What is Softchestrate?
Softchestrate, from the Laragear package organisation, is a set of plain Markdown subagent and skill definitions meant to be imported into any agent harness that supports subagents and skills, such as Claude Code, Codex or Cursor. A planning-implementations skill writes a plan an executor can run with no chat history; orchestrating-plans dispatches one phase-executor subagent per phase and gates each return against a manifest, the diff and the acceptance criteria before any commit.
The 12 subagents include build-doctor, code-explorer, code-reviewer, codemod-runner, test-runner, refactoring-finder and security-auditor; the 10 skills cover devcontainers, deploys, reversible schema migrations, dependency upgrades, local production replicas and documentation. No definition names a model: each carries a complexity tier from mundane to critical and the harness picks the cheapest model that fits. The repository was created on 8 November 2026; version 1.1.0 shipped on 16 November.
Is Softchestrate worth it?
Most skill packs tell one agent how to behave. Softchestrate is about who does the work: a planner writes the plan, an orchestrator hands each phase to a fresh subagent with a narrow role, and nothing is committed until a gate reconciles what was asked, what was reported and what the diff actually shows. The token argument follows: a codemod-runner rated mundane can run on a cheap model while the security-auditor rated judgment gets a strong one.
The definitions are unusually concrete about what success looks like. Each has an It's working if line: build-doctor returns one ranked root cause with log evidence, not a raw log; the refactoring-finder returns at most 15 ranked specs; disproving-audit-findings ends every finding as Confirmed, False positive or Not testable, each backed by a reproduction test. Read-only roles are meant to be restricted to read tools in the harness configuration, not only in prose, and the README asks you to keep deploying-builds from being started by the model alone.
The catch: it is two weeks old with no published evaluation and, at the time of checking, no GitHub stars, so the fewer tokens claim is a design intent, not a measurement. Adoption takes work: the subagents must be translated into each harness's format, which the README suggests asking the agent itself to do, and the installer overwrites local edits on every update.
Our take: for a team already running multi-phase agent work in Claude Code or Codex, the definitions are worth reading as a reference for role boundaries and gates even if you never install them. Superpowers remains the tested default for a single agent's discipline.
Where it earns its place
A tech lead at a 12-person Laravel shop using Claude Code
Run a multi-day framework upgrade without one agent losing the plot
Use upgrading-dependencies to map breaking changes, planning-implementations to write the plan into pending/, then orchestrating-plans to run one phase per subagent with a review gate before each commit. Free; cost is model usage, lower where mundane roles run on cheaper models.
A security-minded backend developer reviewing an agent-written feature
Separate real vulnerabilities from false alarms before a release
Dispatch security-auditor, which reports each finding as a testable claim with an input-to-sink path, then disproving-audit-findings, which writes reproduction tests and marks each Confirmed, False positive or Not testable.
Before you commit
- Setup
- `npx skills add Laragear/Softchestrate` (Node.js 22.20+) for the skills, then the install-agents.sh script, with `--dry-run` first, to copy subagents into ~/.agents/agents; ask your harness to adapt them following AGENTS.md.
- Integrations
- Any harness with native subagent and skill discovery, such as Claude Code, Codex, Cursor or ZCode; MCP tools where available, with plain search and file reads as a fallback.
- Data and privacy
- Plain Markdown definitions and small scripts; nothing is sent anywhere by the library itself. Code goes to whichever models your harness assigns. MIT licence.
- Limits
- No token or quality benchmarks published; per-harness adaptation is manual or agent-assisted; updates overwrite installed copies.
What’s good
- Clear success signal for every subagent and skill
- No model hard-wired; complexity tiers guide cost
- Commit gates check reports against the actual diff
- Read-only roles and a guarded deploy skill
What’s not
- No published evaluation of token savings
- Subagents must be adapted per harness
- Installer overwrites local edits
- Brand new, with almost no outside users
Best for
Teams running multi-phase agent work who want explicit roles, plans and gates rather than one long agent session.
Softchestrate alternatives
Softchestrate: common questions
- Is Softchestrate worth it?
- Softchestrate is a careful, opinionated division of labour for coding agents that nobody else has yet tested, its author included in public. We rate it 3.5 out of 5.
- How much does Softchestrate cost?
- Free and open source (MIT). You pay for your harness's model usage.
- What are the downsides of Softchestrate?
- No published evaluation of token savings; subagents must be adapted per harness; installer overwrites local edits; brand new, with almost no outside users.
- Who is Softchestrate best for?
- Teams running multi-phase agent work who want explicit roles, plans and gates rather than one long agent session.
- Is Softchestrate safe for company data?
- Plain Markdown definitions and small scripts; nothing is sent anywhere by the library itself. Code goes to whichever models your harness assigns. MIT licence.
- What are the best Softchestrate alternatives?
- The closest alternatives we have reviewed are Superpowers, Agent Skills (Addy Osmani) and Spec Kit.
Answers checked Nov 21, 2026.