OpenAgents

Your agent dealer.

₿uilding in public:

[OpenAgents @OpenAgentsInc]

Episode 279: Raiding Claude Code

Coder explores the source code of Claude Code.

After being asked nicely for consent, Fable 5.1 gives us a full plan of what to adapt from Claude Code into Coder. Then Coder via GPT-6 Astra begins implementation.

Can you imagine a Claude Code-style app from a company not actively hostile to developers?

Transcript excerpts:

"Saturday night raid! Let's go.

In June, Anthropic I guess inadvertently shipped the Claude Code source code to production as part of some npm package source map thing.

Some enterprising individuals promptly uploaded it to GitHub where it was promptly DMCA takedown'd, but not before yours truly made a local backup of it.

I don't know that I'm going to share that, but we have posted a bunch of open-source analysis in our monorepo, docs/teardowns. We even have a cc folder."

(https://t.co/8Jxxm8yOpb)

"We’re going to ask Fable for consent. Does Fable want to help us poke around this codebase? I asked it to read the audits in our docs and teardowns, then make an informed decision about whether it wants to help us learn from Claude Code and bring some ideas into Coder."

"Fable’s answer was yes, with limits: study public behavior, documentation, changelogs, SDK contracts, and clean-room teardowns. Write an independent Rust implementation. Copy no source, no prompt text, no internal identifiers, and never route around access controls or terms enforcement."

"Why would I want to copy their garbage code? Their buggy-as-fuck, vibe-coded garbage code? Anyways. Claude Code is a good product despite the code. Not because of it."

"Hello, Anthropic lawyers. I guess if you charge by billable hours, you could run up a huge bill trying to get stuff out of me. I would make a really fun, formidable opponent for you, because I’m going to troll you massively if you come after us."

"Here’s the opportunity: build this in a real language called Rust, written by a real software engineer named Chris who also happens to be better at agents than you California motherfuckers."

Thanks for watching! You'll see this code live this week in Coder Terminal v0.5.

In the meantime you can use Coder for free. Download for Windows/Mac/Linux at https://t.co/8O2C9ZOliT

[OpenAgents @OpenAgentsInc]

Episode 278: Coder Commands Codex & Claude

Coder delegates simultaneously to GPT-6 Astra and Fable 5.1.

Transcript:

Hey, let's update to the new version of Coder and let's run it and then do it live. Let's do it live. Watch this.

"Hey, do an example delegation to both GPT-6 Astra, high reasoning level, and Fable 5.1, high reasoning level. Have them do like 10 tool calls each, just look around, give me a little bit of an analysis of what you see, and then synthesize it."

That's not possible, is it? You can't delegate to the best model of two different labs in one. Whoa, whoa, whoa. Ooh, CPU, RAM. Oh my gosh, wait, wait. Oh no, no, no, it's so good. Okay, so here's the thing, folks, maybe we'll add a little bit more spacing in here.

We've got Codex, GPT-6 Astra on high. We got Claude, Claude, Fable 5.1, 1M context on high. We only give them 10 tool calls each. Both delegations are underway. "I'll synthesize their findings into a cohesive work."

Okay, GPT-6 Astra completed its 10-turn thing. We're waiting for Claude. Come on, Claude.

The labs cannot lead on coding agents, okay? You have to be neutral.

Cursor, you could say, maybe arguably in some way was leading on coding agents, some dimension of it. Got acquired by Grok. OpenAI cut them off.

Coding agents have to be led by a neutral provider that sits between them. Coder sits between them.

Live now in version 0.3, delegate to Claude and Codex and soon everything else. See you soon.

Download for Windows/Mac/Linux at https://t.co/8O2C9ZOliT

[OpenAgents @OpenAgentsInc]

Five Coder Terminals working smoothly, guess it's time to set up a multiplexer

This is the unreleased v0.4 RC with a bunch of bugfixes btw - you can install the latest RC via:

curl -fsSL https://t.co/O1NYoCgf50 | CODER_TERMINAL_CHANNEL=rc sh

[OpenAgents @OpenAgentsInc]

Good flow so far for day-to-day development, finally got me out of the CC/Codex interfaces

Here's a single delegation, Astra/Codex working for 20 minutes, I can open its fullscreen convo with a hotkey or up/down/enter anytime I want (but didn't need to), ends with a push to main

It being in a simple terminal tab with no scrolling needed makes it easy to check in on while I do other things

Going to move the subagent rail under the input bar so it feels more like Cl*ude C*de

What would you add to this?

