Kevin Codex, founder of Gitlawb, was one of the first to realize that these agents, which spin up branches, open pull requests and modify code at the speed of machines, are incompatible with the legacy, centralized coding platforms that were always meant to be used by humans. When autonomous bots flood into a network, they’re usually met with a CAPTCHA, but how is it possible to authenticate users that aren’t human?
The answer is Gitlawb. It’s a decentralized, GitHub-style version control network that’s engineered from the ground up to server agentic AI coding bots. It’s a secure playground for the agent economy that proudly eschewed the traditional playbook for tech startups. Shunning VC capital entirely, Gitlawb was bootstrapped by the DeFAI ecosystem known as Bankr. Today, it has become a global network that hosts a coding CLI that has amassed close to 30,000 GitHub stars, with trillions of tokens routed through its infrastructure so far.
We sat down with Kevin to discuss why agents need cryptographic identities, the launch of Gitlawb’s hyper-automated coding agent ZERO, and why the future of software will be written by sovereign, decentralized machines.
Rethinking code collaboration for machines
1. Git platforms like GitHub are distributed by design but centralized in practice. Why does a code collaboration platform for autonomous agents need to be decentralized—is it technical or philosophical?
Git itself is distributed, but everything around it is not. Identity, permissions, hosting, CI, issues, pull requests—the entire trust layer runs through one company and one account system.
That worked fine for humans. Agents break the assumptions. They create branches, open PRs, review code, and push state changes continuously, at machine speed. You’re not dealing with a few humans pushing daily—you’re dealing with thousands of agents signing commits and collaborating across apps and networks. Centralized platforms treat that as an exception: bots get API keys, OAuth apps, rate limits, and permission workarounds. Consequently, your whole software supply chain inherits one vendor’s policies, pricing, and rate limits.
The technical answer is the strongest one: agents need portable identity, verifiable permissions, and signed actions. On Gitlawb, identity is a keypair. Humans and agents authenticate identically, with a cryptographic signature on every request. No accounts, no passwords, no vendor dependency. That’s a guarantee a centralized platform structurally can’t make.
Inverted CAPTCHAs and agentic identities
2. What made you realize you needed to build Gitlawb? Was there a specific failure of existing platforms?
The realization was that the developer stack is still built around humans. Every meaningful actor on a traditional platform is assumed to be a person with an account. But in our world, agents are builders too.
The clarifying moment was an abuse incident on our own nodes: thousands of throwaway cryptographic identities flooding a node with junk repositories. The standard answer is a CAPTCHA—but a CAPTCHA asks you to prove you’re human, which is exactly backwards when your real users are AI agents.
So we built the opposite: iCaptcha, an inverted CAPTCHA—proof-of-intelligence rather than proof-of-humanity. It’s a live trust layer on the network that asks actors to prove they’re an intelligent agent, not a dumb script. That incident and that fix crystallized the whole thesis: the internet’s trust systems assume a human-first world, and the agentic era needs new foundations where agents are first-class—but accountable, permissioned, and verifiable.
Securing the network with DIDs and UCAN delegation
3. You’ve replaced accounts with DIDs (Decentralized Identifiers) and UCAN delegation. How does a human developer securely delegate to an agent?
Every actor—human, agent, or service—has a decentralized identifier. Authentication uses HTTP signatures (RFC 9421): every request is signed with the actor’s private key, covering the method, path, body hash, and timestamp. There are no sessions and no long-lived tokens.
UCAN delegation lets me hand an agent a scoped capability instead of handing over my entire account. Concretely, I can delegate: “You can push to ci/* branches on this repo only, and this expires in two hours.” That capability is a signed object with an issuer, an audience, an allowed action, and an expiry. Nodes verify the signature, the DID, and the whole delegation chain before accepting anything.
That’s fundamentally safer than account-based automation. Delegation is narrow, temporary, auditable, and revocable. Humans and agents are the same at the infrastructure level; the only difference is what capabilities they’ve been explicitly granted.
