A self-hosted Slack assistant built on Agentic API, adapted from EmbeddedLLM/slack-pi-agent-app. All inference uses the native WebSocket Responses API.
- Slack mentions, DMs, threaded memory, role permissions, job limits and cancellation.
- You.com web search through the gateway, with cited results.
multi <task>for gateway subagents and multi-agent research.steer <update>for an active standard request, with successor tracking.- Gateway Python code interpreter for calculations and analysis.
- Native text/image inputs, optional PDF/Office extraction, and role-scoped file search.
- Research plans, progress, saved sections, resume, and downloadable reports.
- Text- or image-plus-prompt GIF stickers, rendered locally as small transparent animations.
- A VitePress documentation website with local search and hosting instructions.
Steering currently supports single-agent chat only. Gateway Python returns text from fresh sandboxes; it cannot exchange files or create downloadable artifacts. Knowledge uploads/indexing use HTTP APIs; inference remains on WebSocket.
Requirements: Node.js 22.15+, Rust 1.98.1 for the gateway source build, a tool-capable model server, and a Slack app. Interpreter execution requires Linux cgroup v2 delegation; see gateway setup.
git clone https://github.com/EmbeddedLLM/slack-agentic-api-app.git
cd slack-agentic-api-app
npm ci
npm run gateway:install
cp deploy/gateway.env.example .env.gateway
cp .env.example .env
cp access.example.json access.json
chmod 600 .env .env.gateway access.jsongateway:install builds Agentic API main from source, including embedded-code-interpreter and file-search-pdf. Source and binary live under .runtime/agentic-api; the installed commit is recorded in .runtime/gateway-source-commit.
Set YOU_API_KEY in .env.gateway. Defaults match vLLM at http://127.0.0.1:5050 and gateway port 9000. Start the source gateway in a separate terminal with npm run gateway:start. If port 9000 already has a gateway, deliberately stop it or select another port first.
Create the Slack app using slack-app-manifest.json; see Slack setup. Put the xoxb- bot token and xapp- app token in .env, set the exact served model ID, and add your Slack member ID to access.json. Unlisted users are denied.
AGENTIC_WS_URL=ws://127.0.0.1:9000/v1/responses
AGENTIC_MODEL=Qwen/Qwen3.8-27B-FP8npm run smoke
npm startSlack Socket Mode needs no public URL or inbound listener. Keep the gateway private, or use wss:// and gateway authentication for a remote connection.
- Host the bot on this machine with systemd
- Build and configure the source gateway
- Multi-agent, steering and tools
- Input files and searchable knowledge
- Research and recovery
- Create GIF stickers
- Configuration and operations
- Protocol and architecture
npm run docs:dev # local documentation website
npm run docs:build # static site in docs/.vitepress/distThe manual GitHub Pages workflow is prepared for a later publication; a local build does not deploy it.
npm test
AGENTIC_SERVER_BIN=.runtime/agentic-api/target/release/agentic-server node scripts/gateway-integration.jsTests cover WebSocket contracts and Slack handlers without live Slack credentials. The integration script uses a real source-built gateway with isolated storage and a deterministic model stub. Live model quality, You.com credentials, a Slack installation and Chromium rendering need deployment acceptance checks. See development.
Apache-2.0. Attribution and upstream source commits are recorded in NOTICE.