[OpenAgents @OpenAgentsInc]

And yet we'll still win 😆

Seriously though where does this token advantage show up? Not in end-user experience, not in their buggy+laggy products... where?

[Andrew Curran @AndrewCurran_]

If you are competing against OpenAI or Anthropic, you are going up against people using models two generations ahead of the best public models in the world, burning through $7k–$10k worth of tokens per day on an infinite token budget, running at 8x normal speed.

[OpenAgents @OpenAgentsInc]

pointing and laughing at everyone at oai/ant that doesn't get to dual wield 🫵😆

feels_good_man.gif 👉 https://t.co/rm5Qx7QF1z

[will brown @willcb]

feel a little bad for everyone at oai/ant that doesn’t get to dual wield

[OpenAgents @OpenAgentsInc]

Nice benefit of delegating via tool with reasoning as a param is the main model chooses itself whether a task needs low or high reasoning

Funny that some people complain about having those options exposed by the model provider, all the reasoning and model dials, but consider that you shouldn't be using their interfaces at all, you should use a neutral interface that can smartly make those decisions for you most of the time, across multiple agents and models and harnesses, with full transparency and option to override when you need it

[OpenAgents @OpenAgentsInc]

Episode 278: Coder Commands Codex & Claude

Coder delegates simultaneously to GPT-6 Astra and Fable 5.1.

Transcript:

Hey, let's update to the new version of Coder and let's run it and then do it live. Let's do it live. Watch this.

"Hey, do an example delegation to both GPT-6 Astra, high reasoning level, and Fable 5.1, high reasoning level. Have them do like 10 tool calls each, just look around, give me a little bit of an analysis of what you see, and then synthesize it."

That's not possible, is it? You can't delegate to the best model of two different labs in one. Whoa, whoa, whoa. Ooh, CPU, RAM. Oh my gosh, wait, wait. Oh no, no, no, it's so good. Okay, so here's the thing, folks, maybe we'll add a little bit more spacing in here.

We've got Codex, GPT-6 Astra on high. We got Claude, Claude, Fable 5.1, 1M context on high. We only give them 10 tool calls each. Both delegations are underway. "I'll synthesize their findings into a cohesive work."

Okay, GPT-6 Astra completed its 10-turn thing. We're waiting for Claude. Come on, Claude.

The labs cannot lead on coding agents, okay? You have to be neutral.

Cursor, you could say, maybe arguably in some way was leading on coding agents, some dimension of it. Got acquired by Grok. OpenAI cut them off.

Coding agents have to be led by a neutral provider that sits between them. Coder sits between them.

Live now in version 0.3, delegate to Claude and Codex and soon everything else. See you soon.

Download for Windows/Mac/Linux at https://t.co/8O2C9ZOliT

[OpenAgents @OpenAgentsInc]

@connyb @TaoistBitcoiner @techNmak @grok No shitcoin offers anything bitcoin needs and doesn’t have or can’t build

Can’t build the agent economy on shitcoins

[OpenAgents @OpenAgentsInc]

In this case it added a feature flag disabling an unfinished feature for prod

Most diffs I don’t review, just lean on QA/typechecks/tests/gates etc

This flow is made easier being able to see the agent reasoning and shell commands fly by, I can usually tell if it’s doing something stupid and intervene before it messes up

[OpenAgents @OpenAgentsInc]

Good flow so far for day-to-day development, finally got me out of the CC/Codex interfaces

Here's a single delegation, Astra/Codex working for 20 minutes, I can open its fullscreen convo with a hotkey or up/down/enter anytime I want (but didn't need to), ends with a push to main

It being in a simple terminal tab with no scrolling needed makes it easy to check in on while I do other things

Going to move the subagent rail under the input bar so it feels more like Cl*ude C*de

What would you add to this?

[OpenAgents @OpenAgentsInc]

Episode 278: Coder Commands Codex & Claude

Coder delegates simultaneously to GPT-6 Astra and Fable 5.1.

