Slack-native Claude Code runtime. One Slack thread = one persistent session.
Onboard an AI teammate into your Slack workspace. Inspired by hermes-agent, rebuilt Slack-only on the official @anthropic-ai/claude-agent-sdk. One container = one persona. Every thread remembers.
git clone https://github.com/barockok/slaude.git && cd slaude
bun install && bun run manifest > manifest.json # paste into api.slack.com/apps → From manifest
cp .env.example .env && docker compose up -d --build| Section | What you will find | Time |
|---|---|---|
| Getting Started | 5-min quickstart, prerequisites, first DM test | 5 min |
| Configuration | Env vars, Slack tokens, Anthropic auth, SOUL.md schema, slaude.json |
10 min |
| Architecture | Session lifecycle, trust boundary, gateway ↔ agent ↔ MCP, persistence | 8 min |
| Guides | Engagement, approvals, /1on1, slash commands, cron, attachments |
— |
| API Reference | mcp__slaude_* tools, Markdown→mrkdwn, external MCP, skills, KB |
— |
| Deployment | Docker, Kubernetes, scale-out (gateway + node workers), health probes, upgrades, troubleshooting | — |
| Examples | Runnable: first skill, first ingest, custom MCP, sim without Slack | — |
Full site: https://barockok.github.io/slaude/ — source in docs/site/_content/, built by docs/site/build.mjs and published from main by CI. The generated HTML is not committed; run bun run docs to build it into docs/site/_site/.
Source: docs/internal/architecture.html — regenerate PNG via headless Chrome (see file header).
| Feature | What it means |
|---|---|
| Thread = Session | SDK resume: keeps history across idle restarts (SLAUDE_IDLE_MINUTES default 15). |
| Two-layer persona | Hardcoded baseline (output, engagement, approval discipline) + your SOUL.md (name, role, mandate, approvers). |
| MCP-only Slack I/O | Every reply via mcp__slaude_slack__reply / edit / upload / request_approval. Markdown auto-converts to mrkdwn. |
| Approvals | Agent runs YOLO, calls request_approval() for mutating work. Allowlist parsed from SOUL.md — agent never picks IDs. |
| Engagement | @mention engages thread, mentioning someone else disengages. DMs always engaged. /1on1 locks thread. |
| Skills & KB | ~/.slaude/skills/<slug>/SKILL.md hot-reloads; knowledge/ wikis compound via raw/ → wiki/ /ingest. |
| Simulation | bun run sim verifies gates with no Slack — same createGateway, in-memory transport. |
Trust boundary: LLM extracts SOUL.md → typed JSON → sha-cached; gateway verifies every Slack ID against raw text before any gate. Enforcement lives in gateway, never model. See Architecture.
# 1. Slack app from manifest
bun run manifest > manifest.json
# → api.slack.com/apps → Create New App → From manifest → paste
# → App-Level Tokens → connections:write (xapp-…) → Install to workspace (xoxb-…)
# 2. Env
cp .env.example .env # fill SLACK_BOT_TOKEN, SLACK_APP_TOKEN, ANTHROPIC_API_KEY or CLAUDE_CODE_OAUTH_TOKEN
# 3. Run
bun run dev # local — or —
docker compose up -d --build
curl localhost:8080/healthz # {"status":"ok"}Next: Getting Started → Configuration → Deployment
- Releases:
docs/site/_content/releases/— hand-written notes per tag (Features / Fixes / Docs) - Findings:
docs/site/_content/field-notes/— decisions & incident lore - Contributing:
CONTRIBUTING.md· Security:SECURITY.md· License: MIT
slaude v0.44.0 · GitHub · Built with Bun + TypeScript + claude-agent-sdk