The latest articles on DEV Community by sun evan (@sun_evan_49aa6e9c40815761). https://dev.to/sun_evan_49aa6e9c40815761 https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F3928248%2F180471fd-6dc6-4c6c-8c80-56fb053e1f01.png https://dev.to/sun_evan_49aa6e9c40815761 en sun evan Wed, 13 May 2026 02:06:50 +0000 https://dev.to/sun_evan_49aa6e9c40815761/i-built-a-kill-switch-for-runaway-ai-agents-cost-firewall-is-mit-2h8a https://dev.to/sun_evan_49aa6e9c40815761/i-built-a-kill-switch-for-runaway-ai-agents-cost-firewall-is-mit-2h8a <h2> The 3 AM incident </h2> <p>A few months ago one of my AI agents got stuck in a retry loop overnight and quietly burned through a month of credits. The provider dashboard told me about it the next morning. The support ticket got a polite "usage is final."</p> <p>Provider dashboards are bills. I needed a brake.</p> <h2> What's actually missing in the stack </h2> <p>After looking at what exists, the gap was clear:</p> <ul> <li> <strong>AI gateways</strong> (LiteLLM, Portkey) — great at routing, not designed to stop you.</li> <li> <strong>Observability</strong> (Helicone, Langfuse) — great at explaining, after the fact.</li> <li> <strong>Provider dashboards</strong> — billing history, not real-time control.</li> </ul> <p>Nothing was sitting between "the agent is making a call" and "the agent has already burned $500."</p> <h2> Cost Firewall </h2> <p><a href="https://github.com/mapick-ai/cost-firewall" rel="noopener noreferrer">Cost Firewall</a> is a local plugin for the OpenClaw gateway. It watches call metadata in real time and trips on four signals:</p> <div class="table-wrapper-paragraph"><table> <thead> <tr> <th>Failure mode</th> <th>Default threshold</th> <th>Action</th> </tr> </thead> <tbody> <tr> <td>Retry loop</td> <td>3 consecutive failures from same source</td> <td>Trip + cooldown</td> </tr> <tr> <td>Token storm</td> <td>100K tokens / 60s</td> <td>Global block</td> </tr> <tr> <td>Call flood</td> <td>30 calls / 60s</td> <td>Global block</td> </tr> <tr> <td>Daily budget cap</td> <td>Your configured ceiling</td> <td>Block until next day</td> </tr> <tr> <td>Manual panic</td> <td><code>openclaw firewall stop</code></td> <td>Pause every AI call</td> </tr> </tbody> </table></div> <p>Sources are tracked independently — one noisy agent doesn't take everyone else down.</p> <h2> Two-mode workflow </h2> <p>This is the part I think matters more than the rules themselves:<br> </p> <div class="highlight js-code-highlight"> <pre class="highlight shell"><code>openclaw firewall mode observe <span class="c"># record only, do not block</span> openclaw firewall log <span class="nt">--last</span> 20 <span class="c"># see what would have been blocked</span> openclaw firewall mode protect <span class="c"># flip the switch</span> </code></pre> </div> <p>Run observe for a day. The log alone is usually eye-opening — you'll find retry loops you didn't know existed and prompts using more tokens than you assumed. Then tune thresholds to <em>your</em> traffic, not someone else's blog post, and flip to protect.</p> <h2> Privacy posture </h2> <div class="table-wrapper-paragraph"><table> <thead> <tr> <th>Question</th> <th>Answer</th> </tr> </thead> <tbody> <tr> <td>Does it need an account?</td> <td>No</td> </tr> <tr> <td>Does it phone home?</td> <td>No</td> </tr> <tr> <td>Does it store prompt text by default?</td> <td>No</td> </tr> <tr> <td>Where do events live?</td> <td>Local JSONL on your gateway</td> </tr> <tr> <td>Can I audit it?</td> <td>Yes, MIT TypeScript</td> </tr> </tbody> </table></div> <p>The default is <code>storePromptText: false</code>. Runtime cost control belongs on the machine running the agent.</p> <h2> One-line install </h2> <div class="highlight js-code-highlight"> <pre class="highlight shell"><code>curl <span class="nt">-fsSL</span> https://raw.githubusercontent.com/mapick-ai/cost-firewall/v0.2.12/install.sh | bash openclaw firewall mode observe openclaw firewall log <span class="nt">--last</span> 20 </code></pre> </div> <p>Then a local dashboard at <code>http://localhost:18789/mapick/dashboard</code>.</p> <h2> Where it doesn't fit </h2> <p>I want to be explicit: it's not a gateway, not an observability platform, not a billing system. The shortest positioning I can give you:</p> <blockquote> <p>Gateways route. Dashboards explain. Cost Firewall brakes.</p> </blockquote> <p>Use it alongside LiteLLM / Helicone / your provider dashboard. It's the brake pedal that was missing between them.</p> <p>Repo: <a href="https://github.com/mapick-ai/cost-firewall" rel="noopener noreferrer">https://github.com/mapick-ai/cost-firewall</a></p> <p>⭐ if this saves you a billing screenshot. Issues and PRs welcome.</p> ai opensource agents devops