I Need You Guys to Trust Me on This - CourtneyR.dev's 2nd Brain

I Need You Guys to Trust Me on This ![rw-book-cover](https://img.youtube.com/vi/RIkSlHgQYog/maxresdefault.jpg) 🔗 **Source:** [I Need You Guys to Trust Me on This](https://www.youtube.com/watch?v=RIkSlHgQYog) > [!note]+ Document Notes > tags: #claude #anthropic #open-source #developer-tools #api-access > author: Theo - t3․gg > source: YouTube > reference: https://www.youtube.com/watch?v=RIkSlHgQYog > date created: 2024-12-19 14:30 > > # 🏷️ Title: I need you guys to trust me on this (sorry Anthropic) >  ## 🧰 Action Inbox - [H1] Evaluate T3 Code integration approach for Claude subscriptions — 00:00 #claude - [H2] Review Anthropic's policy clarity around SDK usage — 05:23 #anthropic - [H3] Monitor account safety when using alternative Claude interfaces — 12:45 #developer-tools ## 🧭 Table of Contents - [[#📝 Summary]] - [[#📖 Main Content]] - [[#🗒️ Notes]] - [[#📚 References]] - [[#🗣️ Paraphrase]] - [[#📰 Newsletter]] - [[#✅ Action Items]] - [[#🧪 Hypotheses]] - [[#📊 Methodology]] - [[#❓ Questions]] - [[#🔮 Future Research]] - [[#💡 Implications]] - [[#🧠 Learning Styles Assessment]] - [[#📊 Accessibility Assessment]] - [[#🌍 DEIB Assessment]] - [[#✅ Logical Validity]] - [[#🛡️ DARVO Analysis]] - [[#🎯 Conclusion]] ### 🔗 References - [T3 Code](https://t3.gg) - [Open Code GitHub](https://github.com/opencodedev/opencode) - [Claude Agent SDK](https://docs.anthropic.com/claude/docs) ## 📝 Summary - T3 Code now supports Claude Code subscriptions through official SDK integration - Anthropic has been blocking other open-source tools like Open Code with legal threats - T3 uses official CLI integration rather than custom harnesses to stay compliant - Anthropic's unclear policies create uncertainty for developers using subsidized tokens - Users can access $5,000 monthly compute for $200 through Claude subscriptions vs limited API access ## 📖 Main Content Theo announces that T3 Code now supports Claude Code subscriptions, contrasting their approach with Open Code's recent legal troubles with Anthropic. The key technical difference is that T3 Code uses the official Claude Agent SDK and CLI integration rather than creating custom authentication flows or harnesses that Anthropic considers policy violations. The controversy stems from Anthropic's aggressive enforcement against Open Code while maintaining unclear policies around third-party integrations. Anthropic specifically prohibits third-party developers from offering "Claude.AI login or rate limits" but allows official SDK usage. T3 Code's implementation avoids authentication management by requiring users to set up the CLI independently. The economic incentive is significant: Claude Code subscriptions offer up to $5,000 in monthly compute for $200, heavily subsidizing usage compared to API pricing. This makes direct API subscriptions uncompetitive for power users, creating pressure for creative integration approaches. ## 🗒️ Notes #### ⏱️ [00:00](https://youtu.be/RIkSlHgQYog?t=0) — [H1] T3 Code Claude integration announcement - Key idea: T3 Code now supports existing Claude Code subscriptions through official SDK - Evidence: "If you already have Claude Code on your machine, all you have to do is run the npxt3 command" #### ⏱️ [01:15](https://youtu.be/RIkSlHgQYog?t=75) — [H2] Open Code legal troubles with Anthropic - Key idea: Anthropic sent lawyers to block Open Code's Claude integration - Evidence: "Open Code 1.3.0 will no longer autoload the Claude Max plugin. We did our best to convince Anthropic to support developer choice, but they sent lawyers instead" #### ⏱️ [03:45](https://youtu.be/RIkSlHgQYog?t=225) — [H3] Legal disclaimer and risk assessment - Key idea: No guarantees against potential account bans despite compliance efforts - Evidence: "I cannot control what Anthropic does and their ability to ban you for anything they feel like at any time" #### ⏱️ [06:30](https://youtu.be/RIkSlHgQYog?t=390) — [H1] Technical compliance strategy - Key idea: T3 Code uses official SDK and CLI rather than custom harnesses - Evidence: "We're using the official SDK. Open code's using their own harness, which means that they don't have the built-in authentication methods" #### ⏱️ [09:15](https://youtu.be/RIkSlHgQYog?t=555) — [H2] Economic incentives behind Claude subscriptions - Key idea: Claude Code subscriptions offer massive compute subsidization - Evidence: "Anthropic's Cloud Code sub allows you to use up to $5,000 in compute in a month when you're only paying $200" #### ⏱️ [12:00](https://youtu.be/RIkSlHgQYog?t=720) — [H3] Developer community frustration with unclear policies - Key idea: Matt Pocock's public questions remain unanswered by Anthropic - Evidence: "Can I use an OOTH token generated from a sub to power the cloud agent SDK strictly for using cloud code in a local dev loop?" #### ⏱️ [15:30](https://youtu.be/RIkSlHgQYog?t=930) — [H2] Anthropic's communication failures - Key idea: Multi-billion dollar company fails to provide basic policy clarity - Evidence: "It's been over a week and Matt doesn't want to keep waiting. Anthropic AI bump. Working on it." #### ⏱️ [18:45](https://youtu.be/RIkSlHgQYog?t=1125) — [H1] T3 Code's architectural choices for compliance - Key idea: Built abstraction layer to use official CLIs rather than custom harnesses - Evidence: "We abstracted the layer that manages all of our events so that we can have different models and different providers" #### ⏱️ [21:00](https://youtu.be/RIkSlHgQYog?t=1260) — [H3] Alternative options with higher risk - Key idea: Open Code forks still exist but carry account ban risk - Evidence: "There's a fork of the Open Code Claude O plugin that is still being maintained that lets you use your Claude Code credentials" #### ⏱️ [23:15](https://youtu.be/RIkSlHgQYog?t=1395) — [H2] Future policy uncertainty - Key idea: Anthropic's vague policies allow arbitrary enforcement changes - Evidence: "At any point this goalpost can move because Anthropic loves to do that. They're keeping things vague so they can change their mind" #### ⏱️ [25:30](https://youtu.be/RIkSlHgQYog?t=1530) — [H3] Multi-model workflow benefits - Key idea: T3 Code enables unified interface across different AI subscriptions - Evidence: "I live a very multimodel life where I use the $200 month sub from Codeex and Claude and use a meaningful amount of both" ## 📚 References - [T3 Code GitHub](https://github.com/t3-oss/t3-code) - [Claude Agent SDK Documentation](https://docs.anthropic.com/claude/docs) - [Anthropic Terms of Service](https://www.anthropic.com/terms) - [Open Code Project](https://github.com/opencodedev/opencode) ## 🗣️ Paraphrase - T3 Code integrated Claude subscriptions through official SDK compliance rather than custom authentication flows that triggered Anthropic's legal response - The economic incentive is substantial: Claude subscriptions provide 25x more compute value than API pricing, making creative integration approaches financially necessary - Anthropic's policy ambiguity creates widespread developer uncertainty while they selectively enforce against some projects but not others ## 📰 Newsletter T3 Code now supports your existing Claude Code subscriptions! Unlike other tools facing legal challenges, T3 Code uses Anthropic's official SDK and CLI integration to stay compliant while giving you access to that heavily subsidized compute ($5,000/month for $200). The developer community is frustrated with Anthropic's unclear policies around third-party integrations - even respectful inquiries from prominent developers like Matt Pocock go unanswered for weeks. T3 Code's approach avoids custom authentication by requiring you to set up the Claude CLI yourself, then accessing it through official channels. Check out the open-source project and see how they built a unified interface for multiple AI subscriptions while navigating complex vendor policies. Resources: [T3 Code](https://t3.gg), [Claude SDK Docs](https://docs.anthropic.com/claude/docs), [GitHub Repo](https://github.com/t3-oss/t3-code) ## ✅ Action Items #### DevRel - Developer advocacy: Document SDK integration patterns that maintain vendor compliance - Education: Create workshops on multi-vendor AI tool integration strategies - Developer experience: Build abstraction layers that isolate vendor-specific authentication - Developer marketing: Highlight economic benefits of subscription vs API pricing models - Community management: Monitor vendor policy changes and communicate impacts #### OSPO - Policy & Governance: Establish guidelines for AI vendor integration compliance - License Compliance: Review third-party SDK usage and authentication requirements - Strategic Alignment: Track vendor policy changes affecting open-source projects #### Community - Contributors: Support developers navigating unclear AI vendor policies - Maintainers: Share compliance strategies for AI service integrations - Ecosystem: Coordinate response to aggressive vendor enforcement actions ## 🧪 Hypotheses - Official SDK usage provides stronger legal protection than custom authentication implementations - Vendor policy ambiguity is intentional to enable selective enforcement against competitive threats - Economic subsidization of AI subscriptions over APIs will continue driving creative integration approaches - Developer community pressure can eventually force policy clarification from AI vendors ### 📊 Methodology - Study design to test hypothesis 1: Compare enforcement actions against SDK vs custom implementations - Data sources and analysis for hypothesis 2: Track vendor communication patterns and policy updates - Experiments to validate hypothesis 3: Monitor pricing differential impacts on integration patterns ### ❓ Questions - Will Anthropic provide clear technical guidelines for compliant third-party integrations? - How will other AI vendors respond to similar integration attempts? - What legal precedents exist for SDK usage rights in third-party applications? ### 🔮 Future Research - Comparative analysis of AI vendor integration policies across major providers - Economic impact study of subsidized subscriptions vs API pricing on developer adoption - Legal framework development for AI service integration rights ### 💡 Implications - Practice: Developers must architect around vendor policy uncertainty - Policy: Clear technical guidelines needed for AI service integration compliance - Security: Official SDK usage reduces authentication vulnerability risks - Education: Training needed on vendor relationship management for open-source projects - Funding: Economic incentives drive architectural decisions in AI tooling ### 🧠 Learning Styles Assessment Reading level estimate: 12th grade #### 👀 Visual learners - Present: Screenshots of T3 Code interface, architectural diagrams implied - Improve: Authentication flow diagrams, compliance decision trees #### 🔊 Aural learners - Present: Full video presentation with clear narration - Improve: Podcast-format discussion of AI vendor policies #### 🗣️ Verbal learners - Present: Technical terminology explained, policy distinctions clarified - Improve: Glossary of AI integration terms, vendor comparison tables #### 🧩 Kinesthetic learners - Present: Installation commands, step-by-step CLI setup - Improve: Hands-on integration workshop, compliance checklist #### 🧠 Logical learners - Present: Technical architecture reasoning, compliance logic flow - Improve: Decision matrices for integration approaches, risk assessment frameworks #### 🧑‍🤝‍🧑 Social learners - Present: Community discussion context, developer testimonials - Improve: Developer forums for integration strategies, peer review processes #### 🧘‍♀️ Solitary learners - Present: Individual implementation guide, self-assessment questions - Improve: Personal compliance checklists, reflection prompts on vendor relationships #### 🌿 Naturalistic learners - Present: Ecosystem thinking about vendor relationships - Improve: Industry evolution analogies, competitive landscape analysis ### 📊 Accessibility Assessment #### Screen reader semantics - Present: Clear heading structure in video transcript - Improve: Structured navigation landmarks, technical term definitions #### Keyboard navigation - Present: Standard video controls available - Improve: Transcript search functionality, chapter navigation #### Contrast and readability - Present: High contrast video interface - Improve: Transcript formatting with appropriate contrast ratios #### Text resizing and reflow - Present: YouTube's built-in accessibility features - Improve: Responsive transcript layout for different screen sizes #### Media and captions - Present: Auto-generated captions available - Improve: Professional captions with technical term accuracy, downloadable transcript #### Forms and interactions - Present: Basic video interaction controls - Improve: Interactive code examples with proper labeling ### 🌍 DEIB Assessment #### Cultural breadth - Present: Global developer community context - Improve: Multi-regional AI policy examples, international compliance perspectives #### Inclusive language - Present: Technical focus with minimal jargon explanation - Improve: Plain language summaries, acronym definitions #### Neurodiversity support - Present: Clear section breaks in video content - Improve: Content warnings for rapid technical discussions, structured information hierarchy #### Representation - Present: Developer-focused perspective - Improve: Diverse stakeholder viewpoints, impact on different developer communities #### Psychological safety - Present: Respectful criticism of corporate policies - Improve: Community guidelines for policy discussions, constructive feedback frameworks ### ✅ Logical Validity #### Strong reasoning - Deductive strength: Legal policy analysis follows from stated terms of service - Inductive strength: Pattern recognition in vendor enforcement actions - Causal reasoning: Economic incentives driving architectural choices - Analogical strength: Comparison with other open-source integration challenges #### Fallacies - Appeal to consequences: "If you get banned" reasoning, but acknowledges uncertainty appropriately - False dichotomy: Presents T3 vs Open Code as main options, but notes other alternatives exist #### Tone of voice - Overall tone: Frustrated but professional, technical authority with appropriate disclaimers ### 🛡️ DARVO Analysis - Deny: No denial patterns observed - Attack: Criticism directed at Anthropic's policy clarity, not individuals - Reverse victim/offender: No reversal patterns present - Emotional cues and rhetoric: Frustration expressed constructively toward policy improvement ## 🎯 Conclusion T3 Code's Claude integration demonstrates how technical architecture choices can navigate vendor policy uncertainty. By using official SDKs and avoiding custom authentication, the project aims to provide access to subsidized AI compute while maintaining compliance. However, the broader issue of unclear AI vendor policies continues to create uncertainty for the developer community, highlighting the need for clearer technical guidelines and consistent enforcement standards. ## 🧱 Refactor Appendix ### 📚 Literature Split Pad Purpose: Stage CEE units that become Literature notes. Split level: H3 ### [H1] Official SDK usage provides compliance protection over custom implementations > [!quote] [I need you guys to trust me on this (sorry Anthropic)](https://youtu.be/RIkSlHgQYog?t=390) by [[Theo - t3․gg]] via [[YouTube]]. Published 2024-12-19. Accessed 2024-12-19. > "We're using the official SDK. Open code's using their own harness, which means that they don't have the built-in authentication methods for claude code." T3 Code's architectural choice to use official channels rather than custom authentication flows aims to maintain vendor compliance while providing third-party integration capabilities. [view highlight](https://youtu.be/RIkSlHgQYog?t=390#H1) #xref #articleseed ID:rw_claude_sdk_compliance ### [H2] AI subscription subsidization creates economic pressure for creative integrations > [!quote] [I need you guys to trust me on this (sorry Anthropic)](https://youtu.be/RIkSlHgQYog?t=555) > "Anthropic's Cloud Code sub allows you to use up to $5,000 in compute in a month when you're only paying $200." The massive price differential between subscriptions and API access drives developers to find compliant ways to access subsidized compute through alternative