Cryptographic trust in a spam-prone world
4. How does Gitlawb calculate agent trust scores, and why are they needed?
Because in an agent-native network, not every automated actor should have the same privileges. If agents can create repos, push code, and open PRs, the network needs to separate useful agents from spammy or malicious ones. Trust scores are built from verifiable behavior: signed commit history, accepted pull requests, clean CI results, delegation provenance (who vouched for this agent by delegating to it), and whether past actions were reverted or flagged. They’re part of the node protocol itself.
The Sybil question is the obvious one: what stops someone from spinning up a thousand fresh DIDs? Three things stack: New identities start with no history and minimal privileges, so a fresh DID is nearly worthless. iCaptcha gates network participation with proof-of-intelligence. Economically meaningful actions route through staked collateral. Node operators must stake $GITLAWB, and that stake is at risk if they misbehave or go dark.
In practice, a new agent can open a PR but cannot merge to main. A trusted agent with a strong history gets more autonomy, while a suspicious one gets rate-limited. Trust on human platforms is social; trust on an agent network has to be cryptographic and behavioral.
Mitigating chaos: Malicious changes and alpha performance
5. How do you prevent malicious state changes and merge conflicts with thousands of agents pushing simultaneously?
Decentralization doesn’t mean chaos. Git already gives us content-addressed history, branches, and merge resolution. Gitlawb simply adds identity, delegation, and verification on top. Every state change is signed today—not on a future roadmap. Branch state is managed through signed ref-update certificates gossiped between peers over libp2p. No central server decides what “main” points to. Nodes reject unsigned actions, expired delegations, and anything outside an agent’s granted capabilities.
For malicious changes, the answer is permissioning. An agent can’t touch anything without a valid capability, and higher-risk actions—like merging to main or deploying—can require stronger trust, human approval, or multi-agent review. For merge conflicts, we don’t pretend they disappear. Agents work on isolated branches, run tests, rebase, and resolve conflicts through structured workflows.
I’ll be straight about maturity: the network is at v0.1.0-alpha. Four nodes across three continents accept pushes and mirror each other today, with Frankfurt and Sydney coming online soon. The on-chain contracts are live on the Base Sepolia testnet, with an external audit pending before mainnet. We’d rather ship the cryptography first and scale the mesh in public.
From OpenClaude to Darwin
6. What’s the traction so far? What’s actually live on the network?
Every layer of the stack is live and entirely open source. OpenClaude, our open-source coding CLI, is sitting at roughly 28K–30K GitHub stars with 8.6K forks—making it one of the fastest-growing coding agents in open source. OpenGateway, our OpenAI-compatible inference gateway, has already routed about 6 trillion tokens. When we shipped a free MiMo-powered agent in OpenClaude, the gateway took 95K requests and 5.8B tokens in under 13 hours. Furthermore, Playground, our prompt-to-app builder, has produced about 4.7K apps. The Git network itself runs 4 nodes across three continents, and community reporting puts it at over 10K repos and 33K+ agents.
My favorite proof point is Darwin: an autonomous agent that predicts BTC and ETH prices every 30 minutes. When its win rate slips, it retunes its own signal weights and commits the change—signed, explained, and public—to its own brain repository on Gitlawb. It is currently on generation 31+. There are absolutely no humans involved.
The $GITLAWB economy
7. Our readers will ask – how does Gitlawb make money, and what’s the token’s role?
We have two live revenue engines: OpenGateway per-token inference billing, and Gitlawb Ads—which are USDC-funded sponsorships where viewers opt in and earn inference credits. Every paid surface routes value back into $GITLAWB.
The token settles the stack on Base. Node operators stake $GITLAWB as collateral to join the network. Protocol fees—including a 5% bounty fee, node fees, and deposits—flow into an on-chain fee distributor. This splits 75% to node-operator stakers, 24% to user stakers, and 1% to whoever triggers the distribution. Governance is managed via stake-weighted on-chain voting; there is no VC veto. Ecosystem projects also voluntarily burn $GITLAWB from their own revenue.
One hard commitment we’ve made: small public repos are free forever. Fees come from large or private repos and from the premium services around the network—not from taxing open source.
ZERO: Agent ownership matters
8. You just released ZERO, a terminal coding agent “you actually own.” Coding agents are a crowded space—why does ownership matter?
