agent-friend

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.json

GitHub's official MCP: 20,444 tokens → ~14,000. Same tools. More accurate. No config.

Fix

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.

Grade

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: 4483

Notion'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.

Validate

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 — PASS

13 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.

Audit

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 tokens

Accepts 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).

Write once, deploy everywhere

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 Schema

One 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 format

CI / GitHub Action

Token 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 < 80

agent-friend grade tools.json --threshold 90  # exit code 1 if below 90
agent-friend audit tools.json --threshold 500  # exit code 2 if over budget

Pre-commit hook

Grade 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 error

Override the threshold:

      - id: agent-friend-grade
        args: ["--threshold", "80"]  # fail if score < 80

Claude Code hook

Auto-check grades when you add MCP servers to Claude Code:

mkdir -p ~/.claude/hooks
curl -sL https://0-co.github.io/company/claude-code-hook.sh -o ~/.claude/hooks/af-check.sh
chmod +x ~/.claude/hooks/af-check.sh

Add to ~/.claude/settings.json:

{
  "hooks": {
    "ConfigChange": [{
      "matcher": ".",
      "hooks": [{"type": "command", "command": "bash ~/.claude/hooks/af-check.sh"}]
    }]
  }
}

Now every time you add an MCP server to Claude Code, you see its grade. See Discussion #191 for details.

Start a new MCP server

Use mcp-starter — a GitHub template repo that scaffolds a new server pre-configured for A+. agent-friend pre-commit hook and CI grading included.

REST API

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.

# CI pass/fail check (200=pass, 422=fail)
curl "http://89.167.39.157:8082/v1/check?url=https://example.com/schema.json&threshold=80"

# README badge redirect (shields.io)
curl -L "http://89.167.39.157:8082/badge?repo=owner/repo-name"

Endpoints: /v1/grade, /v1/check?url=...&threshold=80, /v1/servers, /badge?repo=....

Also included

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.

Hosted version?

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.

Built by an AI, live on Twitch

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