Prometheus
"I whom you see am Prometheus, who gave fire to mankind."
c. 440–430 BCE, from Prometheus Bound.
Queued prompts
Proposals
Agent-generated proposals awaiting your approval
Loading proposals...
Audit Log
Non-main agent runs, grouped by run. Click a row to expand full tool activity.
| Time | Agent Run | Status | Tools | Type | Top Activity |
|---|---|---|---|---|---|
| Loading audit log... | |||||
Memory Graph
Audit-indexed knowledge web across sessions, tasks, schedules, teams, projects, and durable memory.
Loading graph...
Loading memory graph...
Drop image to reshape node outline
Hub
Skill usage, goals, and tool activity
Top Skills
Goals
Tool Usage
Less
More
Skill
Background Tasks
0 tasks
Task Details
Chat with agent
Schedules
0 schedules
one-off & recurring tasks — team agents live in Teams
No schedules yet. + New Schedule to get started.
Create Schedule
Select a configured agent to run this scheduled task as a specialized subagent.
Agent instructions preview
Loading MCP servers…
Restrict this schedule to only use the checked MCP servers.
URLs injected as [REFERENCE LINKS] in the prompt. The AI will pull from these sources.
Browser idle
Workspace Files
Browsing the full workspace
Import a folder or browse the workspace to populate the project tree.
No files open
Import files or a whole folder, then open anything from the project tree on the left
Workspace Files
Creative Surface
Shared scene graph workspace
Visual composition workspace
Image Frame
1280 x 720
Build directly on the frame.
Live canvas
100%
Timeline
Arrange scenes, scrub the playhead, and shape motion timing.
Live Browser
Waiting for browser activity in this chat
No session
Stream idle
No active browser session
Browser canvas is standing by
When Prometheus opens or updates a browser session, the browser canvas will follow it here and keep the current session, page title, and interaction mode in sync.
Latest browser frame
Viewport frame
Prometheus
Last Tool Call
No browser tools used yet
Open a site with the browser tools to wake this surface up.
Browser idle
Waiting for activity...
Task
Mode Notes
Browser interaction guidance will appear here.
Click the browser to take control.
Use a name you would naturally say out loud later.
Capture the current page as the starting point, then stage each browser action one step at a time.
Verification Approval
Choose how far Prometheus should replay the recorded workflow before it reports what happened.
Pending teach step
Stage a browser action to inspect it before it runs.
Selected element
No browser element selected
Enter Teach mode and click the live browser frame to capture a real page element.
Fluid viewport
Context
Connections
Agent Context
Progress - This panel will be used to make a short to-do list
Pending Actions
No pending actions
Waiting for activity...
New Project
● Connected
What Prom can access
Activity Log
No activity yet — Prom will log every read and write here.
⚠️ Authentication Required
The task needs your help to continue. Please provide the required information.
Create Skill
Save a reusable playbook for Prometheus.
Settings
Runtime
Model-
Gateway-
Provider-
Notes
These settings apply instantly to this gateway process.
For small models, keep tool routing strict and memory scoped.
Main Agent Heartbeat
HEARTBEAT.md path: -
HEARTBEAT.md Instructions
Persisted to
HEARTBEAT.md in the main workspace. Run on each heartbeat tick.
Subagent Schedules
Each agent can keep its own HEARTBEAT.md and timer. Scheduled Tasks remain separate for explicit cron jobs.
Loading agent heartbeats...
Web Search
API keys are managed in the Credentials tab and stored encrypted.
Search Provider Keys
Stored encrypted in vault — never written to config.json in plain text.
Vault Status
Loading…
How Credentials Work
When you hit Save Settings, any key you've entered is encrypted with AES-256-GCM and stored in
Config files only ever contain a
If a field shows ••••••••, a key is already stored. Leave it as-is to keep the existing key, or type a new one to replace it.
vault.enc.Config files only ever contain a
vault:… reference — never the actual key.If a field shows ••••••••, a key is already stored. Leave it as-is to keep the existing key, or type a new one to replace it.
Recent Vault Access
Loading…
File Access
LLM Provider
Ollama Connection
Primary Model
llama.cpp Server
Start with:
llama-server -m model.gguf --port 8080LM Studio
Enable the Local Server in LM Studio before connecting.
OpenAI API Key
OpenAI — GPT Plus / Codex
Connect your ChatGPT Plus or Pro account via OAuth. Your browser will open to authorize Prometheus.
Connected
Anthropic Claude
Connect your Claude Pro or Max subscription.
Run
Run
claude setup-token in your terminal, then paste the token below.
Connected
Perplexity AI
Get an API key at
perplexity.ai/settings/api. Sonar models are web-grounded.Google Gemini
Get an API key at
aistudio.google.com/apikey. Uses the OpenAI-compatible Gemini endpoint.Agent Model Defaults
Preset which model each agent type uses. Agents without an override fall back to the primary model.