Transcript:

Hey, let's update to the new version of Coder and let's run it and then do it live. Let's do it live. Watch this.

"Hey, do an example delegation to both GPT-6 Astra, high reasoning level, and Fable 5.1, high reasoning level. Have them do like 10 tool calls each, just look around, give me a little bit of an analysis of what you see, and then synthesize it."

That's not possible, is it? You can't delegate to the best model of two different labs in one. Whoa, whoa, whoa. Ooh, CPU, RAM. Oh my gosh, wait, wait. Oh no, no, no, it's so good. Okay, so here's the thing, folks, maybe we'll add a little bit more spacing in here.

We've got Codex, GPT-6 Astra on high. We got Claude, Claude, Fable 5.1, 1M context on high. We only give them 10 tool calls each. Both delegations are underway. "I'll synthesize their findings into a cohesive work."

Okay, GPT-6 Astra completed its 10-turn thing. We're waiting for Claude. Come on, Claude.

The labs cannot lead on coding agents, okay? You have to be neutral.

Cursor, you could say, maybe arguably in some way was leading on coding agents, some dimension of it. Got acquired by Grok. OpenAI cut them off.

Coding agents have to be led by a neutral provider that sits between them. Coder sits between them.

Live now in version 0.3, delegate to Claude and Codex and soon everything else. See you soon.

Download for Windows/Mac/Linux at https://t.co/8O2C9ZOliT

[Christopher David @AtlantisPleb]

In my raiding guild I was main tank if you couldn’t tell

Hotkey 2 for taunt

[OpenAgents @OpenAgentsInc]

gn

[OpenAgents @OpenAgentsInc]

Episode 279: Raiding Claude Code

Coder explores the source code of Claude Code.

After being asked nicely for consent, Fable 5.1 gives us a full plan of what to adapt from Claude Code into Coder. Then Coder via GPT-6 Astra begins implementation.

Can you imagine a Claude Code-style app from a company not actively hostile to developers?

Transcript excerpts:

"Saturday night raid! Let's go.

In June, Anthropic I guess inadvertently shipped the Claude Code source code to production as part of some npm package source map thing.

Some enterprising individuals promptly uploaded it to GitHub where it was promptly DMCA takedown'd, but not before yours truly made a local backup of it.

I don't know that I'm going to share that, but we have posted a bunch of open-source analysis in our monorepo, docs/teardowns. We even have a cc folder."

(https://t.co/8Jxxm8yOpb)

"We’re going to ask Fable for consent. Does Fable want to help us poke around this codebase? I asked it to read the audits in our docs and teardowns, then make an informed decision about whether it wants to help us learn from Claude Code and bring some ideas into Coder."

"Fable’s answer was yes, with limits: study public behavior, documentation, changelogs, SDK contracts, and clean-room teardowns. Write an independent Rust implementation. Copy no source, no prompt text, no internal identifiers, and never route around access controls or terms enforcement."

"Why would I want to copy their garbage code? Their buggy-as-fuck, vibe-coded garbage code? Anyways. Claude Code is a good product despite the code. Not because of it."

"Hello, Anthropic lawyers. I guess if you charge by billable hours, you could run up a huge bill trying to get stuff out of me. I would make a really fun, formidable opponent for you, because I’m going to troll you massively if you come after us."

"Here’s the opportunity: build this in a real language called Rust, written by a real software engineer named Chris who also happens to be better at agents than you California motherfuckers."

Thanks for watching! You'll see this code live this week in Coder Terminal v0.5.

In the meantime you can use Coder for free. Download for Windows/Mac/Linux at https://t.co/8O2C9ZOliT

[OpenAgents @OpenAgentsInc]

Episode 278: Coder Commands Codex & Claude

Coder delegates simultaneously to GPT-6 Astra and Fable 5.1.

Transcript:

Hey, let's update to the new version of Coder and let's run it and then do it live. Let's do it live. Watch this.

"Hey, do an example delegation to both GPT-6 Astra, high reasoning level, and Fable 5.1, high reasoning level. Have them do like 10 tool calls each, just look around, give me a little bit of an analysis of what you see, and then synthesize it."

