How Rippletide's Free Tool Helps Ensure Reliable Code Production
Code agents are not inherently reliable. Large language models (LLMs) can hallucinate, and the only way to prevent this is to apply external safeguards outside the model itself. That's the approach taken by Rippletide, a French startup based in Paris. Initially launched for Claude Code, Anthropic's code agent, Rippletide now focuses on Codex, OpenAI's code agent. Its integration with Codex helps make your code generation more reliable—and can even help you avoid technical debt over the long term.
Errors that add up at every step
"If you use an LLM, there's always a 3 to 4% chance it won't follow the instructions you gave it. If you have a 3 to 4% risk at every step, those risks compound, and the error rate explodes exponentially," explains Yann Bilien, co-founder of Rippletide, speaking to JDN. Code agents are especially vulnerable to this, since they chain together dozens of tool calls, commands, and file modifications. The result: an agent can head off in the wrong direction and keep going for a long time before anyone notices. The problem only gets worse at the scale of a large company. "If I put my code agents in the hands of 5,000 or 10,000 engineers, my code quality will suffer. Things will go off in all directions," the expert warns. And monitoring these agents quickly overwhelms human attention: "After three or four Claude Code tabs, the bottleneck is the developer's focus."
This is exactly the flaw Rippletide was created to address. In practice, the tool analyzes the codebase, Markdown files, comments, and the requests a developer repeats from one session to the next, then suggests turning these habits into explicit rules. These rules are stored in a neurosymbolic graph (a system that combines two components: a symbolic part, made up of logical rules applied strictly, and a neural part, an LLM). Rippletide then connects to Codex hooks. "Every time the agent wants to call a tool, write a command, or modify a file, Rippletide intercepts it. We decide whether it's allowed or not, and we explain why," Yann Bilien says. The rules cannot be bypassed.
According to the startup, the benefits are clear. Since the agent is redirected as soon as it strays, it no longer wastes cycles. "We cut token consumption by half or even two-thirds," the expert claims. He says this figure was measured on Claude Code, not Codex. Over the long term, systematically applied rules produce more consistent code and, if the initial guidelines are sound, less technical debt. Rippletide also lets you monitor the agent's actions in real time, so you can catch issues or vulnerabilities as soon as they're detected. This is a valuable feature, since code agents don't natively offer any centralized, reviewable history of their actions.
Installation in just a few commands
Rippletide is available on macOS, Windows, and Linux, and installs with a single command:
npm install -g rippletide-package
Next, you need to log in using the following command:
rippletide login
This step opens a browser window, allowing you to log in to the Rippletide website (which is required to store your rules' neurosymbolic graph).
To then connect Rippletide to Codex:
codex plugin marketplace add "$(npm root -g)/rippletide-package/codex-plugin" codex plugin add rippletide@rippletide rippletide codex setup
What about Claude Code users? The startup still claims compatibility, but the dedicated version—a separate tool that hasn't been updated since April—did not work during our test. For now, Rippletide is best reserved for Codex.
The command connects Rippletide with Codex and creates four different hooks:
- SessionStart: triggers at the start of each Codex session. Rippletide then sends its platform the list of tools available to the agent, as well as project instructions pulled from files like AGENTS.md.
- PreToolUse: this is the main hook, and the only point where Codex can still be stopped. Before each tool call (a shell command, a file modification, etc.), Rippletide checks the action against its rules. If a rule is violated, the call is blocked and Codex receives the reason for the refusal, along with the rule's identifier, so it can adjust its course.
- PostToolUse: this runs after the action is executed. Rippletide logs the actual result of the call (success or failure, parameters, output).
- Stop: triggers when Codex finishes its turn and is about to respond. Rippletide also checks the agent's final response. If it violates a rule, Rippletide asks Codex to revise it.
The setup command also creates a default web environment linked to your Codex, in the form of a URL like: https://app.rippletide.com/build/yourID
Here, you can review your Codex agent's action logs and create new rules. By default, Rippletide doesn't enforce any rules. You have to provide them, either by creating them manually in the interface (which we don't recommend), or simply by asking Codex (which the Rippletide plugin installed above enables). For example: "Add a Rippletide rule: forbid file deletion." Or, even smarter: "Analyze my AGENTS.md and suggest a list of rules to hardcode in Rippletide. I'll approve them one by one, and we'll add them all."
Turning conventions into rules
As you've probably guessed, the quality of your code will (as always) depend on how relevant your rules are. Coding conventions, naming rules, and your company's architectural choices are all rules you can add to Rippletide. Codex will follow them from start to finish, so you don't have to check. In a way, it's a great complement to Codex's auto mode, which only uses basic security rules to decide what it can do without your approval.
As for pricing, the Rippletide Codex plugin is still in beta and is currently free.
On the security front, the system does what it promises: every agent action is checked before execution, then logged. There is a tradeoff, however: decisions aren't made locally, but in Rippletide's cloud, which receives the project instructions, details of commands and code changes the agent wants to make, the result of each action, and the agent's final response. Rippletide is a French company with its technical team based in Paris, but it's still an additional intermediary. The startup offers data hosting in the European Union, but doesn't specify whether this is the default.
This article was automatically translated from French. Read the original version: https://www.journaldunet.com/intelligence-artificielle/1557763-cet-outil-gratuit-permet-de-fiabiliser-votre-production-de-code-avec-codex/