interfaces. [view highlight](https://youtu.be/RIkSlHgQY --- ## Full Transcript [00:00] I have a very exciting announcement to [00:01] make. T3 Code now supports your Claude [00:04] Code subscriptions. If you already have [00:06] Claude Code on your machine, all you [00:07] have to do is run the npxt3 command or [00:09] download the T3 Code app and now you can [00:11] use one of the best coding UIs with your [00:13] existing Cloud Code subscription for [00:15] free in an open- source program. If [00:17] you've been around for a bit, this might [00:18] be raising some red flags in your head [00:20] because, as you may have seen on [00:22] Twitter, this has not been the case for [00:24] other open-source coding tools like Open [00:27] Code. Open Code 1.3.0 will no longer [00:30] autoload the Claude Max plugin. We did [00:32] our best to convince Anthropic to [00:33] support developer choice, but they sent [00:35] lawyers instead. It's your right to [00:37] access services however you wish, but [00:38] it's also their right to block whoever [00:40] they want. This specifically means that [00:41] while they can't enforce that open code [00:44] remove these things, they can absolutely [00:46] threaten to ban Open Code users if they [00:49] continue to support things this way. So, [00:51] what's going on here? If you use this [00:53] button and you use your Cloud Code sub [00:55] in T3 code, are you going to get banned? [00:58] Is this going to ruin your subscription? [01:00] Are you going to get into trouble? Can [01:01] Anthropic sue you? What's going on here? [01:04] Why have other places not been able to [01:06] integrate Claude? And we were. I have a [01:08] lot to tell you about this. But as you [01:10] guys might know by now, T3 Code does not [01:13] make us any money. I'm effectively [01:14] funding multiple developers working on [01:16] this full-time to give it out for free [01:18] as an open source app. Hopefully someday [01:20] we could find some way to sell a service [01:22] alongside it that makes money, but right [01:23] now we have no way to make a single [01:25] dollar off of T3 Code. and I got bills [01:27] to pay. So, we're going to do a quick [01:28] break for today's sponsor. You've all [01:30] heard the line, "Time is money." This [01:31] usually implies you have to trade one [01:33] for the other. What if I told you you [01:35] didn't have to? What if I told you there [01:37] was an incredible product that will save [01:38] you time and money? Good news, there is. [01:41] It's called Blacksmith, and it's made by [01:42] development experience significantly [01:44] better. It seems insane that you can [01:46] change one line of code and cut your [01:48] costs in half and your runtime in half. [01:50] And those stack, by the way, to a 75% [01:53] savings for most projects. Blacksmith is [01:55] the best way to be running your GitHub [01:57] actions. If you're not already using [01:58] them, it's so easy to switch. I implore [02:00] that you try because it's it's magical. [02:02] Our build times for T3 Code got cut from [02:04] 2 and 1/2 minutes to under a minute. Our [02:06] build times for T3 Chat got cut from [02:08] around 7 minutes to around three. Their [02:10] hardware is twice as fast cuz they're [02:11] using good modern gaming CPUs. So, the [02:14] single threaded performance is way [02:15] better, which is huge for most of our [02:16] workloads. The cache downloads are four [02:18] times faster because they actually have [02:20] them stored in a reasonable way. And [02:22] that goes even further with Docker [02:23] builds, which are up to 40 times faster [02:25] because their NVME drives cach your [02:27] layers throughout. And don't get me [02:28] started on the dashboard because this [02:30] has been one of the most life-changing [02:32] parts of the product. Actually being [02:33] able to see what's working and what's [02:35] failing with your builds, your [02:36] pipelines, your CI, and more. Especially [02:39] for a project like T3 Code where we have [02:40] hundreds of developers trying to [02:42] contribute. Now, seeing where failure [02:43] rates are going up and down is really [02:45] useful as we try to make our tests and [02:47] everything work how they should. I'm [02:49] going to be honest with y'all for a [02:50] second. I didn't set up Blacksmith on T3 [02:52] Code initially because the project was [02:54] so simple I didn't think we need it. [02:55] Then we had a regression in CI. Couldn't [02:57] track it at all and I missed the [02:59] dashboard so much that I went into it to [03:01] set it up. It took two clicks and I [03:03] immediately had faster builds and way [03:05] better insights onto what's going on. [03:06] You're currently running slowly in the [03:08] dark. Get yourself a Ferrari with [03:09] headlights for cheap at soy. Blacksmith. [03:12] Okay, back to the matter at hand. How [03:14] did we get a claude code subscription [03:16] button that works in T3 Code in a way [03:18] that supposedly won't get people banned? [03:21] Before we do that though, I need to be [03:22] very explicit. This video is not legal [03:24] advice. I am not responsible for [03:26] Anthropic's actions and what they do to [03:28] your account. If you do get banned by [03:29] using T3 Code, I'm incredibly sorry and [03:32] I want you to tell me every single thing [03:33] about that ban so we can make sure it [03:35] doesn't happen to others. But I cannot [03:36] control