skills-not-mcp

80-90% fewer tokens with Skills(/) + shell vs MCP

About

80-90% token savings by replacing MCP with Skills + Shell scripts.

18 MCP tools from pal-mcp-server converted to:

Claude Skills (commands/*.md) - slash command interface
Shell scripts (scripts/*.sh) - actual execution

approach	tokens loaded	savings
MCP (18 tools)	~12,500 always	baseline
Skills + Shell	~50 per call	80-90%

MCP loads all 18 tool schemas (~700 tokens each) into context whether you use them or not.

Skills + Shell loads nothing until called - then only the output.

Conversion Table

#	MCP (before)	Skill (after)	Shell (after)
1	`tools/chat.py`	`commands/chat.md`	`scripts/chat.sh`
2	`tools/thinkdeep.py`	`commands/thinkdeep.md`	`scripts/thinkdeep.sh`
3	`tools/consensus.py`	`commands/consensus.md`	`scripts/consensus.sh`
4	`tools/codereview.py`	`commands/codereview.md`	`scripts/codereview.sh`
5	`tools/debug.py`	`commands/debug.md`	`scripts/debug.sh`
6	`tools/planner.py`	`commands/planner.md`	`scripts/planner.sh`
7	`tools/analyze.py`	`commands/analyze.md`	`scripts/analyze.sh`
8	`tools/refactor.py`	`commands/refactor.md`	`scripts/refactor.sh`
9	`tools/testgen.py`	`commands/testgen.md`	`scripts/testgen.sh`
10	`tools/secaudit.py`	`commands/secaudit.md`	`scripts/secaudit.sh`
11	`tools/docgen.py`	`commands/docgen.md`	`scripts/docgen.sh`
12	`tools/apilookup.py`	`commands/apilookup.md`	`scripts/apilookup.sh`
13	`tools/challenge.py`	`commands/challenge.md`	`scripts/challenge.sh`
14	`tools/tracer.py`	`commands/tracer.md`	`scripts/tracer.sh`
15	`tools/clink.py`	`commands/clink.md`	`scripts/clink.sh`
16	`tools/precommit.py`	`commands/precommit.md`	`scripts/precommit.sh`
17	`tools/listmodels.py`	`commands/listmodels.md`	`scripts/listmodels.sh`
18	`tools/version.py`	`commands/version.md`	`scripts/version.sh`

Source: pal-mcp-server

Pattern

Deterministic Delegation - Skills orchestrate, Shell executes

┌─────────────────────────────────────────────────────────────┐
│                    MCP APPROACH                             │
│                                                             │
│  ┌────────┐ ┌────────┐ ┌────────┐ ┌────────┐              │
│  │ tool 1 │ │ tool 2 │ │ tool 3 │ │ ...18  │              │
│  │ ~700   │ │ ~700   │ │ ~700   │ │ tokens │              │
│  └────────┘ └────────┘ └────────┘ └────────┘              │
│                                                             │
│  COST: ~12,500 tokens ALWAYS LOADED                        │
│  SAVINGS: 0%                                                │
└─────────────────────────────────────────────────────────────┘

┌─────────────────────────────────────────────────────────────┐
│            SKILLS + SHELL (Deterministic Delegation)        │
│                                                             │
│  ┌────────┐                                                 │
│  │ /skill │ ────> shell script ────> output                │
│  │ ~50 tok│       (external)         (only this)           │
│  └────────┘                                                 │
│                                                             │
│  COST: ~50 tokens + output only                            │
│  SAVINGS: 80-90%                                            │
└─────────────────────────────────────────────────────────────┘

flow:

USER                    AI                      SHELL
 │                       │                        │
 │  "/chat review this"  │                        │
 │ ───────────────────>  │                        │
 │                       │  exec chat.sh          │
 │                       │ ────────────────────>  │
 │                       │                        │
 │                       │      ┌─────────────────┤
 │                       │      │ - parse args    │
 │                       │      │ - detect model  │
 │                       │      │ - call API      │
 │                       │      │ - format output │
 │                       │      └─────────────────┤
 │                       │                        │
 │                       │  JSON result           │
 │                       │ <────────────────────  │
 │                       │                        │
 │  formatted response   │                        │
 │ <───────────────────  │                        │

layers:

┌────────────────────────────────────────────────┐
│  LAYER 1: SKILL (orchestrator)                 │
│  commands/*.md                                 │
│  - slash commands, routing, ~50 tokens each    │
└────────────────────────────────────────────────┘
                    │ delegates to
                    ▼
┌────────────────────────────────────────────────┐
│  LAYER 2: SHELL (execution)                    │
│  scripts/*.sh                                  │
│  - actual logic, 0 tokens til called           │
└────────────────────────────────────────────────┘
                    │ calls
                    ▼
┌────────────────────────────────────────────────┐
│  LAYER 3: PROVIDER (external)                  │
│  Anthropic / OpenAI / Ollama / etc             │
└────────────────────────────────────────────────┘

property	MCP	Skills + Shell
schema loading	always (~12.5k)	never
token savings	0%	80-90%
execution	probabilistic	deterministic
debugging	opaque	`bash -x script.sh`

How to build

git clone https://github.com/vdutts7/skills-not-mcp.git
cd skills-not-mcp/scripts

API keys:

export ANTHROPIC_API_KEY="sk-..."
export CEREBRAS_API_KEY="..."
# or local Ollama (no key)

Usage

# claude
./chat.sh -m claude-sonnet-4-20250514 "structure this API?"

# ollama (free/local)
./chat.sh -m qwen2.5-coder:7b "review this"

# file context
./chat.sh -f src/main.py -f src/utils.py "explain flow"

# multi-turn
./chat.sh -c sess1 "design cache"
./chat.sh -c sess1 "redis vs memcached?"

# code review
./codereview.sh -m llama-3.3-70b -f src/auth.py

# demo
./demo.sh

provider	models	env
Anthropic	sonnet-4, opus-4	`ANTHROPIC_API_KEY`
Cerebras	llama-3.3-70b, qwen-3-32b	`CEREBRAS_API_KEY`
Ollama	qwen2.5-coder, llama, mistral	local
OpenAI	gpt-4o, o1, o3	`OPENAI_API_KEY`
Gemini	2.0-flash	`GEMINI_API_KEY`
OpenRouter	any	`OPENROUTER_API_KEY`

skills-not-mcp

skills-not-mcp

Table of Contents

About

Conversion Table

Pattern

How to build

Usage

Tools

Contact

Reviews (0)