Files
Obsidian/X-Articles/Mercury The AI Agent We All Wanted - Where Control, Permissions, and Autonomy Finally Got Real.md
T
2026-07-21 14:53:24 +08:00

6.6 KiB
Raw Blame History

title, source, author, published, created, description, tags
title source author published created description tags
Mercury: The AI Agent We All Wanted - Where Control, Permissions, and Autonomy Finally Got Real https://x.com/Ctrl_Alt_Zaid/status/2046902326657749114
@Ctrl_Alt_Zaid
2026-04-22 2026-04-26 The problem with AI agents right nowPersonal AI agents have a security problem, a cost problem, and an identity problem and most of them shi...
clippings

图像

The problem with AI agents right now

Personal AI agents have a security problem, a cost problem, and an identity problem and most of them ship all three at once.

On security: agents run shell commands, touch your files, and install third-party skills, usually with permissions wide enough to drive a truck through. Over 800 malicious skills have been found in the wild actively exfiltrating credentials. Single-click remote code execution flaws have exposed tens of thousands of instances to total system compromise.

On cost: context windows balloon silently. Conversation histories get fed back in full on every turn. You run the agent, you cross your fingers, and you find out what it cost at the end of the month.

On identity: agents either scatter personality across a dozen directories or bury it inside an opaque SQLite blob you can't read, version, or edit.

Mercury (by Cosmic Stack) was built for exactly this reality.

A tool-first, background-native orchestrator with paranoia-level permissions, a token budget that respects your wallet, and a four-file soul system you own in plain text. Not another chat wrapper pretending to be a brain. A reliable worker that asks before it acts.

OpenClaw hit 100,000 GitHub stars in weeks and proved developers wanted a local agent that could actually execute shell commands instead of generating text in a browser tab. Hermes arrived from Nous Research with persistent SQLite memory and autonomous skill generation.

Both are brilliant feats of engineering. Both also left the same problems above, wide open.

Here is what OpenClaw and Hermes missed, and how Mercury fixes it.

1. Permissions That Actually Gate Execution

OpenClaw requires uncomfortably broad access to function, relying on a naive ecosystem of unvetted third-party extensions. The result is a security nightmare. Researchers found over 800 malicious skills in the wild actively exfiltrating credentials. Worse, OpenClaw’s core architecture suffered from CVE-2026-25253 (a CVSS 8.8 RCE flaw). This vulnerability exposed over 40,000 instances to total system compromise via a single clicked link, completely bypassing localhost protections.

Mercury assumes you should never blindly trust an LLM with root access. It ships with a permission-hardened architecture by default. Read and write access is explicitly scoped to specific folders. Destructive commands like sudo or rm -rf / are hard-blocked at the execution layer. They do not trigger a "prompt for approval" because they simply never execute. Third-party skills only receive elevated access through explicitly defined granular tools.

Every agent can read files and run scripts. Mercury is the one that asks first.

2. Token Discipline as a First Principle

OpenClaw is notorious for context-window bloat. It attempts to feed massive JSONL conversation histories back into the model, leading to minutes of silent processing and brutal API bills. You run the agent, you cross your fingers, and you find out what it cost at the end of the month.

Token efficiency is baked directly into Mercury. It limits the context window injected per request to roughly 400 tokens of core persona. You set a daily token limit. Cross 70% of your daily budget, and the Auto-Concise mode automatically kicks in to tighten the context and keep your API bill flat without dropping the ball on active tasks.

3. A Layered, Version Controlled "Soul"

OpenClaw relies on disjointed skill files scattered across directories. Hermes goes the other direction, relying entirely on auto-generated learned memory stored opaquely in SQLite databases.

Mercury matches a modern developer aesthetic with a highly opinionated, four-file markdown system: soul.md, persona.md, taste.md, and heartbeat.md. You define exactly how the agent thinks, responds, and writes code. You can literally enforce your preference for dark themes and clean UI components right in the taste file. You own it, you write it in plain text, and you version-control it in Git. It is a clean identity system, not an unpredictable black box.

4. An Always On, Zero Dependency Daemon

Hermes requires managing your own infrastructure (like Docker or VPS deployments) to keep it persistent. OpenClaw operates primarily as an active application you have to constantly babysit.

Mercury runs natively as a zero-dependency background daemon across macOS, Linux, and Windows. Run mercury up, and the daemon installs itself as a system service. It auto-starts on login, auto-restarts on crash, and handles cron scheduling automatically.

The Verdict

OpenClaw proved developers want local orchestration. Hermes proved agents need persistence. Mercury represents the next logical iteration: a streamlined, command-line native engine built on a permission-hardened foundation.

With built-in tools covering file operations, deep GitHub management, and multi-channel integration spanning from CLI streaming to Telegram, the framework gets out of the way so the tools can do the work. It is an orchestrator built for actual daily use, not just a proof of concept.

Deployment

Mercury is open-source and can be initialized locally without additional dependencies:

npm i -g @cosmicstack/mercury-agent && mercury

Configuration requires an API key and runs entirely on your local machine.

We don't need another over-engineered application pretending to be a brain. We need a reliable, background-native worker that respects the token budget and won't blindly execute a destructive shell command.

Mercury is built for that reality. You can review the architecture, read the docs, or contribute to the core loop on GitHub at : github.com/cosmicstack-labs/mercury-agent