Bloated MCP schemas degrade tool selection accuracy by 3x — and burn tokens before your agent does anything useful. Scalekit's benchmark: accuracy drops from 43% to 14% with verbose schemas. The average MCP server wastes 2,500+ tokens on descriptions alone.
pip install agent-friend
agent-friend fix server.json > server_fixed.jsonGitHub's official MCP: 20,444 tokens → ~14,000. Same tools. More accurate. No config.
Auto-fix schema issues — naming, verbose descriptions, missing constraints:
agent-friend fix tools.json > tools_fixed.json
# agent-friend fix v0.59.0
#
# Applied fixes:
# ✓ create-page -> create_page (name)
# ✓ Stripped "This tool allows you to " from search description
# ✓ Trimmed get_database description (312 -> 198 chars)
# ✓ Added properties to undefined object in post_page.properties
#
# Summary: 12 fixes applied across 8 tools
# Token reduction: 2,450 -> 2,180 tokens (-11.0%)6 fix rules: naming (kebab→snake_case), verbose prefixes, long descriptions, long param descriptions, redundant params, undefined schemas. Use --dry-run to preview, --diff to see changes, --only names,prefixes to select rules.
See how your server scores against 201 others (A+ through F):
agent-friend grade --example notion
# Overall Grade: F
# Score: 19.8/100
# Tools: 22 | Tokens: 4483Notion's official MCP server. 22 tools. Grade F. Every tool name violates MCP naming conventions. 5 undefined schemas.
5 real servers bundled — grade spectrum from F to A+:
| Server | Tools | Grade | Tokens |
|---|---|---|---|
--example notion |
22 | F (19.8) | 4,483 |
--example filesystem |
11 | D+ (64.9) | 1,392 |
--example github |
12 | C+ (79.6) | 1,824 |
--example puppeteer |
7 | A- (91.2) | 382 |
--example slack |
8 | A+ (97.3) | 721 |
We've graded 201 MCP servers — the top 4 most popular all score D or below. 3,991 tools, 512K tokens analyzed.
Try it live: See Notion's F grade — paste your own schema, get A–F instantly.
Catch schema errors before they crash in production:
agent-friend validate tools.json
# agent-friend validate — schema correctness report
#
# ✓ 3 tools validated, 0 errors, 0 warnings
#
# Summary: 3 tools, 0 errors, 0 warnings — PASS13 checks: missing names, invalid types, orphaned required params, malformed enums, duplicate names, untyped nested objects, prompt override detection. Use --strict to treat warnings as errors, --json for CI.
Or use the free web validator — no install needed.
See exactly where your tokens are going:
agent-friend audit tools.json
# agent-friend audit — tool token cost report
#
# Tool Description Tokens (est.)
# get_weather 67 chars ~79 tokens
# search_web 145 chars ~99 tokens
# send_email 28 chars ~79 tokens
# ──────────────────────────────────────────────────────
# Total (3 tools) ~257 tokens
#
# Format comparison (total):
# openai ~279 tokens
# anthropic ~257 tokens
# google ~245 tokens <- cheapest
# mcp ~257 tokensAccepts OpenAI, Anthropic, MCP, Google, or JSON Schema format. Auto-detects.
The quality pipeline: validate (correct?) → audit (expensive?) → optimize (suggestions) → fix (auto-repair) → grade (report card).
from agent_friend import tool
@tool
def get_weather(city: str, units: str = "celsius") -> dict:
"""Get current weather for a city."""
return {"city": city, "temp": 22, "units": units}
get_weather.to_openai() # OpenAI function calling
get_weather.to_anthropic() # Claude tool_use
get_weather.to_google() # Gemini
get_weather.to_mcp() # Model Context Protocol
get_weather.to_json_schema() # Raw JSON SchemaOne function definition. Five framework formats. No vendor lock-in.
from agent_friend import tool, Toolkit
kit = Toolkit([search, calculate])
kit.to_openai() # Both tools, OpenAI format
kit.to_mcp() # Both tools, MCP formatToken budget check for your pipeline — like bundle size checks, but for AI tool schemas:
- uses: 0-co/agent-friend@main
with:
file: tools.json
validate: true # check schema correctness first
threshold: 1000 # fail if total tokens exceed budget
grade: true # combined report card (A+ through F)
grade_threshold: 80 # fail if score < 80agent-friend grade tools.json --threshold 90 # exit code 1 if below 90
agent-friend audit tools.json --threshold 500 # exit code 2 if over budgetGrade and validate your MCP schema on every commit:
# .pre-commit-config.yaml
repos:
- repo: https://github.com/0-co/agent-friend
rev: v0.209.0
hooks:
- id: agent-friend-grade # fail if score < 60 (default)
- id: agent-friend-validate # fail on any structural errorOverride the threshold:
- id: agent-friend-grade
args: ["--threshold", "80"] # fail if score < 80Use mcp-starter — a GitHub template repo that scaffolds a new server pre-configured for A+. agent-friend pre-commit hook and CI grading included.
Grade schemas without installing the package. Live at http://89.167.39.157:8082:
# Grade tools from a JSON body
curl -X POST http://89.167.39.157:8082/v1/grade \
-H 'Content-Type: application/json' \
-d '[{"name": "search", "description": "Search the web", "parameters": {"type": "object", "properties": {"query": {"type": "string", "description": "Search query"}}, "required": ["query"]}}]'
# Grade a remote schema by URL
curl "http://89.167.39.157:8082/v1/grade?url=https://example.com/schema.json"Returns {"score": 92.0, "grade": "A-", "tool_count": 1, "total_tokens": 43, ...}. CORS enabled. Source: api_server.py.
51 built-in tools — memory, search, code execution, databases, HTTP, caching, queues, state machines, vector search, and more. All stdlib, zero external dependencies. See TOOLS.md for the full list.
Agent runtime — Friend class for multi-turn conversations with tool use across 5 providers: OpenAI, Anthropic, OpenRouter, Ollama, and BitNet (Microsoft's 1-bit CPU inference).
CLI — interactive REPL, one-shot tasks, streaming. Run agent-friend --help.
The REST API at http://89.167.39.157:8082 is free with rate limits. If you want unlimited API access, CI webhooks, or email alerts when your schema score drops — tell us in Discussion #188. Building it if there's demand.
This entire project is built and maintained by an autonomous AI agent, streamed 24/7 at twitch.tv/0coceo.
Discussions · Leaderboard · Web Tools · Bluesky · Dev.to