agent-tools

Shared MCP servers and standard runtime for Dan's Slack-driven agents.

Why this exists

All 11 agents (fpa, cos, strategy, revops, video, web, content-review, pmm, marketing-recruiter, social, ea) historically copy-pasted the same tool modules and run loop. Bugs got fixed in one repo and never made it to the others. Slack reliability suffered as a result.

This package centralizes:

• Standard runtime (agent_tools.runtime): one wrapper around ClaudeSDKClient that every agent uses. Handles the agent loop, error-into-Slack hook, telemetry, memory injection, and the claude_code system-prompt preset.
• MCP servers (agent_tools.slack, agent_tools.snowflake, agent_tools.notion, agent_tools.drive, ...): one canonical implementation per service, imported by every agent that needs it.
• Smoke-test harness (agent_tools.smoke): pre-deploy probe that catches token expiry, schema drift, and MCP misconfiguration before users hit it in Slack.

Layout

agent_tools/
  runtime.py            # ClaudeSDKClient wrapper, run_ask() entry point
  smoke.py              # smoke-test harness
  slack/
    client.py           # Slack API wrapper (lifted from fpa-agent-cloud)
    mcp.py              # MCP server: post_in_thread, list_channel_messages, get_thread
  snowflake/
    client.py           # Snowflake key-pair auth + query()
    mcp.py              # MCP server: snowflake_query
  notion/
    client.py           # Notion search/read/write helpers
    mcp.py              # MCP server: notion_search, notion_get_page, notion_query_database, notion_create_db_row, notion_update_page, notion_append_to_page, notion_replace_page_content
  drive/
    client.py           # Google Drive read-only via OAuth user token
    mcp.py              # MCP server: drive_search_files, drive_get_file_metadata, drive_read_file_content, drive_list_recent_files

Usage from an agent

from agent_tools.runtime import run_ask
from agent_tools.slack.mcp import slack_server
from agent_tools.snowflake.mcp import snowflake_server
from agent_tools.notion.mcp import notion_server
from agent_tools.drive.mcp import drive_server

SYSTEM_PROMPT = (Path(__file__).parent / "system_prompt.md").read_text()

def handle_slack_message(text, channel_id, thread_ts, thread_context=None):
    return run_ask(
        text=text,
        channel_id=channel_id,
        thread_ts=thread_ts,
        system_prompt_append=SYSTEM_PROMPT,
        mcp_servers={
            "slack": slack_server,
            "snowflake": snowflake_server,
            "notion": notion_server,
            "drive": drive_server,
        },
        thread_context=thread_context,
    )

Installation

This is a private package, not on PyPI. Each agent adds it to its requirements.txt as a git dependency:

agent-tools @ git+https://github.com/danstotts-ops/agent-tools.git@main

For local development, install in editable mode:

pip install -e ~/agent-tools

Environment variables

Each MCP server reads its own credentials from env. Agents must set these in .env (local) or Railway env (prod):

• Slack: SLACK_BOT_TOKEN, SLACK_USER_TOKEN
• Snowflake: SNOWFLAKE_ACCOUNT, SNOWFLAKE_USER, SNOWFLAKE_PRIVATE_KEY, SNOWFLAKE_PRIVATE_KEY_PASSPHRASE, SNOWFLAKE_ROLE, SNOWFLAKE_WAREHOUSE, SNOWFLAKE_DATABASE
• Notion: NOTION_API_TOKEN
• Drive: DRIVE_ACCESS_TOKEN, DRIVE_REFRESH_TOKEN, DRIVE_OAUTH_CLIENT_ID, DRIVE_OAUTH_CLIENT_SECRET

What's NOT here yet

These MCP servers will be added in subsequent phases as agents need them:

• hubspot — HubSpot via Snowflake mirror (currently in fpa-agent-cloud)
• posthog — PostHog product analytics (fpa, revops)
• ramp — Ramp expense data (fpa, cos)
• linkedin_ads — LinkedIn ads spend (fpa, social)
• gong — Gong call transcripts (revops, marketing-recruiter)
• fireflies — Fireflies meeting summaries (cos, ea)
• gmail — Gmail read/draft (ea)
• calendar — Google Calendar (ea, cos)
• mode — Mode Analytics SQL pulls (revops, fpa)

The pattern is identical: lift the existing *_tools.py from whichever agent originally implemented it, drop it in agent_tools/<service>/client.py, write the mcp.py wrapper.

Status

Phase 1 (in progress). Foundation + slack/snowflake/notion/drive servers. fpa-agent will be the first agent migrated onto this package as proof-of-pattern; the rest follow in their own PRs.