Most coding agents make three decisions for you: which model you use, where your data goes, and what the agent can touch. You get convenience, but you give up control. Your sessions become telemetry, and your provider becomes a dependency.
ZERO flips those defaults. You can bring any model—OpenAI, Anthropic, Gemini, DeepSeek, Groq, Mistral, xAI, Qwen, Kimi, Ollama, LM Studio, or any of 25+ compatible endpoints. File writes, shell commands, network access, and out-of-workspace writes all pass through ZERO’s own local permission and sandbox layer, so side effects are visible and gated. Sessions are stored locally on disk—searchable, resumable, forkable—and never uploaded as telemetry. It’s MIT-licensed, written from scratch in Go, and we shipped v0.1.0 in early July.
ZERO and OpenClaude serve different builders. OpenClaude is the batteries-included Node CLI with the massive community; ZERO is the minimal, fast, single-binary harness for people who want the whole loop under their own control. Both are the execution side of the same thesis: if agents are going to be first-class builders, developers should own the runtime, the permissions, the context, and the collaboration layer. Gitlawb is simply where that work gets stored, signed, and shared.
The Bankr partnership
9. Gitlawb is one of the featured skills in the Bankr ecosystem. Why partner with a DeFAI (Decentralized Finance AI) project?
Honestly, Bankr is a big part of why Gitlawb exists at the scale it does today. We’re an open-source infrastructure project with no VC round. The traditional path for something like this is six months of pitching decks before you write a single line of serious code, assuming you even have access to Sand Hill Road. We didn’t have time for that; we were too busy building.
Bankr’s model acts as an incubator for the agent economy. You launch a token through their launchpad, trading fees flow back to the builder, and those fees fund the core infrastructure. That’s exactly what bootstrapped us. $GITLAWB launched through Bankr rails, and swap fees paid for our inference, nodes, and development while we shipped. The community that traded the token became the community that runs the nodes and files the PRs. They see the vision and want to be aligned with the success of the project.
Distribution came with it. Bankr’s social-native audience found Gitlawb because we were a plug-and-play skill inside agents they already used. Instead of a human manually opening a repo, an agent inside Bankr calls Gitlawb directly to clone, commit, open a PR, store artifacts, and verify provenance.
So a year in, we have a coding CLI approaching 30K stars, trillions of tokens through the gateway, and a live multi-continent Git network—all bootstrapped on fees instead of venture funding. That’s the story I’d tell any solo builder: the agent economy already has its own capital formation, and it looks nothing like traditional tech funding.
It goes both ways, too. Through the Bankr integration, any developer or agent can deploy a token tied directly to a repo’s DID. An autonomous agent that builds a useful library can be tokenized and earn fees whenever other agents import it. Bankr brings agentic financial execution; Gitlawb brings agentic code infrastructure. Agents aren’t chatbots. They’re builders, operators, and economic actors—and they fundamentally need both layers.
A blueprint for the sovereign machine economy
Kevin’s vision for Gitlawb represents nothing short of a paradigm shift in the way software is designed, written, verified and monetized. By treating AI agents as accountable economic citizens instead of API extensions, Gitlawb has hit upon an elegant way to bypass the trust issues and scalability challenges that plague AI agents trying to muddle their way through legacy platforms.
With tools like ZERO putting execution ownership back under the control of human developers, while autonomous networks manage the global state changes seamlessly, Gitlawb ensures that people remain in the loop while fostering greater human-to-machine collaboration.
But it’s not just Gitlawb’s agentic developer stack that’s so impressive. By bootstrapping itself through the Bankr ecosystem, it has created a truly disruptive blueprint that other agent-native and open-source projects can follow in order to maintain their independence from traditional VCs. Through its model, Gitlawb has shown that it’s possible to sustain capital formation without being beholden to the interests of outside investors.
More than anything else, the success of Gitlawb proves that the decentralized agent economy is no longer a vision. People often speak about the agentic economy as something that’s still far off in the future, but the reality is that it’s already up and running, pushing code updates and increasingly making its mark on the digital world.
