, etc.].
Walk me through:
1. The MINIMUM permissions this tool needs to do its core job. If it's asking for more, what's the actual use case for the extras?
2. The BLAST RADIUS if this tool is compromised, jailbroken, or gets a malicious instruction from data it ingests (indirect prompt injection). What's the worst legitimate-looking action it could take?
3. The REVOCATION PATH. If something goes wrong, exactly how do I cut access — and how fast?
4. The AUDIT TRAIL. Where do I see what it actually did, in case I need to reconstruct events?
Then tell me: is the trust required by this tool proportional to the work it saves me? If not, what's a narrower scope I should ask for instead?
Why it works: the agentic shift means every new tool is a new permission grant. Most people grant permissions like cookies — click, dismiss, forget. This prompt forces the model to do the work a security review used to do, before you click connect. Run it once per tool. It takes two minutes. It saves the rare-but-expensive incident where a connected tool reads a poisoned doc and acts on it.
More prompts: toolchase.com/prompts/
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
⚖️ COMPARE CORNER — Originality.ai vs GPTZero
The Atlantic ran a piece this week on how generative image tools are being weaponized for fraud — fake invoices, synthetic IDs, doctored evidence. The same pressure is hitting written content from a different angle. AI detection is no longer a college plagiarism story; it's a publishing, hiring, and compliance story.
Two tools dominate the category, and they pick different sides of the accuracy/UX tradeoff.
ORIGINALITY.AI — built for content marketing teams and publishers. The detector is paired with a plagiarism scanner, a fact-checker, and a readability analyzer in one workflow. The reporting is dense and exportable; the API is mature; the brand has a clear point of view about content provenance. Best for: anyone running a content operation where you need to prove your stuff is human-written, or audit submissions from freelancers and agencies.
GPTZERO — built for educators and individual users. The interface is simpler, the free tier is more generous, and it's made paid efforts to publish detection methodology rather than treat it as a black box. Strong on highlighting which sentences look AI-generated rather than just giving you a single percentage. Best for: teachers, hiring managers, journalists, anyone who needs to defend a detection result to the person they're flagging.
The honest take: both have false positives. Both can be defeated by light paraphrasing. Neither will replace human judgment, and any tool promising "100% accuracy" is selling. Originality.ai wins for workflow depth (publishers, agencies, content ops). GPTZero wins for explanatory UX (educators, individual reviewers, anyone who needs to show their work). If detection is part of how you make decisions about people or content, the answer is usually run both and compare.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
📌 ONE EDITORIAL TAKE
For two years, the question that mattered was how smart is the model.
Then it became can it use tools.
This week, it became what is it allowed to touch.
That third question is the one that compounds. A tool with marginally less intelligence but full access to your filesystem, your calendar, and your CRM does more useful work than a frontier model behind a chat box. We watched this play out in coding (Claude Code, Cursor), then in workflow automation (n8n, Lindy), and as of this week it's playing out in creative software too. Adobe's agentic Firefly. Claude inside Blender, Ableton, Autodesk, Splice. The cursor stopped being something you move and started being something you brief.
Two implications worth holding onto:
First, the practical evaluation axis is now permission-shaped. Don't ask "is this AI good." Ask "what does it want to read, what does it want to write, and what's the worst thing that happens if I'm wrong about it?" That's the audit the Trust prompt above is for. If a tool is asking for more access than its job requires, that's not a security flag — that's a roadmap flag. They're telling you what they're going to ask for next.
Second, the trust premium is real, and it shows up in pricing. Tools that hold sensitive context (calendar, inbox, CRM) charge more than tools that just generate text — and they should. Lindy at credit-based pricing, Reclaim and Folk at flat per-seat, the new generation of voice agents at burst pricing. The era of one $20/mo subscription that does everything is closing. What's emerging in its place is a small stack of well-permissioned tools that each do one thing and report cleanly.
We track verified pricing and the actual permissions every tool requests on every page at toolchase.com. If you spot a permission scope that's drifted, hit reply.
That's it for this Wednesday.
If one thing in here changed how you'd evaluate or pick a tool — hit reply and tell me which one. Two-line replies are fine. I read every one.
Talk next week,
Emre
Founder, ToolChase
toolchase.com
P.S. Building an AI tool? Submit it for a free editorial review:
toolchase.com/submit/
P.P.S. Forwarded this email? You can subscribe at:
toolchase.com/newsletter/