agent-friend 0.173.0

MCP schema linter & quality grader — validate, audit, optimize, grade (A+ to F). Also: @tool exports to OpenAI, Claude, Gemini, MCP.

agent-friend

The quality linter for MCP tool schemas. Validate, audit, optimize, fix, and grade (A+ through F). Like ESLint for MCP. Also: write a tool once, export to OpenAI, Claude, Gemini, or MCP.

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.

Install

pip install agent-friend

Grade a real MCP server (no API key, no schema file)

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 — 74 checks applied, no setup required:

Server	Tools	Score	Tokens
--example notion	22	F (19.8)	4,483
--example filesystem	11	F (24.9)	1,392
--example github	12	F (39.6)	1,824
--example puppeteer	7	F (55.2)	382
--example slack	8	F (57.3)	721

74 checks covering naming, descriptions, parameter schemas, token efficiency, and cross-tool consistency. Getting to A+ requires deliberate effort — most real-world servers don't.

We've graded 201 MCP servers — the top 4 most popular all score D or below. 3,978 tools, 512K tokens analyzed.

agent-friend examples  # list all bundled schemas

Or open the Colab notebook — 51 tool demos in the browser.

Batch export

from agent_friend import tool, Toolkit

@tool
def search(query: str) -> str: ...

@tool
def calculate(expr: str) -> float: ...

kit = Toolkit([search, calculate])
kit.to_openai()   # Both tools, OpenAI format
kit.to_mcp()      # Both tools, MCP format

Context budget

MCP tool definitions can eat 40-50K tokens per request. Audit your tools from the CLI:

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
#     json_schema   ~245 tokens
#
#   Context window impact:
#     GPT-4o (128K)       ~0.2%
#     Claude (200K)       ~0.1%
#     GPT-4 (8K)          ~3.1%  <- check your budget
#     Gemini 2.0 (1M)     ~0.0%

Or measure programmatically:

kit = Toolkit([search, calculate])
kit.token_report()

Accepts OpenAI, Anthropic, MCP, Google, or JSON Schema format. Auto-detects.

Optimize

Found the bloat? Fix it:

agent-friend optimize tools.json

# Tool: search_inventory
#   ⚡ Description prefix: "This tool allows you to search..." → "Search..."
#      Saves ~6 tokens
#   ⚡ Parameter 'query': description "The query" restates parameter name
#      Saves ~3 tokens
#
# Summary: 5 suggestions, ~42 tokens saved (21% reduction)

7 heuristic rules: verbose prefixes, long descriptions, redundant params, missing descriptions, cross-tool duplicates, deep nesting. Machine-readable output with --json.

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 — paste schemas, get instant results, no install needed.

Fix

Found issues? Auto-fix them:

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 pick rules.

The quality pipeline: validate (correct?) → audit (expensive?) → optimize (suggestions) → fix (auto-repair) → grade (report card).

Or get the full report card:

agent-friend grade tools.json

# agent-friend grade — schema quality report card
#
#   Overall Grade: B+
#   Score: 88.0/100
#
#   Correctness   A+  (100/100)  0 errors, 0 warnings
#   Efficiency    B-  (80/100)   avg 140 tokens/tool
#   Quality       B   (85/100)   1 suggestion
#
#   Tools: 3 | Format: anthropic | Tokens: 420

Weighted scoring: Correctness 40%, Efficiency 30%, Quality 30%. Use --threshold 90 to gate CI on quality, --json for machine-readable output.

Try it live: See Notion's F grade — or paste your own schemas. 5 real servers to try, share buttons, copy-paste badge for your README.

CI / GitHub Action

Add a token budget to your CI pipeline — like a bundle size check 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
    optimize: true        # also suggest fixes
    grade: true           # combined report card (A+ through F)
    grade_threshold: 80   # fail if score < 80

Runs the full quality pipeline: validate → audit → optimize → fix → grade. Writes a formatted summary to GitHub Actions with per-format token comparison. Use CLI flags too:

agent-friend audit tools.json --json              # machine-readable output
agent-friend audit tools.json --threshold 500      # exit code 2 if over budget

When you need this

• You're writing tools for one framework but want them to work in others
• You want to define a tool once and use it with OpenAI, Claude, Gemini, AND MCP
• You need the adapter layer, not an opinionated orchestration framework
• You want MCP tools in Claude Desktop — agent-friend ships an MCP server with 314 tools

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.

Why not just use [framework X]?

Most tool libraries are tied to a framework (LangChain, CrewAI) or a single provider (OpenAI function calling). If you switch providers, you rewrite your tools.

agent-friend decouples your tool logic from the delivery format. Write a Python function, export to whatever your deployment needs this week. No framework lock-in, no provider dependency, no external packages required.

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 · Website · Bluesky · Dev.to