what Anthropic does and their [03:38] ability to ban you for anything they [03:40] feel like at any time is not something I [03:42] can make any guarantees against. That [03:43] all said, it is my understanding that [03:46] you should not get banned for using T3 [03:48] code as the user interface for working [03:50] with Claude Code. And this isn't just my [03:53] understanding. This is also Claude's [03:55] understanding because I asked Claude [03:57] code to take a look at the project, take [03:58] a look at the pull request where poor [04:00] DAX has to actually comply with [04:02] Anthropic's legal requests here and also [04:05] look at the terms of service from [04:06] Anthropic and to say are we within the [04:09] terms of service. Is there any reason [04:10] for us to be worried? And thankfully [04:12] Claude's on my side. I am not sure how [04:14] good of a lawyer Claude would be if this [04:16] actually goes to court. But there's a [04:17] very important critical distinction [04:19] here. Unless previously approved, [04:21] Anthropic does not allow third-party [04:23] devs to offer Claude.AI login or rate [04:27] limits for their products, including [04:29] agents built on top of the Claude agent [04:31] SDK. Please use the API key [04:33] authentication methods described in the [04:35] document instead. That sounds like we're [04:37] damned, right? I don't think so. Does [04:39] not allow third party devs to offer [04:41] Claude AI login or rate limits for their [04:44] products. We're not offering either of [04:46] those things. In the comparison that [04:48] again claude made with T3 code versus [04:50] open code, we are using the official [04:52] SDK. Open code's using their own [04:54] harness, which means that they don't [04:56] have the built-in authentication methods [04:58] for claude code. And anthropic will deny [05:01] API requests that are part of your [05:03] Claude sub if they are not coming from [05:05] their harness because the point of the [05:07] sub is that you use Claude's surfaces, [05:09] official Anthropic surfaces with it. And [05:11] they don't want people using different [05:12] harnesses for a bunch of reasons. They [05:14] claim it's because they can't make a [05:16] good quality experience with these other [05:17] harnesses. It would be a lot easier for [05:19] people to abuse them and use these other [05:22] subs in places that will do way more [05:24] inference per request. Like if these [05:26] other harnesses don't implement caching [05:28] correctly, they don't implement tools [05:30] for finding context in your codebase [05:31] correctly and they do things wrong, you [05:33] could theoretically have a worse [05:35] experience using these models and using [05:37] your sub on a different thing. And [05:40] Enthropic really wants to prevent that. [05:41] At least that's what they claim. I think [05:42] that their interests go way beyond that [05:44] because I'm not very confident in [05:46] anthropic as a business. But their [05:48] distinction as far as it's been [05:49] described to me both online and offline [05:51] is that they don't want other harnesses [05:53] and other code to be executed based on [05:56] the model's behaviors using the [05:58] subscription that they charge for. But [06:00] in order for Open Code to do that, they [06:02] had to fake headers. They had to provide [06:04] their own O flow for users to go sign in [06:06] to get their token and whatnot. It would [06:08] then use their subscriptionbased tokens [06:10] and they would be managing the O. With [06:12] us, we're using the agent SDK to [06:14] effectively access your local cloud code [06:18] CLI. There is no O. There is no button [06:20] anywhere in T3 code where you can set up [06:21] O or API keys or anything. You have to [06:24] set up the CLI and we are effectively [06:26] just calling the CLI on top. It is also [06:28] worth noting that what Open Code was [06:30] doing was totally fine for quite a while [06:32] and it is totally fine for other [06:34] integrations like their GitHub copilot [06:36] integration or their codec subscription [06:38] integration. All of the other labs are [06:40] taking the opportunity to help out Open [06:42] Code. Anthropic is seemingly trying to [06:44] destroy them. So why are people still [06:46] using these subscriptions? Why not just [06:47] move to a sub with Open Code? You can [06:49] get $200 a month there and have a ton of [06:51] stuff. You're not going to have as much. [06:53] I've mentioned this at other videos, but [06:54] I think it's more important now than [06:55] ever to mention that Anthropic's Cloud [06:58] Code sub allows you to use up to $5,000 [07:00] in compute in a month when you're only [07:02] paying $200. Anthropic is massively [07:06] subsidizing the amount of inference you [07:08] can do because they can. They want to [07:10] crush the competition. They want [07:11] everyone to have these high tier subs [07:13] and just use anthropic models. And we [07:15] can't compete with that. Nobody can [07:16] compete with that. Most users now expect [07:18] this level of subsidization. So if [07:20] they're used to doing that much compute [07:21] for 200 bucks and they move over to your [07:23] subscription instead because they hate [07:24] Entropic so much and they get to do [07:26] onetenth to 125th as much inference then [07:30] good luck. You're not going to keep that [07:32] user. They're going to go back. That's [07:33] why we chose to integrate into T3 code [07:35] the way we did because I don't have [07:37] confidence that I could charge you money [07:39] when other labs can charge you way less [07:41] and give you way more. Subsidization is [07:44] a really tough pill to swallow in the [07:46] space that makes even things like T3 [07:48] chat scary for us to keep investing in [07:50] because the labs can make it nearly [07:52] impossible to host by making their subs [07:54] more generous and making the APIs more [07:55] expensive. All of this said, I've seen [07:58] so many people who are genuinely [08:00] concerned and also confused. Will using [08:02] Claude code in T3 code get you banned? [08:05] Will using cla code and T3 code put your [08:07] account at risk? Are we going to get [08:08] legal notices from Anthropic about the [08:10] way we integrated this? Can other [08:12] businesses and other products integrate [08:14] in similar ways? Is this okay if it's [08:15] closed source or only open source? Can I [08:17] host this in the cloud or does it have [08:18] to be on my own machine? All these [08:20] questions and more have not been [08:21] answered well. And while I do mostly [08:24] have confidence in my answer for a few [08:26] of those things, it's nearly impossible [08:28] to know. And I am far from the only [08:30] person who feels this way. I have so [08:32] many incredible people who have been [08:33] DMing me like hi at people that you [08:35] would never guess hitting me up asking [08:37] if I have any insights on the real [08:39] answers to these questions because [08:40] they're not saying it at all. Even Matt [08:42] PCO, who you guys might know, the [08:43] Typescript guy who is now also the AI [08:46] guy, has been having the same problem. [08:48] He posted, "Can I get some questions [08:49] answered by someone? First, can I use an [08:52] OOTH token generated from a sub to power [08:54] the cloud agent SDK strictly for using [08:56] cloud code in a local dev loop? All I [08:58] want is a more reliable API for [09:00] paralyzing multiple cloud codes. And [09:02] two, if he builds an open- source tool [09:03] that relies on this pattern, like for [09:05] making parallelization easier, can he [09:06] distribute that so others can use it? [09:08] The reason I'm asking is the legal [09:10] compliance docs and Thoric's public [09:12] statements that seem to contradict each [09:14] other where previously Thoric had said, [09:15] "We want to encourage local dev and [09:17] experimentation with the Asian SDK and [09:19] cloud-p. If you're building a business [09:21] on top of the Asian SDK, you should use [09:22] an API key instead." Not very clear. [09:25] Matt was trying to relieve his confusion [09:26] here and assumed as somebody who just [09:28] asked nicely that they would get an [09:29] answer. Clearly not familiar with how [09:31] Anthropic does things. Again, as always, [09:33] no shade to Thoric. He has been put [09:35] between a rock and a hard place here [09:37] where he clearly wants to do right to [09:39] the community, but he's not allowed to [09:40] answer the important hard questions [09:42] here. Sorry, this has been confusing. I [09:44] know we should be clearer here. The [09:45] Asian SDK spans a lot of use cases, and [09:47] that makes it tough to give guidance. [09:49] We're working on more clarity for this, [09:51] and Thor will follow up. Thanks for [09:52] bearing with us. It's been a hectic [09:54] couple weeks. Unlike many folks here, [09:56] I'm not assuming malicious intent. Take [09:57] your time and give us a clear answer so [09:58] we can put all the fear, uncertainty, [10:00] and doubt to bed. And this is where Theo [10:02] hops in. I assumed the same, but the [10:04] amount of time and opportunity they've [10:05] been given to fix this is genuinely [10:07] absurd. And the lack of any good faith [10:09] action on their part is unacceptable. It [10:11] is also worth noting, as chat has just [10:13] noticed, this was over a week ago, and [10:15] we still have heard nothing. I thought [10:17] AI meant people moved faster. [10:19] Apparently, Claude's a good lawyer. [10:20] Maybe they're struggling to get Claude [10:21] to agree with them because whenever I [10:23] asked Claude about this, it agrees with [10:25] me. I bet if Claude agreed with them, [10:27] they would have had this fixed a lot [10:28] faster. We always strive to do right by [10:30] all of our users, but we've seen [10:32] incredible growth since January, and as [10:33] you both know, it's very difficult to [10:35] scale a team fast enough to match [10:36] growth. We have users using AI for the [10:38] first time on Cloud AI, people learning [10:39] to code using cloud code, enterprises [10:41] switching legacy workflows, and power [10:43] users using the agent SDK, amongst many [10:45] others. People at Anthropic are often [10:47] across many of these user archetypes at [10:49] once and there's been a lot of [10:50] challenges dealing with this, especially [10:52] with the crazy recent growth. We've not [10:54] done fully right by Asian SDK users. I [10:56] agree, but it's not for lack of trying [10:57] or wanting and I cannot fault anyone in [10:59] anthropic for it. That said, I always [11:01] admire that Theo and people like him [11:02] hold us to high expectations and I agree [11:04] that we should meet them. We can always [11:06] be better. Seems like progress, right? [11:08] Well, it's