Slack

Search Slack conversations, read threads, and post messages.

Two tools: query_slack for searching and reading, update_slack for posting messages.

1from kern.agent import Agent
2from kern.context.slack import SlackContextProvider
3
4slack = SlackContextProvider()
5
6agent = Agent(
7    model=...,
8    tools=slack.get_tools(),
9)
10
11agent.print_response("What did the team discuss about the auth migration?")

The agent calls query_slack("auth migration"). Behind the scenes, a sub-agent searches messages, fetches relevant threads, and returns a synthesized answer.

When to use this vs SlackTools

Use SlackContextProvider when...	Use SlackTools directly when...
Slack is one of several context sources	Slack is the primary task surface
You want to reduce tool clutter (2 tools vs 12)	You need fine control over individual API calls
The agent should ask questions, not orchestrate API calls	You're building a Slack-specific agent

Setup

1export SLACK_BOT_TOKEN=xoxb-...

Or pass directly: SlackContextProvider(token="xoxb-...").

See SlackTools for OAuth scope requirements.

Example queries

Queries that work well give the provider a topic to search and context to narrow results:

Query	Why it works
"What did the team decide about billing migration?"	Topic-driven search with synthesis
"Summarize the thread where Alex discussed OAuth scopes"	Combines search + thread reading
"What open questions came up in #launch this week?"	Channel + time constraints
"Post a short update to #support saying the issue is resolved"	Clear channel + action

Queries that are too vague:

Less effective	Better
"Search Slack"	"Search for recent discussion about webhook retries"
"What happened?"	"What happened in #incidents about the API outage?"
"Post it"	"Post this summary to #engineering"

Configuration

Parameter	Type	Default	Description
`id`	`str`	`"slack"`	Changes tool names to `query_<id>` and `update_<id>`.
`model`	`Model`	`None`	Model for the sub-agents. Defaults to Kern's default model.
`read`	`bool`	`True`	Expose `query_slack`.
`write`	`bool`	`True`	Expose `update_slack`.
`enable_media_tools`	`bool`	`False`	Enables `download_file` for read tools and `upload_file` for write tools.
`mode`	`ContextMode`	`default`	`default` exposes both tools. `tools` exposes read-only SlackTools.

Read/write modes

Control what the agent can do:

1# Research agent: can search but not post
2slack = SlackContextProvider(write=False)
3
4# Notification agent: can post but not search
5slack = SlackContextProvider(read=False)

Use write=False for research, triage, audit, and eval agents. The read sub-agent physically cannot post because send_message isn't in its toolkit.

Multi-provider example

The main value of context providers is reducing tool surface when an agent has multiple sources:

1from kern.context.slack import SlackContextProvider
2from kern.context.gdrive import GoogleDriveContextProvider
3
4slack = SlackContextProvider(write=False)
5drive = GoogleDriveContextProvider()
6
7agent = Agent(
8    model=...,
9    tools=[*slack.get_tools(), *drive.get_tools()],
10    instructions="Use Slack for team discussion, Drive for docs. Note when they disagree.",
11)
12
13agent.print_response("What's the current auth spec and did engineering raise concerns?")

The agent sees 2 tools (query_slack, query_gdrive) instead of 20+.

Tips

Token naming: SLACK_BOT_TOKEN is preferred; SLACK_TOKEN is a fallback.
Private channels: The bot must be invited to access private channel history.
Thread context: Ask for "the thread" or "the decision," not just the parent message.
Channel names: Natural names like #general work. The sub-agent resolves them to IDs.
Write clarity: "Post this to #team" is safer than vague "send it."

Resources

wrench

SlackTools Reference

All 12 methods and OAuth scope requirements

book

Cookbook Example

Working example with search and posting