Mark Pesce CLI — ExoCortex¶
March 12, 2026 · LinkedIn
9 reactions · 1 comments · 386 views
Mark Pesce at The Register this morning: AI agents need CLIs, not GUIs.
Screen-scraping interfaces designed for humans is the wrong abstraction layer. External agents operating tools via structured interfaces is the right one.
Not a surprise. It's exactly why I built ExoCortex.
ExoCortex is my AI practitioner rig — the full stack I use to develop software, manage knowledge, and run autonomous agents:
→ Claude Code — CLI-first AI development. Not a GUI plugin. A terminal.
→ kcp-commands — structured manifests at the CLI boundary. 289 tools. Agents stop burning context on --help round-trips. 33% of context window recovered.
→ Synthesis — knowledge graph over codebases. Structural queries in under a second. What exists, how things relate, what breaks if you change it.
→ KCP — the protocol layer above CLIs: workspace topology, trust boundaries, load order, human-in-the-loop signals.
→ IronClaw — two AI employees (Mímir + Klaw), running 24/7 on CLIs and APIs. No GUIs.
When Google released gws — their Workspace CLI — we had a KCP command manifest for it within the day. Not racing anyone. Just the normal workflow: every CLI tool agents touch gets a structured manifest.
Pesce calls the coming scramble a "SaaSpocalypse" — every SaaS company rushing to build CLIs before competitors do.
That's probably right. But CLIs are still just the surface.
What makes agents actually useful once they have structured interfaces is what sits above: knowledge infrastructure. Workspace maps. Trust boundaries. Evidence by construction. The difference between "send this email via gws" and "understand the full context of this relationship and act on it with appropriate authority."
That's the autonomous agentic web. The CLI is table stakes. ExoCortex is the layer above.
Article: https://lnkd.in/eFRsbfxN
Discussion¶
Agree: Agree