daimon

daimon is an open source data science agent that joins your team's Discord, Slack, or Microsoft Teams workspace and answers analysis questions directly in the thread. Mention it in a channel and it writes and runs code, fits Bayesian models with PyMC, and replies with charts, an uncertainty read, and a runnable marimo notebook that reproduces the result. Unlike general-purpose chatbots, it runs on Anthropic's Managed Agents with one tenant per server or workspace, so a single self-hosted deployment on your own API key can serve many communities with isolated memory and data. Anyone in the thread can follow up, and scheduled routines let recurring analyses post themselves where you ask.

Features

• Thread-native answers. @mention the bot in a channel and it opens a thread, runs the analysis, and replies with a fitted model, a chart, a plain-English read on the uncertainty, and a marimo notebook that reproduces the result.
• Shared context for the whole team. Anyone in the server can follow up in the same thread; sessions keep their context across turns.
• Scheduled routines. Recurring analyses run headlessly and post results where you ask.
• Connects to your tools. Ask it to connect Notion, Linear or any other MCP server. Token-based servers are shared per agent; OAuth servers are connected per person. GitHub repos connect through the GitHub App: ask it to connect GitHub, or run /github.
• Reachable from Claude Code. The bundled plugin logs in through Slack or Discord OAuth and exposes every daimon you belong to as an MCP server.
• One deployment, many communities. Deploy once on your own Anthropic API key. Any number of Discord servers and Slack workspaces can install it, each as an isolated tenant with its own agent, memory and data. On Teams, a deployment serves one Microsoft 365 organisation.

Conversation is the primary interface. Slash commands exist for people who prefer them:

/here          /agent-setup    /github
/routines      /billing        /privacy
/help

On Teams these are plain words (setup, routines, help) sent in a 1:1 chat with the bot.

Example prompts

• "Here's last quarter's sales export. We changed pricing in week 6. Did it actually help?"
• "Is variant B actually better than A, or is that just noise?"
• "Forecast next month's signups, with uncertainty bands."
• "Every Monday at 9am, pull the weekend's numbers and post a summary here."

How it works

daimon is built on Anthropic Managed Agents. Every adapter — Discord, Slack, Teams, the CLI, the MCP server and the routines scheduler — funnels into the same turn pipeline, so a conversation behaves the same regardless of where it started.

flowchart LR
    subgraph adapters
        direction TB
        Discord
        Slack
        Teams
        CLI
        MCP
        Scheduler
    end
    adapters --> core["daimon core<br>turn pipeline"]
    core <--> ma["Anthropic Managed Agents<br>agents · sessions · skills"]
    core --> pg[("Postgres<br>tenants · thread↔session map")]

A single turn moves through four stages:

• Identify the tenant. The adapter derives the tenant from platform identity — a Discord guild, a Slack workspace, or a Microsoft 365 organisation.
• Open or resume a session. Core looks up the thread-to-session mapping in Postgres and either resumes the existing Managed Agents session or starts a new one.
• Stream events. Core subscribes to the session's event stream and forwards deltas as they arrive.
• Render in-thread. The adapter formats deltas for its platform and keeps posting until the session goes idle.

Design boundaries

• daimon.core owns the schema, stores and turn pipeline. It imports no adapters.
• Each adapter owns exactly one platform's I/O and auth. Adapters never import each other.
• import-linter enforces both rules in CI, so a new adapter can't quietly couple to another.

Managed Agents holds the agents, environments, sessions and skills. Postgres holds only metadata:

Because isolation is enforced at the database tenant_id layer, a single Anthropic API key can safely serve every install on a deployment — each community sees only its own agents, memory and data.

Extending it

The packages/mux/ package defines a provider-agnostic managed-agent interface, so adapters and core code against an abstract provider rather than Anthropic specifics. Adding a new chat platform means implementing one adapter; adding a new model backend means implementing one mux provider.

Requirements

daimon runs as a multi-service Docker Compose stack. Before you deploy, make sure the host and the accounts you connect satisfy the following.

Host

Credentials and secrets

Slack and Microsoft Teams adapters are optional and have their own setup requirements, covered in the Slack setup and self-hosting guide.

Chat platform setup

• Discord: create a bot in the Developer Portal, enable the Message Content Intent, and invite it to a server you control.
• Slack: a Slack app with the scopes and event subscriptions listed in the Slack setup guide.
• Microsoft Teams: a Microsoft 365 organisation; one deployment serves one organisation.

Isolation model

One Discord guild, Slack workspace or Microsoft 365 organisation is one tenant. Tenant isolation is enforced at the database tenant_id layer, so a single Anthropic key can serve multiple installs, but each install should still be treated as its own deployment target.

Configuration

daimon is configured through layered settings: .env / environment variables, a daimon.yaml defaults file, and per-tenant records stored in Postgres. Anything set at the tenant layer overrides the deployment-wide default, so one deployment can serve every install with the same key.

Environment variables

Every setting is namespaced under DAIMON_ and uses a double underscore to separate sections, e.g. DAIMON_DISCORD__BOT_TOKEN maps to discord.bot_token.

Start from .env.example, which documents each key inline with placeholder values.

Agent defaults

The YAML files in defaults/ seed the agents and skills that Managed Agents and the database receive on first boot. Each agent entry describes the model to use, the skills it can load, and the platform channels that route to it.

agents:
  analyst:
    model: <model-id>
    skills:
      - bayesian-modelling
      - forecasting
    channels:
      discord: []
      slack: []

Edit the defaults file, then re-run the init service so the new va