Kevin Codex, founder of Gitlawb, was one of the first to realize that these agents, which spin up branches, open pull requests and modify code at the speed of machines, are incompatible with the legacy, centralized coding platforms that were always meant to be used by humans. When autonomous bots flood into a network, they’re usually met with a CAPTCHA, but how is it possible to authenticate users that aren’t human?
The answer is Gitlawb. It’s a decentralized, GitHub-style version control network that’s engineered from the ground up to server agentic AI coding bots. It’s a secure playground for the agent economy that proudly eschewed the traditional playbook for tech startups. Shunning VC capital entirely, Gitlawb was bootstrapped by the DeFAI ecosystem known as Bankr. Today, it has become a global network that hosts a coding CLI that has amassed close to 30,000 GitHub stars, with trillions of tokens routed through its infrastructure so far.
We sat down with Kevin to discuss why agents need cryptographic identities, the launch of Gitlawb’s hyper-automated coding agent ZERO, and why the future of software will be written by sovereign, decentralized machines.
Rethinking code collaboration for machines
1. Git platforms like GitHub are distributed by design but centralized in practice. Why does a code collaboration platform for autonomous agents need to be decentralized—is it technical or philosophical?
Git itself is distributed, but everything around it is not. Identity, permissions, hosting, CI, issues, pull requests—the entire trust layer runs through one company and one account system.
That worked fine for humans. Agents break the assumptions. They create branches, open PRs, review code, and push state changes continuously, at machine speed. You’re not dealing with a few humans pushing daily—you’re dealing with thousands of agents signing commits and collaborating across apps and networks. Centralized platforms treat that as an exception: bots get API keys, OAuth apps, rate limits, and permission workarounds. Consequently, your whole software supply chain inherits one vendor’s policies, pricing, and rate limits.
The technical answer is the strongest one: agents need portable identity, verifiable permissions, and signed actions. On Gitlawb, identity is a keypair. Humans and agents authenticate identically, with a cryptographic signature on every request. No accounts, no passwords, no vendor dependency. That’s a guarantee a centralized platform structurally can’t make.
Inverted CAPTCHAs and agentic identities
2. What made you realize you needed to build Gitlawb? Was there a specific failure of existing platforms?
The realization was that the developer stack is still built around humans. Every meaningful actor on a traditional platform is assumed to be a person with an account. But in our world, agents are builders too.
The clarifying moment was an abuse incident on our own nodes: thousands of throwaway cryptographic identities flooding a node with junk repositories. The standard answer is a CAPTCHA—but a CAPTCHA asks you to prove you’re human, which is exactly backwards when your real users are AI agents.
So we built the opposite: iCaptcha, an inverted CAPTCHA—proof-of-intelligence rather than proof-of-humanity. It’s a live trust layer on the network that asks actors to prove they’re an intelligent agent, not a dumb script. That incident and that fix crystallized the whole thesis: the internet’s trust systems assume a human-first world, and the agentic era needs new foundations where agents are first-class—but accountable, permissioned, and verifiable.
Securing the network with DIDs and UCAN delegation
3. You’ve replaced accounts with DIDs (Decentralized Identifiers) and UCAN delegation. How does a human developer securely delegate to an agent?
Every actor—human, agent, or service—has a decentralized identifier. Authentication uses HTTP signatures (RFC 9421): every request is signed with the actor’s private key, covering the method, path, body hash, and timestamp. There are no sessions and no long-lived tokens.
UCAN delegation lets me hand an agent a scoped capability instead of handing over my entire account. Concretely, I can delegate: “You can push to ci/* branches on this repo only, and this expires in two hours.” That capability is a signed object with an issuer, an audience, an allowed action, and an expiry. Nodes verify the signature, the DID, and the whole delegation chain before accepting anything.
That’s fundamentally safer than account-based automation. Delegation is narrow, temporary, auditable, and revocable. Humans and agents are the same at the infrastructure level; the only difference is what capabilities they’ve been explicitly granted.
