one source of truth for your agents
CLAUDE.md, AGENTS.md, .cursorrules — edit one, forget two. Mossbear keeps your guides — rules, skills, documents — as one copy that lands back on disk wherever your agents read it, local or remote. It grades every Claude Code run against them, and when you correct an agent mid-session that correction comes back as a drafted edit with its evidence. You say yes or no.
npm install -g mossbearReads your sessions on your machine. Your code, your diffs and your raw transcripts never leave it — see exactly what syncs.
CLAUDE.md, AGENTS.md, .cursorrules — edit one, forget two. Your conventions drift apart per tool.
Your remote Codex agent spins up in a fresh container while your local Claude Code knows all your conventions. Same project, split brain.
Notes, specs, references — scattered across apps and devices where no agent can reach them. So you paste the same background into every session.
You write instructions, the agent does… something. The only feedback loop is your own frustration.
Your rules reach every one of them. Grading starts with Claude Code — today it is the one agent that reports what it did in enough detail to grade.
auto-detects your platforms, imports the rules and skills you already have. You don’t start from scratch — you bring your CLAUDE.md with you.
Mossbear puts your rules and skills back on disk at the paths your agents already read — including in the repos your remote agents clone. Agents pull the context a task needs on demand, through the built-in MCP server.
Grading starts when a session ends — close your next one and its verdicts are already being worked out. And when you correct an agent mid-session, that correction comes back as a drafted edit to the rule, carrying its reasoning and the actions behind it. It applies only when you accept it.
created ~/.mossbear/config.json ~/.mossbear/config.json: dashboardUrl unset — runtime falls back to https://app.mossbear.io .mossbear/config.json: project proj_8f2c1d47a9b30e56 (derived from this repo's root commit — commit the file so every clone agrees) created .claude/settings.json with PostToolUse hook updated .claude/settings.json with PostToolUseFailure hook updated .claude/settings.json with Stop hook (mossbear sync) updated .claude/settings.json with SessionEnd hook (mossbear sync --session-complete) updated .claude/settings.json with SessionStart hook (mossbear session-start) In environments rebuilt between sessions — cloud containers, Codespaces, devcontainers — this hook also installs the current mossbear release before the first turn. It never updates this machine. To stop that, set MOSSBEAR_NO_UPDATE_CHECK=1; the hook still pulls this repo's context. To drop the hook, remove the SessionStart entry from .claude/settings.json — `mossbear init --no-session-start-hook` skips writing it in the first place. created AGENTS.md created .agents/ created .claude/skills/mossbear/SKILL.md created .claude/skills/mossbear-context/SKILL.md created .claude/skills/mossbear-cli/SKILL.md
Done. Next steps: Edit AGENTS.md to add your rules mossbear sync --token <token> — connect your dashboard (Settings → Developer → CLI Token); remembered after the first run mossbear link — adopt this repo's existing guide files (2 found) into your dashboard mossbear guides active — list cached guides mossbear doctor — verify your environmentevery graded action lands as one of three
one copy
Import the files you already keep — CLAUDE.md, .cursorrules, .agents/ — and each one becomes a single canonical guide. Edit it in a repo and the edit comes back to that one copy instead of forking a second. Pull it into another repo and it lands at the path it came from, so your agents read it where they already look.
cross-repo
Capture rules, notes and references from any device. Agents pull exactly what a task needs — mossbear pull, or the built-in MCP server — and push results back. Planning docs sit alongside your skills, so parallel local and remote agents work from the same page.
Your code and your diffs never leave your machine, and your raw transcripts stay there too — what syncs from them is redacted excerpts. The CLI holds no model credentials and produces no verdict of its own: grading runs on the server, from what it syncs.
Never uploaded, never stored on our servers.
Enough for the dashboard to grade your alignment — nothing more.
This isn’t a policy you have to trust — it’s where the code runs. The CLI is MIT licensed and ships as plain compiled JavaScript, no bundler and no minifier, so you can read the hook before you trust it:cat $(npm root -g)/mossbear/dist/hook-bin.js
$0
everything the product does today. multi-platform and synced from day one — no card, no trial clock.
Mossbear is in alpha, and free while it stays there. Paid tiers will come later — spend tracking is the first one planned — and nothing you set up now gets taken away when they do. If you want out,mossbear unlink hands back everything you linked.