Connection health

Snapshots

A point-in-time copy of config + the database — take one before a risky change, restore it if the change turns out wrong. Creating one never stops the app; restoring briefly reopens the database connection, not a full restart.

Hot Reload

Applies an edit to BotServer's own bot/*.py to this already-running process — most files take effect on the very next call; Discord/Slack/Matrix code gets a brief reconnect; a documented set of core files (routing, the database connection, the dashboard itself) still need a full restart, reported here as "restart required" rather than silently skipped.

TURN Server (WebRTC mesh relay)

Used by the Android app's WebRTC fallback for mesh APK transfers, when two devices can't reach each other directly (symmetric NAT, restrictive firewalls). BotServer never runs the relay itself — this only mints short-lived credentials for a real TURN server (coturn) you run separately with the same shared secret. See docs/turn-server-setup.md.

Edit .env contents

Editing writes straight to the file above — this is the actual bot token, API key, etc. in plain text. A timestamped backup is made automatically before every save and every restore, so nothing is ever overwritten without one. Takes effect after the server process restarts.

Backups

Every save and restore snapshots the previous file — kept indefinitely, nothing here is auto-deleted.

MCP servers

Servers Claude Desktop itself has configured — enable/disable each below.

Control this app over MCP

Registers this app's own control server (bot/mcp_server.py) with Claude Desktop as an MCP server named bot-server, so Claude Desktop (or Claude Code via claude mcp add) can check status, list jobs, flip backends, and start/stop/restart Claude Desktop itself — the same things this dashboard can do, but callable as MCP tools instead of clicks.

Config change history

Add a bot

Plugins

Local Python files that register extra agent tools and/or slash commands — trusted local code with the same privileges the app already has (not sandboxed, not fetched from a network). Enter the path to a plugin.py already on this machine's disk.

Pending pairings

Someone messaged a bot from an unrecognized chat ID and got a one-time code — approve here to add them to that bot's allowed users.

Configured bots

Schedules

Recurring prompts a bot runs on its own — the same store /cron, /loop, and /heartbeat already write to in chat, now visible and manageable here too.

Backups

Every create/update/delete snapshots the full bot list first — kept indefinitely, nothing here is auto-deleted.

Providers

Named endpoints for the "custom_model"/"native_agent" backends — any OpenAI-compatible server (a local Ollama/LM Studio/vLLM/llama.cpp instance, OpenRouter, or real OpenAI). A bot instance points at one with a model of the form "<name>/<model_id>".

Models

Add a swarm

Configured swarms

Hermes swarm tools — which Hermes agents can organize a swarm themselves

A Hermes-backed instance can always spawn its own sub-agents via its native delegate_task tool. Enabling swarm tools additionally registers this app's own MCP server into that instance's Hermes config, so its agent can reach OTHER bot instances too — ask_instance, run_swarm, dispatch_swarm_goal, and more, mid-turn. Takes effect on that instance's next message, not immediately.

Project context — shared markdown docs every agent can read/write

Small, named documents any registered instance (any backend) can read via read_project_context or write via write_project_context — the one place a swarm of workers and their manager keep something like a project-status update in sync, independent of any one instance's own conversation history. Capped at 8KB per doc, 50 docs total.

Run a swarm

Run history

Swarm budget — pre-dispatch cost ceiling for dispatch_swarm_goal

A worst-case ESTIMATE gate (max children × assumed tokens/child × the resolved model's known price) checked before a dispatch is ever sent — not a real spend tracker. A refused dispatch is logged below as swarm_dispatch_blocked and never touches Hermes.

Delegation activity — who asked whom to do what

Every cross-instance call — ask_instance, dispatch_swarm_goal, and an api-backend agent's own delegate_to_instance — logged here, newest first. A dispatch row can be expanded for its live tool-call status and post-hoc per-child breakdown.

Model health — hybrid classifier self-monitoring

Every message the Support Bot classifies runs through two independent models — the deterministic TF-IDF centroid model and a trained neural network (MLP) — and the hybrid trusts whichever is right more often. This is that behavior, measured live over real traffic, not a static claim.

Support Bot phrases

Add example phrasings for an intent the Support Bot misrecognized or missed entirely — retrains both the TF-IDF model and the neural network immediately, no restart. These are layered on top of the built-in baseline (training_data.py), never replacing it.

Generate a key

Leave all three blank to auto-fill this machine's own detected LAN/Tailscale/Funnel addresses. Each is an independent path to this server — the app tries them in order and switches automatically if one stops answering, so the phone stays connected from any network without you doing anything.

Shown once — this key can't be recovered after you leave this page. Copy it or scan the QR now.

Key

"Send APK" queues the last APK built on the desktop app for that device — it downloads and prompts to install next time the Android app is opened (there's no way to push to a closed/backgrounded phone without Firebase configured).

1. Generate a pairing token (for someone linking into THIS server)

One click — this server auto-detects its own reachable address and bakes it into the token, so whoever links in never needs to know or type it. Share the token — never your dashboard token — with them; it works once and expires in 10 minutes, so it's safe to send over chat.

2. Link a server (using ITS pairing token)

Just a name and a token — no address to look up. On the other server's own Linked Servers tab, have its admin generate a pairing token (step 1 above, done on their side) and paste it below. Your own dashboard token never leaves this machine — the pairing token is the only thing sent over the network, and it stops working the instant it's used.

Linked servers