oathe

oathe / harnesses / claude code

Claude Code hits its limit. Pick up where you left off.

You're mid-refactor. Claude Code prints You've reached your usage limit and stops. Your files are fine. What you were doing is in a context window that no longer exists.

Oathe focuses on your goal, not just your session. Start Claude Code through Oathe and your task is claimed when work begins, checkpointed as it runs, and still yours after the session ends.

Start

Node 22 or newer.

npm install -g @oathe/oathe cd your-project oathe init # onboards your harnesses oathe claude # normal session, board first

Nothing about Claude Code changes. Same prompts, same tools, same subagents, same permission modes. Oathe sits underneath and records what you're doing.

Continue in another agent

Ctrl-C anytime. Open a different harness in the same folder.

oathe codex # Codex picks up your task oathe agent # or Cursor's CLI

Why should you have to paste a transcript or re-explain the goal? Whichever harness you open gets a compiled context bundle: what you claimed, what's already done, what's left, and the test that decides whether it's finished. It reads the diff and reruns the test before trusting anything the last agent told it.

Works the other way too: start in Codex and finish in Claude Code.

"Done" is a claim until something checks it

Claude Code is good. It can also be wrong and confident at the same time.

When it says your work is finished, Oathe records that as an assertion, not a fact. A verifier that didn't write the code reads the recorded session traces and checks what the agent did against what you asked for. Pass and your task settles. Fail and it reopens. You can run that check through a different harness entirely.

oathe claim fix-login "users can sign in again" oathe done fix-login "guard rewritten, test added" src/auth.test.js oathe verify fix-login # a non-author checks the traces

In-session, Claude Code can do all of this itself through the oathe_* MCP tools.

Before Claude Code installs a skill

Every MCP server you add runs with the same access Claude Code has. There is no permission model between them.

Oathe Scan runs a skill in a sandbox first and reports what it actually did — file access, network calls, process activity. We've audited 2,986 of them, including 24 that were outright malicious.

Next

Oathe is open source (Apache-2.0) · What Oathe is · What's proven per harness today · Claude Code