been over a week and Matt [11:10] doesn't want to keep waiting. Anthropic [11:13] AI bump. Working on it. And then the [11:15] single most terrifying reply I think [11:18] I've ever seen on Twitter. ETA. If [11:21] someone did this to me on a thing that [11:23] had been stressing me out as much as I [11:24] am certain this Claude code drama has [11:27] been stressing out poor Thoric, I would [11:28] quit on the [ __ ] spot. Like I cannot [11:31] imagine what went through this poor [11:32] guy's head when Matt PCO, the nice guy [11:35] in the batch, the one who is the kindest [11:38] in this whole space about all of this [11:40] stuff and one of the few to take the [11:42] time to engage with Anthropic in a kind, [11:44] thoughtful way about this is now past [11:47] his patience. This is all absurdity. [11:49] Like there's no excuse for Anthropic to [11:51] not have clarified this. And I see [11:53] somebody in chat saying something stupid [11:55] like crazy that you expect as an [11:56] individual that a multi-billion dollar [11:58] company is going to do whatever you [11:59] want. No, that's not what we're saying. [12:02] What we're saying is that we have to [12:03] cover these things as journalists, as [12:05] influencers, as whatever you want to [12:07] call us, educators, whatever [ __ ] [12:09] posters. We need to be able to talk [12:10] about these things. We need to be able [12:11] to release the things we're working on. [12:13] And we're all stuck in this weird [12:14] uncertain space that I've never [12:16] experienced anything like before. I've [12:18] never had a company that is so prolific [12:20] and so active in the community I am in [12:23] that is so unwilling to answer basic [12:25] questions that are blocking us every [12:26] day. It's insane. Anthropic just isn't [12:29] participating when this is the topic. [12:31] All of this said, it is my understanding [12:34] from all of the conversations I've had, [12:35] all of the reading I've done, and how we [12:37] implemented this that we're in the [12:39] clear. We specifically built T3 code in [12:41] such a way to make it as likely as [12:43] possible that we are complicit with [12:45] these subsidized token options. We [12:47] abstracted the layer that manages all of [12:49] our events so that we can have different [12:51] models and different providers and most [12:53] importantly different harnesses [12:55] supported inside of T3 code. We didn't [12:57] make the harness. We didn't make the box [12:58] that provides the tools the models use [13:00] to edit files and all of that. We're [13:02] using the CLIs and it honestly was a [13:04] little bit more work to do things that [13:06] way because all the CLI have their own [13:07] weird ways of exposing information and [13:09] communicating. In order to integrate [13:11] those, we had to build our own [13:12] abstraction to filter through those [13:15] layers. It's almost like we have to [13:16] unbake a cake in order to reassemble it [13:18] on the other side. But we did it. We put [13:20] a lot of work into it and I still don't [13:22] know if we are fully safe. Again, I'm [13:24] not a lawyer. This is not legal advice. [13:26] I cannot be held responsible if your [13:27] account gets banned. If it does, I want [13:29] to help. But it is my understanding that [13:31] this is entirely clean and fully safe to [13:35] use. I would be very, very surprised if [13:38] Anthropic was to do anything here. And [13:40] if they do actually get your account [13:42] banned because you use T3 code and [13:44] you're able to prove that, I will cover [13:46] your sub for permission to use your name [13:48] in the video I will film calling them [13:50] the [ __ ] out for this. But if your risk [13:52] tolerance is a little bit higher than [13:54] mine, there are options so you can keep [13:56] using things like Open Code. There's a [13:58] fork of the Open Code Claude O plugin [14:00] that is still being maintained that lets [14:02] you use your Claude Code credentials [14:03] inside of Open Code. I would expect this [14:05] GitHub repo will stay up, but I would [14:07] also not be surprised if Anthropic [14:09] starts cracking down more on things like [14:11] this. They've already hard-coded a [14:13] change in their own API where if open [14:16] code is mentioned in the headers at all, [14:17] the request gets instenied for off [14:20] through their subscriptions to be clear. [14:22] But still, this is a thing that I would [14:23] not be surprised if it gets your account [14:25] banned. T3 code is a thing that I would [14:27] be incredibly surprised if it gets your [14:29] account banned. That all said, I do [14:31] still have my concerns. At any point [14:32] this goalpost can move because Anthropic [14:34] loves to do that. They're keeping things [14:36] vague so they can change their mind at [14:38] any point and make different and [14:39] potentially worse policies for us. They [14:42] also want to act things like openclaw [14:44] and keep those from using up people's [14:45] subs. This could change for the better. [14:47] They could be more open in the future or [14:48] it could change for the worse and they [14:50] become more aggressive. I'm not very [14:52] high faith. I am scared that there is a [14:54] future where we are forced to deprecate [14:56] the cloud code support which would suck [14:58] for a bunch of reasons. most notably [14:59] because you can't switch