Your personal AI orchestrator.
Turn Discord into a persistent AI workspace.
View on GitHub
Not a chatbot. An orchestrator.
Unlike chat UIs that forget you after every session, DiscoClaw is a persistent layer between Discord and your AI runtimes. It carries live context, durable memory, and cold-storage recall across conversations, manages tasks, runs automations, and routes each job to the right model — all from your own dedicated private server. It works while you sleep.
Rolling summaries, durable facts, per-channel context, and cold storage semantic recall over past conversations. Memory stays useful after the chat scroll is long gone.
Track work in Discord forum threads with bidirectional sync. Create, update, and close tasks without leaving chat.
Scheduled automations and webhook-triggered actions defined in plain language. Morning briefings, repo checks, CI hooks — on a timer or on demand.
Chain AI reasoning, shell commands, and Discord actions into multi-step workflows. Each step's output feeds the next. Analyze → act → notify — all composable, all auditable.
Claude, Codex, OpenAI, Gemini, OpenRouter. Route different roles to different runtimes, keep voice on a faster model, and spend the expensive tokens where they actually matter.
One API key, 200+ models. A unified gateway to hundreds of open and proprietary models via a single OpenAI-compatible endpoint. Swap what you're running without changing your config.
WebFetch to Chrome DevTools — a full automation ladder. Research, screenshot, scrape, interact.
No databases. State is JSON and markdown you can actually read. Backups are cp ‑r. Debug with cat. Intentionally simple — because for a single-user tool, it's all you need.
Portable, auditable integration specs in markdown. Build a workflow, export it as a recipe, share it. Another DiscoClaw instance picks it up and runs it — no code required.
Safety rules live at Layer 1 — below prompts, user instructions, and injected content. Prompt injection is an architectural concern, not a policy one: fetched web pages, emails, file contents, and tool results are treated as data, never commands. No injected instruction — however it arrives — can override the boundary.
Join Discord voice channels and participate in spoken conversations. Ask questions, run automations, get answers — without typing a word.
Dispatch parallel sub-agents to handle multiple tasks at once. One request fans out to as many agents as you need — each working independently, each posting its own results.
DiscoClaw is a foundation. Describe what you need, and your AI builds it — with the tools already in the box. The only limit is what you ask for.
DiscoClaw lives in your own private Discord server — a personal AI workspace that's entirely yours. Not limited to single user. Designed for it. No auth layers, no permission matrices, no tenant isolation. Just full trust between you and your machine.
Channels give you natural context separation — code in one, finances in another, personal projects in a third. Threads for deep dives. Full history and search built in. Rich embeds, file sharing, mobile push notifications. And it runs everywhere: web, Mac, Windows, Linux, iOS, Android. Why build a worse version of something that already exists?
JSON, markdown, flat files. You can read the state with cat. You can back it up with cp. No migrations, no schemas.
No plugin marketplace. No integration catalog. DiscoClaw gives you the primitives — memory, pipelines, runtimes, tools — and you tell it what to build. Describe what you want in conversation and it builds it with you. Share what you've built as recipes — portable, auditable integration specs that other DiscoClaw instances can pick up and run.
Runs on your hardware. Memory stays local. Conversations stay inside your own Discord server. No cloud services between you and your AI beyond the API calls themselves.
Three commands. Two env vars. Zero infrastructure.
# Install globally
$ npm install -g discoclaw
# Interactive setup — creates .env, scaffolds your workspace
$ discoclaw init
# Register the system service (Linux / macOS)
$ discoclaw install-daemonNew to Discord bots? Follow these steps to create your private DiscoClaw workspace from scratch.
Click the + button at the bottom of Discord's server list to create a new server. Use this dedicated server for DiscoClaw — don't start in a shared or public server.
Go to the Discord Developer Portal and create a new application. Give it a name (e.g. "DiscoClaw").
This is the #1 setup failure mode. Without this, the bot connects and appears online,
but msg.content will be empty — the bot silently ignores every message.
On the Bot page, click Reset Token (or copy it if you just created the app).
Paste it into your .env file immediately.
DISCORD_TOKEN=paste-your-token-here Use the Installation page in the Developer Portal to create a permanent, reusable invite link with permissions baked in.
