sladk-agents helps you run AI agents in your Slack workspace. It is powered by Google ADK and Bolt for Python. Multi-agent assistant in the side panel, threads, @mentions, and DMs, with streaming and tools.

• Slack-native - Uses Slack’s AI Agent surfaces (side panel, threads, @mentions, DMs).
• Google ADK - Root agent (Gemini) + sub-agents (e.g. search, code execution) and custom tools.
• Sessions - Conversation state per thread; configurable context compaction.
• Streaming - Responses streamed in real time.

Prerequisites: uv, Slack workspace (admin), Google API key with Gemini.

git clone https://github.com/jonigl/sladk-agents.git
cd sladk-agents
cp .env.sample .env
# Edit .env: SLACK_APP_TOKEN, SLACK_BOT_TOKEN, GOOGLE_API_KEY, AGENT_MODEL (e.g. gemini-2.5-flash)
cp AGENTS.md.sample AGENTS.md
# Edit AGENTS.md to define your agent's persona (optional — a default is used if omitted)
uv sync

uv manages the virtual environment and dependencies automatically — no manual venv or pip install needed.

Slack app: Create an app from manifest → paste manifest.json → Install to workspace. See SLACK_BOLT_TEMPLATE_README.md if you need step-by-step.

Run:

uv run python app.py

Or using Slack CLI:

slack run

Two .env knobs control response length:

• AGENT_MAX_OUTPUT_TOKENS — hard cap on ADK/Gemini output tokens; set 0 to disable (default 0)
• AGENT_TARGET_OUTPUT_CHARS — character budget applied both to the concise-response guidance sent to the LLM and to each Slack streaming message (default 9000)

When a response exceeds the streaming budget, the app continues automatically in the same thread with a follow-up message.

Attachment handling (optional overrides): ATTACHMENT_MAX_CHAR_BUDGET (default 200000), ATTACHMENT_MAX_FILES (default 5), ATTACHMENT_DOWNLOAD_TIMEOUT (default 15 seconds).

Slack: Preferences → Navigation → App agents & assistants → enable Show app agents. Then use the agent via the side panel, @mention in a channel, or DM.

Example prompts: search the web, run Python snippets, or use built-in tools like weather (see ai/tools/custom_tools.py).

You can give the agent a completely different role or personality without touching any code — just drop an AGENTS.md file in the project root.

cp AGENTS.md.sample AGENTS.md
# Edit AGENTS.md with any role, e.g. a security & compliance reviewer or a legal assistant

The full file content becomes the system instruction sent to the LLM. Restart the app to pick up changes.

Resolution order (highest wins):

To load the file from a different path, set AGENTS_MD_PATH in your .env:

AGENTS_MD_PATH=/path/to/my-agent.md

AGENTS.md is git-ignored (like .env) so each deployment can have its own persona. Commit AGENTS.md.sample as a template.

Docker / Kubernetes — either bake AGENTS.md into the image or mount it as a volume / ConfigMap and set AGENTS_MD_PATH accordingly.

Add tools in ai/tools/custom_tools.py and register them on the agent in ai/llm_caller.py:

# ai/tools/custom_tools.py
def my_tool(param: str) -> str:
    """What the tool does. Args: param. Returns: result."""
    return result

# ai/llm_caller.py — add to tools=[...]
from ai.tools.custom_tools import get_weather, my_tool
tools=[get_weather, my_tool, AgentTool(agent=search_agent), ...]

Slack (UI) → Bolt app (Socket Mode, listeners) → Google ADK (LlmAgent + sub-agents + tools, session store) → Gemini API. Each Slack thread maps to one ADK session; responses are streamed back.

Search, weather tool, threads & mentions

Python code execution via sub-agent

• Memory Bank across sessions
• MCP (Model Context Protocol) tools
• Agent Engine / Cloud Run deployment
• Observability (e.g. OpenTelemetry)
• A2A protocol for multi-agent workflows

MIT

Built on Slack’s bolt-python-assistant-template and Google ADK. A capstone version lives on the kaggle-project branch and was also shared in the 5-Day AI Agents Intensive with Google capstone.

Made with ❤️ by jonigl