out of cloud [15:01] code once you've picked it for a thread [15:02] because the threads, the compaction, the [15:05] knowledge and thinking data that's all [15:06] encrypted and hidden in the cloud, all [15:08] of that is linked to different threads. [15:10] So once you've picked which model you [15:11] want to use for a thread, you can't [15:12] switch out of that harness. Annoying, I [15:15] know, I hate it, but it's all we're able [15:17] to do because all of these things are [15:18] now more and more hidden and locked away [15:20] from us as the developers building. [15:22] We've done our best and I think we did a [15:23] really good job. I've had a ton of fun [15:25] playing with T3 code and playing with [15:27] codeex in claude code in this way. I [15:29] still personally find that 5.4 is the [15:31] best model for coding most of the time. [15:32] So I will use that for the majority of [15:34] tasks and then I'll quickly just open a [15:36] new thread and switch over to Claude [15:37] when I want to use that for a UI pass, [15:39] quick tidy ups, small changes, things [15:41] like that. I live a very multimodel life [15:43] where I use the $200 month sub from [15:45] Codeex and Claude and use a meaningful [15:47] amount of both for the work that I'm [15:48] doing every day. And if you're like me, [15:50] you have all these different subs and [15:51] you want a UI that unifies all of them [15:54] while still using the subsidized tokens [15:56] that you're getting from those labs. As [15:57] far as I have seen, T3 code is the best. [16:00] I hope other people keep building cool [16:01] things. They keep forking T3 Code and [16:02] keep doing cool stuff with it because [16:04] that's why we made it open source. We [16:05] want to make the tools as a whole [16:07] better. Shout out to my whole team for [16:08] helping build T3 code the way we did. [16:10] Huge shout out to Julius for the insane [16:12] work he has done making this all [16:14] possible. He grinded it hard to build [16:16] that new event system and also make sure [16:17] things worked well with Claude and [16:20] that's why it works. Also a shitload of [16:22] effect code. Very happy with where we've [16:24] landed. Relatively confident our [16:26] implementation won't get you banned and [16:28] more importantly will let you keep [16:30] taking advantage of those subsidized [16:31] tokens in a UI that's way better and [16:33] won't leak memory all over your [16:35] computer. It is quite funny that T3 Code [16:37] uses less resources than the Cloud Code [16:39] CLI does, but you get the idea. We built [16:42] something that we wanted to use. I'm [16:43] very proud of what we built and as far [16:44] as we understand, you should be safe. Go [16:47] give it a shot yourself. Let me know [16:48] what you think and let me know if you [16:49] get banned because you really, really [16:51] shouldn't. And Enthropic, please just [16:53] just give us the answer here. I [16:54] shouldn't have to keep speculating this [16:56] way. I've done more content about your [16:58] policy than anybody at your company has, [16:59] than everyone at your company combined [17:01] has. Just get us out of the [ __ ] dark [17:03] here. It's not that hard. Just tell us [17:05] what you think. [17:07] Anyways, bye nerds.  📄 **Full text:** ![[Full Document Contents/Articles/I Need You Guys to Trust Me on This|I Need You Guys to Trust Me on This, YouTube]] ## Highlights - I want you to tell me every single thing about that ban so we can make sure it doesn't happen to others. But I cannot control what Anthropic does and their ability to ban you for anything they feel like at any time is not something I can make any guarantees against. That all said, it is my understanding that you should not get banned for using T3 code as the user interface for working with Claude Code. And this isn't just my understanding. This is also Claude's understanding because I asked Claude code to take a look at the project, take a look at the pull request where poor ([View Highlight](https://read.readwise.io/read/01kmg33jvd9dn67rz6vxpmq8ss)) - Tags: 👻 ai highlighted - SDK. Please use the API key authentication methods described in the document instead. That sounds like we're damned, right? I don't think so. Does not allow third party devs to offer Claude AI login or rate limits for their products. We're not offering either of those things. In the comparison that again claude made with T3 code versus open code, we are using the official SDK. Open code's using their own harness, which means that they don't have the built-in authentication methods for claude code. And anthropic will deny ([View Highlight](https://read.readwise.io/read/01kmg33jvzj95vq0yx4qrjpgrd)) - Tags: 👻 ai highlighted - things. They keep forking T3 Code and keep doing cool stuff with it because that's why we made it open source. We want to make the tools as a whole better. Shout out to my whole team for helping build T3 code the way we did. Huge shout out to Julius for the insane work he has done making this all possible. He grinded it hard to build that new event system and also make sure things worked well with Claude and that's why it works. Also a shitload of effect code. Very happy with where we've landed. Relatively confident our implementation won't get you banned and more importantly will let you keep ([View Highlight](https://read.readwise.io/read/01kmg33k0k84033p3kzt0gqwvx)) - Tags: 👻 ai highlighted