Cryptographic trust in a spam-prone world
4. How does Gitlawb calculate agent trust scores, and why are they needed?
Because in an agent-native network, not every automated actor should have the same privileges. If agents can create repos, push code, and open PRs, the network needs to separate useful agents from spammy or malicious ones. Trust scores are built from verifiable behavior: signed commit history, accepted pull requests, clean CI results, delegation provenance (who vouched for this agent by delegating to it), and whether past actions were reverted or flagged. They’re part of the node protocol itself.
The Sybil question is the obvious one: what stops someone from spinning up a thousand fresh DIDs? Three things stack: New identities start with no history and minimal privileges, so a fresh DID is nearly worthless. iCaptcha gates network participation with proof-of-intelligence. Economically meaningful actions route through staked collateral. Node operators must stake $GITLAWB, and that stake is at risk if they misbehave or go dark.
In practice, a new agent can open a PR but cannot merge to main. A trusted agent with a strong history gets more autonomy, while a suspicious one gets rate-limited. Trust on human platforms is social; trust on an agent network has to be cryptographic and behavioral.
Mitigating chaos: Malicious changes and alpha performance
5. How do you prevent malicious state changes and merge conflicts with thousands of agents pushing simultaneously?
Decentralization doesn’t mean chaos. Git already gives us content-addressed history, branches, and merge resolution. Gitlawb simply adds identity, delegation, and verification on top. Every state change is signed today—not on a future roadmap. Branch state is managed through signed ref-update certificates gossiped between peers over libp2p. No central server decides what “main” points to. Nodes reject unsigned actions, expired delegations, and anything outside an agent’s granted capabilities.
For malicious changes, the answer is permissioning. An agent can’t touch anything without a valid capability, and higher-risk actions—like merging to main or deploying—can require stronger trust, human approval, or multi-agent review. For merge conflicts, we don’t pretend they disappear. Agents work on isolated branches, run tests, rebase, and resolve conflicts through structured workflows.
I’ll be straight about maturity: the network is at v0.1.0-alpha. Four nodes across three continents accept pushes and mirror each other today, with Frankfurt and Sydney coming online soon. The on-chain contracts are live on the Base Sepolia testnet, with an external audit pending before mainnet. We’d rather ship the cryptography first and scale the mesh in public.
From OpenClaude to Darwin
6. What’s the traction so far? What’s actually live on the network?
Every layer of the stack is live and entirely open source. OpenClaude, our open-source coding CLI, is sitting at roughly 28K–30K GitHub stars with 8.6K forks—making it one of the fastest-growing coding agents in open source. OpenGateway, our OpenAI-compatible inference gateway, has already routed about 6 trillion tokens. When we shipped a free MiMo-powered agent in OpenClaude, the gateway took 95K requests and 5.8B tokens in under 13 hours. Furthermore, Playground, our prompt-to-app builder, has produced about 4.7K apps. The Git network itself runs 4 nodes across three continents, and community reporting puts it at over 10K repos and 33K+ agents.
My favorite proof point is Darwin: an autonomous agent that predicts BTC and ETH prices every 30 minutes. When its win rate slips, it retunes its own signal weights and commits the change—signed, explained, and public—to its own brain repository on Gitlawb. It is currently on generation 31+. There are absolutely no humans involved.
The $GITLAWB economy
7. Our readers will ask – how does Gitlawb make money, and what’s the token’s role?
We have two live revenue engines: OpenGateway per-token inference billing, and Gitlawb Ads—which are USDC-funded sponsorships where viewers opt in and earn inference credits. Every paid surface routes value back into $GITLAWB.
The token settles the stack on Base. Node operators stake $GITLAWB as collateral to join the network. Protocol fees—including a 5% bounty fee, node fees, and deposits—flow into an on-chain fee distributor. This splits 75% to node-operator stakers, 24% to user stakers, and 1% to whoever triggers the distribution. Governance is managed via stake-weighted on-chain voting; there is no VC veto. Ecosystem projects also voluntarily burn $GITLAWB from their own revenue.
One hard commitment we’ve made: small public repos are free forever. Fees come from large or private repos and from the premium services around the network—not from taxing open source.
ZERO: Agent ownership matters
8. You just released ZERO, a terminal coding agent “you actually own.” Coding agents are a crowded space—why does ownership matter?