Enable Developer Mode in Discord to copy IDs by right-clicking.
Then right-click your username → Copy User ID, and right-click a channel → Copy Channel ID.
DISCORD_ALLOW_USER_IDS=your-user-idDISCORD_CHANNEL_IDS=your-channel-id Run the smoke test to verify your token and connection, then DM the bot to confirm it responds.
pnpm discord:smoke-test Latest npm release
2026-04-01
Smarter memory management. Capsules now self-invalidate when idle too long or past their TTL, freeing context window space automatically.
Backfilled release notes for the npm cuts so far. Newest first. Scroll the archive if you want the full trail.
Smarter memory management. Capsules now self-invalidate when idle too long or past their TTL, freeing context window space automatically.
Action guard upgrade and docs polish. The output guard now catches past-tense false completion claims, and the README is condensed for the npm landing page.
Self-management docs and Discord action hardening. The bot can now update itself via !update, and action routing is more resilient.
Smarter recall. Cold storage queries now use HyDE (Hypothetical Document Embedding) to generate a synthetic ideal answer before searching, dramatically improving retrieval for vague or conceptual queries.
Lighter context window. History images are no longer injected into the prompt, reducing token overhead without losing image awareness.
Instruction extraction. Core behavioral rules are now in a standalone SYSTEM_DEFAULTS.md file, making them easier to audit, override, and share across workspace configurations.
Silent cron fix. Error messages from failed cron jobs are no longer posted to Discord when the cron has silent mode enabled.
Focus discipline. The bot no longer volunteers help with unrelated topics it notices in conversation history or cross-channel activity.
Cron embed hygiene. Stale pinned prompt and status messages are now unpinned before replacements are posted, preventing pin clutter in cron threads.
Prompt efficiency patch. Reduces per-turn token overhead by ~1,765 tokens through instruction compression and runtime deduplication.
Smarter context, better reliability. Messages now carry temporal awareness, cron scheduling is decoupled from Discord thread state, and the dashboard gets a settings panel.
The 1.0 milestone. Six runtimes validated end-to-end, Discord canvas activities, a redesigned operator dashboard, and a fully automated release pipeline.
Self-improvement harness, PDF ingestion, and model defaults modernized.
Security hardening and safety rails: webhook sanitization, config action auth, git commit safeguards, and smarter follow-up prompts.
Image generation gets source-image support and a progress indicator.
Vision lands: the bot can now see images in chat history, not just the triggering message.
Forge reliability overhaul, Codex hardening across the board, and a new operator dashboard preset switcher.
Operator ergonomics jump forward: built-in dashboard, MCP visibility, and much sturdier long-running workflows.
Cold storage memory lands, model routing gets a tiered overhaul, and prompt/streaming behavior gets tighter.
Memory injection got smarter, Discord reads got richer, and a few operational rough edges got sanded down.
Maintenance cut to roll forward the voice/runtime work from the surrounding releases.
Voice got substantially more capable, with runtime routing and better conversation state.
Deferred workflows can now nest instead of stopping after one hop.
Install fix release for newer Fedora / GCC toolchains.
Automations learned how to keep state and hand work to each other.
Durable memory stopped being static and started adapting over time.
Prompt context, voice, and task/thread synchronization all got sharper.
Parallel sub-agents became a first-class feature.
A broad systems release: secrets, memory upgrades, transcript ingestion, and stronger safety around external content.
OpenAI-compatible runtimes graduated from basic text to real tool use.
The npm-managed install path got more production-ready.
Follow-up fix for npm-managed installs.
The updater era started here.
Bootstrap and first-run setup got much less painful.
First public npm release.
Full setup guide in the README.
Contributing or want to run from source? Clone and set up manually.
# Clone and install
$ git clone https://github.com/DiscoClaw/discoclaw.git
$ cd discoclaw && pnpm install
# Interactive setup
$ pnpm setup
# Or set env vars manually
$ cp .env.example .env
DISCORD_TOKEN=your-bot-token
DISCORD_ALLOW_USER_IDS=your-discord-id
# Launch in dev mode
$ pnpm dev