Claude KISS

v0.7.1 Same Claude Code. Your defaults.

Love the model.
Control the harness.

You can love Claude Code and still want control over what surrounds it. Claude KISS keeps the same binary, account, and sessions, while giving you a focused prompt, six core tools, and explicit switches for everything else. Your normal claude stays untouched.

Install macOS + Linux 
curl -fsSL https://claude-kiss.com/install.sh | sh
Inspect the script Release v0.7.1

Requires Claude Code. No sudo. No wrapper telemetry. No auto-update.

06 core defaults 04 quick switches

The principle

Keep the Claude Code you value. Choose what surrounds it.

Claude KISS does not replace the model or build another agent runtime. It controls the layer where prompts, tools, integrations, and handoffs become ambient defaults. KISS points at the harness, never at the quality of the work.

  • Same Claude Code.The installed binary, your account, and your sessions stay in place.
  • Your boundaries.Questions stay questions; small fixes stay small; unrelated work stays yours.
  • Capability on purpose.Add web, skills, delegation, or MCP when the task actually needs them.

The surface

Six core defaults. Four quick switches.

The launcher requests six coding tools by default. Add a common capability when the task needs it, and combine switches when it needs more.

Core default

06
  • Bash
  • Glob
  • Grep
  • Read
  • Edit
  • Write

Quick switches

04
CLAUDE_KISS_WEB=1
WebFetch + WebSearch
CLAUDE_KISS_SKILLS=1
Your skills
CLAUDE_KISS_DELEGATION=1
Agent delegation
CLAUDE_KISS_MCP=1
MCP discovery, tool search + resources

Managed defaults suppress bundled workflows, official marketplace autoinstall, Claude.ai skill sync, Chrome integration, agent view, artifacts, connectors, remote control, and auto memory. Native CLI overrides, settings, hooks, and plugins remain live and can change the effective surface. EndConversation may also appear conditionally as a safety tool; it does not add file access.

Observed starting context

More room for the work.

In this captured Opus 5 comparison, the Claude KISS default started with 33.4k fewer estimated tokens in context—about 80% less than ordinary Claude Code.

41.6k → 8.2k
estimated tokens used
958.4k → 991.8k
free context
24.5k → 3.9k
system-tool context
Ordinary Claude Code41.6k used
Claude Code context view showing 41.6k of 1m tokens used, including 24.5k for system tools, 8.9k for memory files, and 2.7k for skills
Claude KISS default8.2k used
Claude KISS context view showing 8.2k of 1m tokens used, including 3.9k for system tools and 1.9k for memory files

One observed snapshot, not a universal benchmark or billing claim. /context reports estimates; versions, models, account features, settings, and memory change the totals. Opting into more capability intentionally uses more context.

The control

Plain files. Visible choices.

Inspect the effective launch before it runs, or copy the managed defaults once and make them yours. Updates never overwrite user-owned configuration.

See the launch

doctor reports the effective prompt, settings, tools, capability state, memory source, and Claude Code checks.

claude-kiss doctor
CLAUDE_KISS_DRY_RUN=1 claude-kiss

Own the defaults

Run claude-kiss init. From then on, these three files belong to you:

~/.config/claude-kiss/
  prompt.md
  settings.json
  memory/CLAUDE.md

The research

A small relevant surface is a reliability choice.

Recent 2026 studies do not prove that one launcher is ideal. They do show that instruction count, visible tools, handoffs, and interface design change how coding agents behave.

  1. Across two scaffolds and four frontier models, the simpler mini-SWE-agent scaffold was more robust to semantics-preserving repository changes than OpenCode.

  2. Across 11,700 trajectories, useful structured tools improved consistency and Python CodeAct cut steps and tokens; scratchpad-style scaffolding had little effect.

  3. Across 247,694 instruction lifetimes, agent prompts grew 226%; preserving rationale enabled dramatic pruning and improved instruction following by up to 23.1%.

  4. For repository questions, direct retrieval scored 65.2% versus 46.2% for planner-to-subagent search; each correct answer cost less than half as much, and 41.8% of failures happened at handoff.

  5. Instruction adherence fell from about 96% with one simultaneous rule to as low as 20% with twenty. More permanent policy can mean less policy followed.

  6. An adaptive system exposed about seven tools instead of fifty while retaining nearly identical retrieval coverage: 90.3% versus 90.8%.

The limit is part of the claim. Structured tools can outperform Bash-only minimalism, and difficult work can justify orchestration. The evidence supports default-minimal and task-relative—not “complexity is always bad.”

Read all eleven studies and the caveats

Evidence

Measure the files, not the pitch.

The repository includes eight paired tasks: same fixture, prompt, model, and budget; ordinary claude versus claude-kiss. The evaluator checks the resulting files, not just the prose.

Read the evaluator Read the case

Trust boundary

Strong defaults are not a sandbox.

Bash remains powerful. User, project, and local settings—and their hooks and plugins—still load. Explicit native flags can replace managed defaults. Tool limits shape attention, not filesystem access. Review permissions, repository settings, plugins, and MCP configurations as code.

Understand the boundary

Start small

Read it. Install it. Ship.

curl -fsSL https://claude-kiss.com/install.sh | sh

Then run claude-kiss. Your ordinary claude command remains exactly as it was.