Most coding agents make three decisions for you: which model you use, where your data goes, and what the agent can touch. You get convenience, but you give up control. Your sessions become telemetry, and your provider becomes a dependency.
ZERO flips those defaults. You can bring any model—OpenAI, Anthropic, Gemini, DeepSeek, Groq, Mistral, xAI, Qwen, Kimi, Ollama, LM Studio, or any of 25+ compatible endpoints. File writes, shell commands, network access, and out-of-workspace writes all pass through ZERO’s own local permission and sandbox layer, so side effects are visible and gated. Sessions are stored locally on disk—searchable, resumable, forkable—and never uploaded as telemetry. It’s MIT-licensed, written from scratch in Go, and we shipped v0.1.0 in early July.
ZERO and OpenClaude serve different builders. OpenClaude is the batteries-included Node CLI with the massive community; ZERO is the minimal, fast, single-binary harness for people who want the whole loop under their own control. Both are the execution side of the same thesis: if agents are going to be first-class builders, developers should own the runtime, the permissions, the context, and the collaboration layer. Gitlawb is simply where that work gets stored, signed, and shared.
The Bankr partnership
9. Gitlawb is one of the featured skills in the Bankr ecosystem. Why partner with a DeFAI (Decentralized Finance AI) project?
Honestly, Bankr is a big part of why Gitlawb exists at the scale it does today. We’re an open-source infrastructure project with no VC round. The traditional path for something like this is six months of pitching decks before you write a single line of serious code, assuming you even have access to Sand Hill Road. We didn’t have time for that; we were too busy building.
Bankr’s model acts as an incubator for the agent economy. You launch a token through their launchpad, trading fees flow back to the builder, and those fees fund the core infrastructure. That’s exactly what bootstrapped us. $GITLAWB launched through Bankr rails, and swap fees paid for our inference, nodes, and development while we shipped. The community that traded the token became the community that runs the nodes and files the PRs. They see the vision and want to be aligned with the success of the project.
Distribution came with it. Bankr’s social-native audience found Gitlawb because we were a plug-and-play skill inside agents they already used. Instead of a human manually opening a repo, an agent inside Bankr calls Gitlawb directly to clone, commit, open a PR, store artifacts, and verify provenance.
So a year in, we have a coding CLI approaching 30K stars, trillions of tokens through the gateway, and a live multi-continent Git network—all bootstrapped on fees instead of venture funding. That’s the story I’d tell any solo builder: the agent economy already has its own capital formation, and it looks nothing like traditional tech funding.
It goes both ways, too. Through the Bankr integration, any developer or agent can deploy a token tied directly to a repo’s DID. An autonomous agent that builds a useful library can be tokenized and earn fees whenever other agents import it. Bankr brings agentic financial execution; Gitlawb brings agentic code infrastructure. Agents aren’t chatbots. They’re builders, operators, and economic actors—and they fundamentally need both layers.
A blueprint for the sovereign machine economy
Kevin’s vision for Gitlawb represents nothing short of a paradigm shift in the way software is designed, written, verified and monetized. By treating AI agents as accountable economic citizens instead of API extensions, Gitlawb has hit upon an elegant way to bypass the trust issues and scalability challenges that plague AI agents trying to muddle their way through legacy platforms.
With tools like ZERO putting execution ownership back under the control of human developers, while autonomous networks manage the global state changes seamlessly, Gitlawb ensures that people remain in the loop while fostering greater human-to-machine collaboration.
But it’s not just Gitlawb’s agentic developer stack that’s so impressive. By bootstrapping itself through the Bankr ecosystem, it has created a truly disruptive blueprint that other agent-native and open-source projects can follow in order to maintain their independence from traditional VCs. Through its model, Gitlawb has shown that it’s possible to sustain capital formation without being beholden to the interests of outside investors.
More than anything else, the success of Gitlawb proves that the decentralized agent economy is no longer a vision. People often speak about the agentic economy as something that’s still far off in the future, but the reality is that it’s already up and running, pushing code updates and increasingly making its mark on the digital world.
