Two open-source tools spotted on Hacker News solve two concrete problems for those running agents in production: keeping credentials secure and avoiding token waste when sessions go cold.

OneCLI is a gateway written in Rust that sits between the agent and external services. Instead of giving each agent real API keys, you assign them a placeholder. When the agent makes an HTTP call, the gateway swaps the placeholder for the real key, injects it into the request, and lets it go. The agent never sees the secret. You have a single place to manage keys, rotate them, and audit what each agent does.

claude-thermos solves a different but equally operational problem. Claude Code’s cache has a 5-minute TTL: if more than five minutes pass between one request and the next on the same prefix, the cache expires and the system re-encodes the entire conversation at full price. This often happens when the main agent waits for a subagent doing long work, because the subagent has a different cache prefix and doesn’t refresh the main agent’s. claude-thermos launches a local proxy that sends keep-alive traffic to keep the cache warm. Its developer measures savings around 22% of the bill across 185 local sessions.

Why this matters to you. If you’re taking agents past the demo stage, credential management and cost control for long sessions are two problems you’ll hit early. Both tools come from people who’ve lived through those problems, not from a lab theorizing.

If you want to try it. The README files in both repositories have a quick start. The commands haven’t been tested on this site’s workbench yet: start there for your first step.

In detail

OneCLI tackles the attack surface problem of distributed credentials. When you have ten agents calling twenty different APIs, keys end up scattered everywhere: in .env files, container environment variables, system prompts. Every copy is an exposure point. OneCLI’s pattern is an authentication gateway, an architecture used for years in enterprise infrastructure (HashiCorp Vault, AWS Secrets Manager) but rarely accessible to those working with AI agents at small scale.

The gateway is written in Rust and runs as an HTTP proxy. The agent authenticates to the gateway with an access token (header Proxy-Authorization), makes its HTTP call to the external service with a placeholder, and the gateway does the swap before the request leaves. Secrets are encrypted with AES-256-GCM and decrypted only at request time, matched by host and path pattern. There’s a Next.js dashboard to manage agents, secrets, and permissions.

The limits are concrete. The repo has 67 open issues out of 304 commits, a sign of a young project with plenty to fix. The quick start mode is single-user with no authentication: multiuser requires OAuth setup. The threat model assumes the agent can’t directly access the gateway or read its filesystem, but in many real deployments this isn’t guaranteed. If the agent runs on the same host as the gateway, a compromised agent could attempt to reach the secret store directly.

claude-thermos solves a narrower but economically significant problem. Claude Code’s prompt caching works with a 5-minute TTL: while the cache is alive, each turn costs 0.1x of the input price instead of full price. But when the main agent delegates a task to a subagent, the subagent has a different system prompt and tool set, so it works on a separate cache prefix. While the subagent runs for more than 5 minutes, the main agent’s prefix ages and expires. When the subagent returns, the main agent resumes with identical history but no cache: it must re-encode 200K-500K tokens at write rate (1.25x instead of 0.1x).

claude-thermos launches a local reverse proxy that intercepts traffic to Anthropic’s API, identifies cache lines (prefix chains), and sends keep-alive requests when it detects the main agent has been idle too long. Parameters are configurable: idle threshold, interval between warm-ups, max cycles per episode.

The claimed 22% savings is measured across 185 local sessions by the tool’s developer, not an independent benchmark. The sample is one person with their own usage patterns. The tool requires Python 3.11+ and works only with Claude Code’s CLI, not other clients or APIs.

As we noted about the bug that deleted the home directory, the weak point of agents in production isn’t the model: it’s the infrastructure around it. OneCLI and claude-thermos are handcrafted answers to problems that providers haven’t yet solved in their products. For those building agents, the lesson is methodological: first understand where your operational chain breaks (exposed credentials, cache expiry), then find or write the specific tool.