Changelog

/benchmark is back, as a waitlist for oathe test — the command that reports what happens to an agent’s work after process death, a lost session, a duplicate run, compaction, or a false “done”. Same format as the Cloud waitlist: sign in with Google or GitHub, or leave an email. Joining either waitlist now sends a confirmation email.

The homepage install block is now two tabs — an AI Prompt your agent can run end to end (prerequisites included), and the CLI commands. Docs lead with the Runtime Quickstart, rewritten with the real requirements (Node 22.13+, Postgres) and expected output at every step; the scanner guide is now the Skill Scanner Quickstart. And Oathe Cloud opens its early-access waitlist at /cloud — sign in with Google or GitHub, or just leave an email.

Three releases (0.4.1 → 0.4.4) tighten the daily loop. Breaches now render as one budgeted digest everywhere — session start, launcher splash, tool attention, the notch — at most 8 rows, sharpest first. oathe update makes the documented upgrade one verb: it installs @latest through the right npm and re-runs init through the new bin. And the notch survives upgrades: every re-init restarts its LaunchAgent cleanly, so the glass comes back on its own.

@oathe/oathe 0.3.1 is the first full release of the runtime package (0.3.2 followed a day later). oathe init brings up the local substrate — vendored, sha-manifested DDL, additive-only — and onboards every detected harness (Claude Code, Codex, Cursor) onto one board. Claims are speech acts: claim before you build, record progress as statements, and no agent grades its own homework. On macOS, init seats the notch: a quiet glass on the camera housing that pulses verdicts and opens the board on click.

One page per harness: /claude, /codex, /cursor. Each shows what happens when a session hits its limit and the task continues in another harness — real commands, the real failure text. The harness chips on the homepage link to them.

/reports lists every public audit — 2,986 skills and MCP servers — fifty to a page, browsable by verdict. Every report page now has a path from the homepage instead of living only in the sitemap.

Three runtime docs land: Runtime Quickstart, How a Handoff Works, How Verification Settles Work. llms.txt and the site’s structured data lead with the runtime and name Oathe Scan as its audit engine. What Is Oathe? is rewritten around the runtime, with Scan as the measurement it stands on.

We realized how unreliable and unsafe agent skills make agents with oathe scan. We also found hundreds of other ways long running agents fail. All of those learnings are packaged into our open-source harness runtime for true multiplayer agent collaboration. The runtime repository is public under Apache-2.0: github.com/oathe-ai/oathe.

The site is rebuilt around the runtime. New homepage with the kill demo and system schematic. Scan gets one page — live scan bar, OpenClaw benchmark, reports — with /audit redirecting there. New pages: /primitives, /matrix, and the Agent Reliability Benchmark at /benchmark.

The scanner we shipped in February is now Oathe Scan. It is the first thing we built: run a third-party skill in a sandbox and measure what it does to an agent before an agent installs it. /audit moves to /scan and redirects there. Same engine, same free tier — now the audit engine of the runtime that follows.

Redesigned homepage: blog poster strip, chat-window Ask-an-LLM tiles, and a Trusted-by section with real testimonials — Neo4j featured, with AI·SLE and Storm Media Marketing.

Signup, pricing, and dashboard are live. Session-based auth with Google OAuth, email verification, and password reset; pricing CTAs adapt to your account. Privacy policy, terms, and a contact page shipped alongside. Follow-up hardening: OAuth token refresh no longer races after login, and ClawHub skills verify via API before falling back to the GitHub mirror.

Published two new posts: What Is Oathe? is the canonical introduction covering trust scores, verdicts, and the sandbox approach. Getting Started with oathe-mcp is a step-by-step tutorial for setting up the MCP server in Claude, Cursor, and Windsurf.

First public release of the Oathe MCP server on npm. Five tools available: submit_audit, check_audit_status, get_audit_report, get_skill_summary, and search_audits. Install with npx oathe-mcp — no API key required.

Strengthened all five MCP tool descriptions with directive pre-install security language. Added server-level description so AI assistants understand Oathe’s purpose on connection. Published to npm via OIDC Trusted Publisher — no tokens or OTP needed.

Dedicated setup guides for each AI coding platform. Each page includes platform-specific MCP configuration examples and explains how to use Oathe to check tools before installing them.

Launched skill marketplace pages with browsable registry of audited MCP servers and AI agent skills. Each skill shows its trust score, verdict, and link to the full audit report.

Published full API reference at oathe.ai/docs with endpoint documentation for submitting audits, checking status, retrieving reports, and searching the audit database. Includes request/response schemas and rate limit details.