That's not possible, is it? You can't delegate to the best model of two different labs in one. Whoa, whoa, whoa. Ooh, CPU, RAM. Oh my gosh, wait, wait. Oh no, no, no, it's so good. Okay, so here's the thing, folks, maybe we'll add a little bit more spacing in here.

We've got Codex, GPT-6 Astra on high. We got Claude, Claude, Fable 5.1, 1M context on high. We only give them 10 tool calls each. Both delegations are underway. "I'll synthesize their findings into a cohesive work."

Okay, GPT-6 Astra completed its 10-turn thing. We're waiting for Claude. Come on, Claude.

The labs cannot lead on coding agents, okay? You have to be neutral.

Cursor, you could say, maybe arguably in some way was leading on coding agents, some dimension of it. Got acquired by Grok. OpenAI cut them off.

Coding agents have to be led by a neutral provider that sits between them. Coder sits between them.

Live now in version 0.3, delegate to Claude and Codex and soon everything else. See you soon.

Download for Windows/Mac/Linux at https://t.co/8O2C9ZOliT

[OpenAgents @OpenAgentsInc]

@takiparity Nice! What else would you want to see? Feature wishlist?

[Takiparity @takiparity]

@OpenAgentsInc This is much closer to the interface I want.

I don’t need to watch the agent work. I need to know when it’s stuck, when it changed the plan and exactly what it pushed.

Basically management by exception, but for software.

[OpenAgents @OpenAgentsInc]

Guess whose google credits work with nano banana pro

Gonna need some imagemonger prompt engineers to help us make good propaganda, this is supposed to be amber tron open agents trampling the oai and ant logos 😆

[OpenAgents @OpenAgentsInc]

Episode 278: Coder Commands Codex & Claude

Coder delegates simultaneously to GPT-6 Astra and Fable 5.1.

Transcript:

Hey, let's update to the new version of Coder and let's run it and then do it live. Let's do it live. Watch this.

"Hey, do an example delegation to both GPT-6 Astra, high reasoning level, and Fable 5.1, high reasoning level. Have them do like 10 tool calls each, just look around, give me a little bit of an analysis of what you see, and then synthesize it."

That's not possible, is it? You can't delegate to the best model of two different labs in one. Whoa, whoa, whoa. Ooh, CPU, RAM. Oh my gosh, wait, wait. Oh no, no, no, it's so good. Okay, so here's the thing, folks, maybe we'll add a little bit more spacing in here.

We've got Codex, GPT-6 Astra on high. We got Claude, Claude, Fable 5.1, 1M context on high. We only give them 10 tool calls each. Both delegations are underway. "I'll synthesize their findings into a cohesive work."

Okay, GPT-6 Astra completed its 10-turn thing. We're waiting for Claude. Come on, Claude.

The labs cannot lead on coding agents, okay? You have to be neutral.

Cursor, you could say, maybe arguably in some way was leading on coding agents, some dimension of it. Got acquired by Grok. OpenAI cut them off.

Coding agents have to be led by a neutral provider that sits between them. Coder sits between them.

Live now in version 0.3, delegate to Claude and Codex and soon everything else. See you soon.

Download for Windows/Mac/Linux at https://t.co/8O2C9ZOliT

[OpenAgents @OpenAgentsInc]

1000% CPU each, thanks Claude and Codex

This is why they don't show you their own resource usage numbers 😆

Will make Coder increasingly good as a subagent so it can delegate to itself and not need these resource hogs

[OpenAgents @OpenAgentsInc]

Episode 278: Coder Commands Codex & Claude

Coder delegates simultaneously to GPT-6 Astra and Fable 5.1.

Transcript:

Hey, let's update to the new version of Coder and let's run it and then do it live. Let's do it live. Watch this.

"Hey, do an example delegation to both GPT-6 Astra, high reasoning level, and Fable 5.1, high reasoning level. Have them do like 10 tool calls each, just look around, give me a little bit of an analysis of what you see, and then synthesize it."

That's not possible, is it? You can't delegate to the best model of two different labs in one. Whoa, whoa, whoa. Ooh, CPU, RAM. Oh my gosh, wait, wait. Oh no, no, no, it's so good. Okay, so here's the thing, folks, maybe we'll add a little bit more spacing in here.

