Vibeleaderboard
Index / agent
Visit github.com
Category
AI Agents
Rank
No. 1822Tools index
Pricing
Open Source
Type
AGENT
GitHub
8 stars
Latest release
v0.1.3
Date

About

Shark is a Claude Code plugin that turns the main session into an operator that briefs subagents instead of writing code. Subagent lanes build in parallel against a shared contract, merges are audited, and bugs need a root cause before a fix. It also bundles a step-debugger MCP server and a hook that blocks tool calls on secret files named in the project's AGENTS file.

At a glance

shark is an MIT-licensed Claude Code plugin that turns the main session into an operator for a team of subagents. The operator never writes code: it writes a shared contract and briefs, sends lanes to build in parallel git worktrees, has a separate agent audit each merge against your written invariants, and requires a root cause before any bug fix. It is one developer's working setup, at version 0.1.4.

Good fit

Claude Code users building multi-file features with parallel agents who keep hitting conflicting edits, business rules lost in merges, symptom-level bug patches or one huge final commit. The skills call commands by their key in your Commands table, so the same kit is meant to work on Python, JS/TS, Go or any stack with a test command.

Consider something else when

Your work is mostly small fixes, which the author says a single Claude Code session already handles well. Also skip it if you use another agent host, or if you already run your own orchestration loop and do not want a second set of start rules (init asks which one leads, and shark then stays on call).

Start here

In Claude Code, run /plugin marketplace add tusharsrivastava/shark-skills, then /plugin install shark@shark-skills. In your repo, run /shark:init. It finds or creates AGENTS.md, adds sections such as commands, invariants, test scope and secrets, and asks for each value it cannot find. Fill in your Invariants carefully: the merge and invariant audits check diffs against them.

Cost and requirements

Free under MIT; the AGENTS.md template and logo are CC0. It needs Claude Code, git, and Node.js with npx, which fetches the separate @debugmcp/mcp-debugger server at run time. The debugger supports Python, JS/TS, Go, Rust, Java, .NET and C/C++. The shark:architect reviewer is described as costly and best saved for big designs.

Limits to know

Init changes the project's output style in .claude/settings.json to shark:asd-ste100, a terse Simplified Technical English register; change it back if you prefer normal replies. The deny-secrets hook only blocks files listed in your Secrets section (.env, *.pem and similar by default). The debugger writes .claude/mcp-debugger.log, which you must add to .gitignore. Lanes are timeboxed to 20 minutes plus a 20% grace period.

Still unclear

  • How much extra model usage do the parallel lanes, separate audits and debugger agent add compared with one session doing the same feature?

Sources checked 2026-11-17. Source-based assessment.

Tags

claude-codepluginmulti-agentorchestrationdebuggingmcp

Media

Shark-skills

Comments (0)

No comments yet

Editorially curated, with community endorsements as a secondary signal. Corrections welcome.