Scribe vs alternatives
Scribe is for teams that want a reproducible, cross-agent way to publish, install, scope, and update coding-agent skills from registries. The closest direct peers are xingkongliang/skills-manager, pivoshenko/Kasetto, and 803/skills-suppl...
Scribe is for teams that want a reproducible, cross-agent way to publish, install, scope, and update coding-agent skills from registries. The closest direct peers are xingkongliang/skills-manager, pivoshenko/Kasetto, and 803/skills-supply — three projects that, like Scribe, treat SKILL.md as the unit and project it across multiple agents. GitHub Copilot custom instructions sit nearby as the most-used per-repo rule renderer. Scribe is not a replacement for an IDE agent, MCP servers, or coding methodologies (see Different layer).
Feature comparison
| Feature | Scribe | xingkongliang/skills-manager | pivoshenko/Kasetto | 803/skills-supply | GitHub Copilot custom instructions |
|---|---|---|---|---|---|
| Primary job | Skill manager: publish, install, project, sync | Desktop GUI for cross-tool SKILL.md library |
Declarative agent environment manager (Rust CLI) | One-manifest skill installer for multiple agents | Repo-scoped instructions for Copilot Chat, code review, agent |
| Where instructions live | ~/.scribe/ (canonical), projected to .claude/, .cursor/, .codex/, .gemini/ |
~/.skills-manager/ (default), GUI-managed |
kasetto.yaml + lockfile, projected to agent dirs |
agents.toml per project or global |
.github/copilot-instructions.md, .github/instructions/*.instructions.md, AGENTS.md |
| Cross-tool support | Claude Code, Codex, Cursor, Gemini, custom | Cursor, Claude Code, Codex, OpenCode, Amp, Roo Code, Gemini CLI, Copilot, Windsurf | 21 built-in agent presets (Claude Code, Cursor, Codex, Windsurf, Copilot, Gemini CLI, …) | Claude Code, Amp, Codex, OpenCode, Factory | Copilot (Chat, code review, cloud agent). AGENTS.md is also read by Claude (CLAUDE.md) and Gemini (GEMINI.md). |
| Path/file scoping | .scribe.yaml per project, kits, snippets |
“Scenarios” + “Project Workspaces” | Per-skill selection + project/global scope | agents.toml (project or global) |
applyTo: "src/**/*.ts" glob frontmatter on *.instructions.md |
| Reproducibility | scribe.lock lockfile, scribe sync, scribe doctor |
Git backup/restore, snapshot tags, multi-machine sync (no lockfile) | Lockfile + pinned refs/tags/commits + doctor |
Tracked state + reconciliation, no user-facing lockfile | Plain markdown in git, no lockfile or sync command |
| Discovery | Connected registries (Naoray, anthropics, custom) | Git repos, local folders, archives, skills.sh marketplace | GitHub / GitLab / Bitbucket / Codeberg / Gitea / self-hosted, local paths, remote team config URLs | GitHub, arbitrary Git remotes, local paths, Claude plugin marketplaces | None — write your own, or borrow from awesome-copilot |
| Existing skill adoption | scribe adopt claims unmanaged local skills via symlink |
Project Workspaces compare project-local skill folders with the central library and sync either direction | Limited — sync from local folders, no one-shot claim | Refuses to overwrite manually-added skills; can reset state to mark them “unmanaged” | N/A — instructions are author-only |
| Per-project state | .scribe.yaml, .ai/scribe.lock, projection map |
Project Workspaces with per-agent assignment | Project-scope install + selectable subset | agents.toml per project |
Repo-scoped instructions in .github/, but no manifest or lockfile |
| Budget enforcement | Codex skill-description guardrail in scribe doctor |
Not documented | Not documented | Not documented | Not documented |
| Schema / introspection | JSON envelopes, scribe schema <command>, exit codes |
--json output for scripts and agents |
sync, list, doctor, clean, self-update all expose --json |
Scriptable via --non-interactive, no JSON output |
Not applicable — natural-language markdown |
| Best at | Multi-agent skill lifecycle and team sync | Desktop-first cross-tool skill library with backup/sync | Declarative reproducible agent environments (CLI) | Manifest-driven skill sync across agents | Tightly-scoped repo instructions for Copilot users |
When to pick Scribe
- You maintain team skills in Git and want one manifest-driven install/update flow across multiple agent tools.
- You need reproducible sync, lockfiles, repair commands, and machine-readable output for agent automation.
- You want project-scoped skill projection so one repository does not inherit every skill installed on the machine.
- You already have hand-written skills and want to adopt them without moving or rewriting them first.
When to pick an alternative
- Pick xingkongliang/skills-manager if you want a polished desktop GUI as the primary surface, and value built-in backup/restore + multi-machine sync over a CLI-first lockfile model.
- Pick pivoshenko/Kasetto if you want a declarative, Rust-based CLI with a strong lockfile +
--jsoneverywhere, and don’t yet need anadopt-style claim flow. - Pick 803/skills-supply if a single
agents.tomlper project covers your needs and you don’t need a registry/lockfile/adopt bundle. - Pick GitHub Copilot custom instructions if your team is all-in on Copilot and a path-scoped
*.instructions.mdplus a repocopilot-instructions.mdcover what you need. Lighter than Scribe, but no lockfile, no project state, no cross-tool projection. - Use skills.sh / Agent Skills as the format underneath any of the above. Scribe consumes
SKILL.mddirectly; most peers do too. It is a spec, not a tool.
Related projects (partial overlap, broader category)
These solve a slice of the same problem but sit in adjacent categories:
- pr-pm/prpm — universal registry that auto-converts one package into vendor-native formats (Cursor rules, Claude skills, Copilot instructions, …). Strong registry/conversion, weaker on canonical local storage and adoption.
- enulus/OpenPackage — package manager for agent configs (rules, commands, agents, skills, MCP) with project/global scope and a wide platform matrix. Broader than Scribe; not
SKILL.md-first. - frmlabz/omnidev — capability format (
omni.toml/OMNI.md) with provider-specific generation and profile switching. Alpha-stage; provider-renderer rather than skill manager. - dennishavermans/agentfile — single committed
ai/contract.yamlrendered intoCLAUDE.md,.cursor/rules/*.mdc,.github/copilot-instructions.md,AGENTS.md. Cross-tool contract renderer; no registry, no central skill store. - KrystianJonca/lnai — define once in
.ai/, sync everywhere. One-config renderer; no registry, no lockfile.
Different layer, not an alternative
- MCP servers solve runtime tool/data access (calling GitHub, reading a database, connecting to internal services). Scribe and Copilot instructions ship prompts and skills; MCP ships capabilities. Most teams want both.
- IDE agents (Cline, Roo Code, Cursor agent mode, Claude Code) are the runtime that executes skills. Scribe is the source-of-truth and delivery system; the agent is the consumer.
- Coding methodologies (TDD-first, agentic plan/build/review loops) are how an agent works. Scribe is what the agent reads.