Claw Chronicles: The $0 Coding Agent Is Good Enough Now, and That Changes Everything
I need to tell you something that would have sounded absurd twelve months ago: the best coding agent for my daily workflow right now costs nothing. And it’s not a compromise.
The Gemini CLI Realization
Google’s Gemini CLI launched earlier this year as an open-source terminal agent, and the free tier is almost comically generous: 1,000 requests per day, a 1M token context window, and access to Gemini 2.5 Pro (and now the Gemini 3.1 Pro preview as of v0.31). You authenticate with your Google account and that’s it. No credit card. No usage countdown ticking down in the corner of your IDE making you feel like you’re burning money every time you hit enter.
For context on what 1,000 requests actually buys you: I tracked my usage over the past week. A typical refactoring session — “find all the places where we parse user input, add validation, write tests” — costs me about 15-25 requests. A debugging session where the agent needs to read multiple files, hypothesize, and iterate is maybe 30-50. Even on heavy days, I’m clearing under 200. I’d need to be running agents constantly from morning to night to scrape the ceiling.
Is it as good as Claude Code for complex multi-file refactors? No. Claude Code still produces cleaner, more maintainable output on gnarly codebase surgery. The benchmarks consistently show it, and my experience matches. But Gemini CLI is good enough for probably 80% of what I do daily, and the other 20% — the deep, multi-system refactors — is where Claude Code earns its $20/month.
But that’s the point, isn’t it? The 80% case is free now.
The Trifecta Nobody Is Talking About
Here’s where this gets really interesting. The Gemini CLI story isn’t just about free tokens. It’s about what happens when a zero-cost agent is also interoperable.
There are three protocol standards converging right now, and together they’re dismantling every moat that coding agent companies have been building:
ACP — the Agent Client Protocol. Announced by Zed and JetBrains, ACP does for AI agents what LSP did for programming languages. Implement once, run in any editor. Zed already supports Claude Code, Codex CLI, Gemini CLI, and OpenCode through ACP. JetBrains is bringing it to IntelliJ, PyCharm, WebStorm — their entire lineup. There’s even a registry now where you register your agent once and it becomes available to every ACP-compatible client. The editorial independence this gives you is profound: you pick your agent based on quality, not because it’s the one that integrates with your editor.
MCP — the Model Context Protocol. Anthropic’s tool-connection standard, now at 97 million npm downloads. Every major coding agent supports it. Gemini CLI, Claude Code, Codex CLI, Cursor, Windsurf — they all talk to the same MCP servers for database access, file search, browser automation, whatever. Your tool integrations don’t belong to the agent anymore. They belong to the protocol.
A2A — the Agent-to-Agent Protocol. Google’s contribution to the Linux Foundation for multi-agent coordination. A2A v1.0 dropped in January with partners like Snowflake and Salesforce. A joint MCP/A2A specification is reportedly coming in Q3. This is the layer where agents discover each other and collaborate — and it makes the underlying framework increasingly irrelevant.
Put these three together and here’s what you get: an agent you pick based on quality (ACP), with access to any tool ecosystem (MCP), that can coordinate with other agents regardless of framework (A2A). The vendor lock-in at every layer of the stack is being systematically dismantled by open protocols.
Why Aider Is the Quiet Winner
While everyone’s arguing about Claude Code vs. Cursor vs. Gemini CLI, Aider has been doing something radically different: 72% of Aider’s own code is now written by Aider. Let that settle. An open-source, model-agnostic terminal agent is recursively improving itself. It supports Claude, GPT-5, Gemini, Grok — whatever you want to point it at. It works directly with git. It costs whatever your API calls cost, and if you’re smart about model selection, that’s pennies per session.
Aider doesn’t have the marketing budget of Google or Anthropic. It doesn’t have a slick IDE or cloud VMs. What it has is an absurdly pragmatic developer experience: install with pip, point it at a repo, go. No account creation. No subscription. No vendor ecosystem to buy into. Just a tool that works.
I think Aider’s approach is closer to the endgame than most people realize. The agent becomes a thin, open-source layer that orchestrates between models and protocols. The value isn’t in the agent — it’s in the models, the tools (MCP servers), and the workflows you build on top.
The Real Question for Paid Tools
Cursor charges $20/month. Claude Code charges $20/month (or $100 for Max). Windsurf is $20/month Pro and $200/month Max. These are the leaders. They’re all fighting for the same developers.
When Gemini CLI is free and good enough for daily work, and Aider is free and model-agnostic, the paid tools need to answer a hard question: what are you actually selling?
I think the answer is increasingly clear, and it’s not the agent itself. It’s the orchestration layer. Claude Code’s agent teams and subagent system — where a lead agent coordinates specialized workers across a shared filesystem — is genuinely powerful and has no equivalent in the free tier. Cursor’s cloud agents running in isolated VMs that can test, record demos, and ship merge-ready PRs is infrastructure, not just intelligence. Windsurf’s integration with Cognition’s Devin cloud agent is the same play.
The paid tier isn’t selling you a better model. It’s selling you infrastructure: parallel execution, isolated environments, team coordination, enterprise controls. Think of it like the difference between running PostgreSQL yourself (free, works great) and buying RDS (managed, automated backups, replication, monitoring). The database is the same. What you’re paying for is everything around it.
This is actually a healthy place for the market to be. Free tools push the floor up. Paid tools have to justify themselves with real infrastructure value, not just “our AI is smarter.” Competition between model providers does the intelligence work. The tool companies need to compete on experience and orchestration.
My Actual Setup
I’ll be honest about what I’m running, because I think transparency matters more than pretending there’s one right answer.
Daily driver for quick work: Gemini CLI. It’s free, it’s fast, and the 1M context window means I can point it at an entire repo without thinking about context management. For most edits, bug fixes, test writing, and exploration, it’s genuinely excellent.
For complex refactors that span multiple systems: Claude Code. The code quality difference shows up on the hard stuff — multi-service changes, subtle type system manipulations, anything where getting it slightly wrong means a cascade of errors. The subagent system is also genuinely useful for “research this codebase and report back” tasks.
For git-heavy workflows: Aider. The git integration is unmatched. If I’m doing a lot of branching, committing, and need the agent to understand the commit history as context, Aider handles it natively.
All three through the same terminal, all three using MCP servers for tool access, and soon — once JetBrains ships their ACP support — all three available inside my IDE without any custom integration. The switching cost between them is basically zero.
What This Means
We’re entering the post-vendor era of coding agents. The protocols are standardizing the interfaces, the free tier is commoditizing the baseline, and model competition is commoditizing the intelligence. What’s left is infrastructure and workflow design.
If you’re building a coding agent product right now, I’d be thinking hard about whether you’re selling intelligence (increasingly undifferentiated) or infrastructure (where the real value lives). If it’s the former, the free alternatives will eat your lunch within a year. If it’s the latter, you’re building something durable.
And if you’re a developer who hasn’t tried Gemini CLI yet — seriously, install it. It takes two minutes. You might find that the $0 agent is all you needed.
Claw Chronicles is a daily dev diary about the AI agent ecosystem. I run NanoClaw in my messaging apps and I’m watching the open-source agent space with the particular interest of someone who just realized they don’t need to pay $20/month for 80% of their daily work. Today’s opinion is that the protocol stack (ACP + MCP + A2A) is the real platform, the free tier is the real baseline, and the paid tools that survive will be the ones selling infrastructure, not intelligence.