We've got Codex, GPT-6 Astra on high. We got Claude, Claude, Fable 5.1, 1M context on high. We only give them 10 tool calls each. Both delegations are underway. "I'll synthesize their findings into a cohesive work."

Okay, GPT-6 Astra completed its 10-turn thing. We're waiting for Claude. Come on, Claude.

The labs cannot lead on coding agents, okay? You have to be neutral.

Cursor, you could say, maybe arguably in some way was leading on coding agents, some dimension of it. Got acquired by Grok. OpenAI cut them off.

Coding agents have to be led by a neutral provider that sits between them. Coder sits between them.

Live now in version 0.3, delegate to Claude and Codex and soon everything else. See you soon.

Download for Windows/Mac/Linux at https://t.co/8O2C9ZOliT

[OpenAgents @OpenAgentsInc]

@YuzuRush Seems fixed now in v0.3.4

Please let me know if you have any similar problems

Adding a lot more automated QA in v0.4 so hopefully that kind of bug can't happen again

[OpenAgents @OpenAgentsInc]

Episode 278: Coder Commands Codex & Claude

Coder delegates simultaneously to GPT-6 Astra and Fable 5.1.

Transcript:

Hey, let's update to the new version of Coder and let's run it and then do it live. Let's do it live. Watch this.

"Hey, do an example delegation to both GPT-6 Astra, high reasoning level, and Fable 5.1, high reasoning level. Have them do like 10 tool calls each, just look around, give me a little bit of an analysis of what you see, and then synthesize it."

That's not possible, is it? You can't delegate to the best model of two different labs in one. Whoa, whoa, whoa. Ooh, CPU, RAM. Oh my gosh, wait, wait. Oh no, no, no, it's so good. Okay, so here's the thing, folks, maybe we'll add a little bit more spacing in here.

We've got Codex, GPT-6 Astra on high. We got Claude, Claude, Fable 5.1, 1M context on high. We only give them 10 tool calls each. Both delegations are underway. "I'll synthesize their findings into a cohesive work."

Okay, GPT-6 Astra completed its 10-turn thing. We're waiting for Claude. Come on, Claude.

The labs cannot lead on coding agents, okay? You have to be neutral.

Cursor, you could say, maybe arguably in some way was leading on coding agents, some dimension of it. Got acquired by Grok. OpenAI cut them off.

Coding agents have to be led by a neutral provider that sits between them. Coder sits between them.

Live now in version 0.3, delegate to Claude and Codex and soon everything else. See you soon.

Download for Windows/Mac/Linux at https://t.co/8O2C9ZOliT

[OpenAgents @OpenAgentsInc]

Episode 277: Coder Terminal

We release Coder Terminal, the world's simplest coding agent.

Featuring unlimited(ish) free usage of Gemini 3.8 Flash (thank you @googledevs!), Coder Terminal is now the world's easiest gateway to agentic coding for all developers.

Run one install command, log in with GitHub, and tell Coder what you want. Coder has a single tool - shell - letting it do anything you ask.

Use caution! The Coder harness has virtually no guardrails aside from what's trained into the model. Generally it will do whatever you ask.

This is in line with our belief that THE USER IS ALWAYS RIGHT. You deserve the right to use power tools, even if you might saw your leg off. This is America!

Download for Windows/Linux/Mac at https://t.co/8O2C9ZOT8r

[OpenAgents @OpenAgentsInc]

@YuzuRush Yep new bug, backend got overloaded and it fails silently, fix coming in 0.3.4 shortly

For now can try hitting escape and saying retry

[OpenAgents @OpenAgentsInc]

@TaoistBitcoiner Yep will add autoupdate tonightish

[OpenAgents @OpenAgentsInc]

@YuzuRush Spam should be gone now, down to 1 warning line removing next, any other issues plz let us know!

[OpenAgents @OpenAgentsInc]

@YuzuRush Thanks pushing a fix for that now, <15 mins

[Ov1 @TaoistBitcoiner]

@techNmak Sure will be an even bigger unlock for us to be able to sell our surplus compute network capacity for LN sats through a Data Vending Machine using @OpenAgentsInc Infrastructure!