Introducing Cusp

Today we're releasing Cusp 0.1, the first public preview of an agentic coding workspace for any model.

Cusp is a coding agent. It reads your code, edits files, runs commands, checks its own work, and asks before doing anything risky. It does this with whichever model you choose: Claude, GPT, Gemini, Grok, DeepSeek, Kimi, GLM, Qwen, MiniMax, Mistral, a model behind OpenRouter or your company's gateway, or one running on your own machine.

Why another agentCopy Anchor Link

Over the past year, coding agents became useful for real work, and most of them are tied to one lab's models. The agent you like decides which models you can use. If you want to try an open model at a fraction of the price, you start over with a different tool, different config, and different habits.

Meanwhile the models keep leapfrogging each other. The strongest model for a hard refactor this month may not be the strongest next month, and a quick rename never needed the expensive one. We think the agent should stay put while the model changes underneath it.

There's a second reason. On terminal-heavy tasks, the harness around a model matters a great deal: in published Terminal-Bench results, the same model scores up to about 30 points apart depending on the agent running it, and in one study a better harness helped a cheaper open model far more than a frontier one. We wrote about that last week in The harness is the product. An agent that works well with any model has to adapt to each one, not just pass messages along.

What's in the previewCopy Anchor Link

A runtime that adapts to the model. Cusp gives each model family the tools it was trained on. GPT models edit with apply_patch, Claude and Gemini with exact string replacement, and weaker or local models with hash-anchored line edits that don't require reproducing text perfectly. Reasoning state is replayed exactly the way each provider expects. Broken tool calls from open models are repaired before a turn is allowed to fail.

Paste a key and go. Cusp recognizes the provider from the key or URL, probes which APIs answer without spending tokens, lists the models, and picks a strong main model and a fast helper. It speaks four wire protocols (OpenAI Chat Completions, OpenAI Responses, Anthropic Messages and Gemini), which covers more than twenty providers and gateways out of the box, plus Ollama, LM Studio and vLLM.

Your existing setup, as is. Cusp reads AGENTS.md, CLAUDE.md and GEMINI.md, .claude/settings.json, skills, subagents, hooks and MCP servers where they already are, and imports the rest from the agents you already use. Hook events and permission rules like Bash(npm run test:*) work unchanged, and you can keep using other agents on the same repository.

A careful permission model. Every tool call goes through hooks and a permission engine that parses shell commands properly, so safe && dangerous doesn't slip past a rule for safe. Writes to .git, agent configuration and shell startup files need your approval in every mode. Shell commands can run in the macOS or Linux sandbox. Every decision is recorded.

Sessions you can rewind. A session is an append-only log of events. Files are snapshotted before every edit. You can resume a session from any surface, fork it to try another approach, or rewind it, code included, to before any prompt.

A CLI and a desktop app on one runtime. The cusp CLI has an interactive terminal UI and a headless -p mode with JSON output for scripts and CI. The desktop app for macOS, Windows and Linux runs the same daemon as a supervised sidecar and keeps your keys in the OS keychain.

What's not in it yetCopy Anchor Link

This is a preview, and we'd rather be clear about the edges than have you find them.

• The web app and cloud sync are coming soon. They'll replicate the same session log, so a session started on your laptop can continue in the browser.
• Crew, AI teammates with a role, goals, a schedule and a budget, isn't in 0.1. The docs describe where it's headed.
• Plugins and marketplaces aren't supported yet.
• One-click import isn't needed to get started, since Cusp reads .claude/ in place, but a tool to copy your setup into ~/.cusp is coming.
• Benchmarks. We're running Cusp through Harbor on Terminal-Bench and patch-style suites, and we'll publish the runs, with configurations and trajectories, rather than a single headline number.

Try itCopy Anchor Link

Install the desktop app from cusp.run/download. For the CLI:

Available with the first public release.

The quickstart walks you through your first session. If you already use another agent, start with Import your setup to see what carries over.

We'd like to hear what works and what doesn't. Open an issue on GitHub or find us on Discord.