Skip to main content
An agent in Pinkfish is a configurable AI assistant. You define how it behaves with instructions, choose which model it runs on, pick the tools it can call (via MCP servers, attached workflows, or other agents), grant it access to resources (datastores, filestores, knowledge bases), and expose it through one or more channels (web chat, Slack, Microsoft Teams, a public URL). You save drafts while iterating and release published versions that become usable from Coworker. The /agents page showing a list of published agents with status badges, filters, search, and a Create new agent button

​
What’s in this section

My Agents

The /agents page — list, filter, search, and monitor all agents in your workspace.

Agent Builder

Anatomy of the builder: tabs on the left, live preview on the right, save and release controls on top.

Instructions

Name, icon, description, system prompt, model choice, and prompt bubbles.

Tools

MCP connections, embedded services, agents-as-tools, and workflows-as-tools.

Resources

Attach datastores, filestores, and knowledge bases with Read / Write / Admin permissions.

Channels

Publish your agent to Slack, Microsoft Teams, or a public External URL.

Monitor

Replay an agent’s chats, inspect tool calls, and debug behavior from inside the builder.

Advanced

Output Schema — force structured JSON responses when calling the agent via API.

Save & Release

The draft-vs-release model, how releases reach Coworker, and how to roll back.

MCPs

How agents discover and invoke MCP tools from your integrations.

​
Agent types

Pinkfish supports two agent types:
  • Chat — A conversational agent that responds to messages. Use for assistants, researchers, and copilots.
  • Workflow — An agent bound to a specific workflow. Use when you want a workflow run to be triggered conversationally.
Both types use the same builder.

​
Where agents run

An agent can answer messages from any of these places once you release it:
  • Coworker — The chat tab inside Pinkfish. @-mention the agent in a Coworker chat.
  • Slack — Connect a Slack workspace in the Channels tab.
  • Microsoft Teams — Connect a Teams tenant in the Channels tab.
  • External (public URL) — A public link that anyone can chat with, with no Pinkfish login.
  • API — Programmatic access with an agent API key. See Agent Management API.

​
How the pieces fit together

1

Write instructions

On the Instructions tab, give the agent a name, an icon, a description users will see, a system prompt that defines its behavior, and a model. Optionally add prompt bubbles — one-click conversation starters.
2

Pick tools

On the Tools tab, select MCP servers (integrations), Embedded Services, other agents you want this agent to be able to delegate to, and workflows it can trigger. Pinkfish adds a few Auto-selected Tools for every agent.
3

Attach resources

On the Resources tab, grant read/write/admin access to specific datastores, filestores, and knowledge bases. Required tools are added to the Tools tab automatically.
4

Pick channels

On the Channels tab, enable Slack, Microsoft Teams, or make the agent public via External. A published release is required before External works.
5

Test in preview

The right-hand Preview pane is a live chat against your current draft. Click Information in the preview to double-check the selected model, tools, and resources.
6

Save, then release

Save commits a draft. Release publishes an immutable snapshot that becomes available in everyone’s Coworker tab (subject to sharing).

​
Agent Memory

Agent Memory

Store and recall user preferences and context across conversations and sessions.

​
Sharing agents

Agents can be shared with other people in your org via the agent’s Share action. Recipients get access to chat with the agent, but they must connect their own integrations — Pinkfish does not share your OAuth tokens. See the API reference for programmatic sharing.