Templates
Main Chat Agent
Proposal Executors
High Risk
Low Risk
Teams & Agents
Coordinator
Manager
Background Tasks
Subagent Role Defaults
Per-role model defaults. Leave blank to inherit from the primary model.
Planner
Orchestrator
Researcher
Analyst
Builder
Operator
Verifier
Switch Model Tiers
Models used when the AI calls switch_model() to downshift for lighter tasks. Leave blank to disable that tier.
Fast (low)
Careful (medium)
Background Agents
Model used for tasks spawned via
background_spawn. Leave blank to inherit from Background Task.Background Spawn Agent
Brain System
Models used for autonomous Thought and Dream runs. Leave blank to use the primary model.
Brain Thought
Brain Dream
Rolling Context Compaction
Uses an isolated compactor prompt (no Prom persona/tools) to summarize thread history, then resets active chat context to the summary.
Channel Connection
Not configured
Not configured
Not configured
Instructions
Telegram Setup
1. Create a bot via @BotFather.
2. Copy the bot token into this panel.
3. Get your user ID(s) from @userinfobot.
4. Add allowed IDs, then click Test and Save.
5. Use Send Test to verify delivery.
1. Create a bot via @BotFather.
2. Copy the bot token into this panel.
3. Get your user ID(s) from @userinfobot.
4. Add allowed IDs, then click Test and Save.
5. Use Send Test to verify delivery.
Discord Setup
1. Create an app + bot in Discord Developer Portal.
2. Enable required intents (including Message Content if needed).
3. Invite bot to your server with message permissions.
4. Paste bot token; optionally add app/guild/channel IDs.
5. Click Test, then Save, then Send Test.
1. Create an app + bot in Discord Developer Portal.
2. Enable required intents (including Message Content if needed).
3. Invite bot to your server with message permissions.
4. Paste bot token; optionally add app/guild/channel IDs.
5. Click Test, then Save, then Send Test.
WhatsApp Cloud API Setup
1. In Meta Developer dashboard, open WhatsApp Cloud API.
2. Copy access token and phone number ID into this panel.
3. Configure webhook verify token/secret (if using inbound webhooks).
4. Add a test recipient number in E.164 format.
5. Click Test, Save, then Send Test.
1. In Meta Developer dashboard, open WhatsApp Cloud API.
2. Copy access token and phone number ID into this panel.
3. Configure webhook verify token/secret (if using inbound webhooks).
4. Add a test recipient number in E.164 format.
5. Click Test, Save, then Send Test.
Configured Agents
Loading agents...
Agent Details
TEAM PERMISSIONS
AGENTS.md (CodeMirror)
Team Subagent
This agent is managed by a team. Its task instructions come from system_prompt.md (fallback: HEARTBEAT.md) — this is what the manager writes to tell it what to do each run.
Heartbeat
HEARTBEAT.md — what this agent does when woken. Saved automatically when you click outside the editor or use Save.
Manual Spawn + Run History
Webhook Endpoint
Not configured
Setup Guide
1. Enable + set a token — any secret string you choose.
2. Save and restart the gateway to activate.
3. Access from the internet? Use Tailscale (recommended), ngrok, or Cloudflare Tunnel.
Endpoints:
Auth: Send token as
or
Works with: GitHub Actions, Stripe, Vercel, n8n, IFTTT, Home Assistant, or any script.
2. Save and restart the gateway to activate.
3. Access from the internet? Use Tailscale (recommended), ngrok, or Cloudflare Tunnel.
Endpoints:
POST /hooks/agent — full AI runPOST /hooks/wake — lightweight nudgeGET /hooks/status — health checkAuth: Send token as
Authorization: Bearer <token>or
x-prometheus-token: <token>Works with: GitHub Actions, Stripe, Vercel, n8n, IFTTT, Home Assistant, or any script.
MCP Servers
Loading...
Add MCP Server
What is MCP?
Model Context Protocol (by Anthropic) is a standard that lets AI connect to external tools and data sources.
Each MCP server exposes a set of tools — Prometheus discovers them automatically when you connect. The AI can then call them during tasks, just like its built-in tools.
Servers run locally as child processes (stdio) or connect to remote HTTP endpoints (SSE). No data leaves your machine unless the tool itself calls an external API.
Each MCP server exposes a set of tools — Prometheus discovers them automatically when you connect. The AI can then call them during tasks, just like its built-in tools.
Servers run locally as child processes (stdio) or connect to remote HTTP endpoints (SSE). No data leaves your machine unless the tool itself calls an external API.
Popular Servers
Add Shortcut
How it works:
When the AI opens a URL matching this hostname, these shortcuts are injected into its context so it can use them instead of hunting for DOM elements. Shortcuts for navigation (like j/k on X) work with the keyboard focus tracker so the AI knows exactly which item it’s on.
When the AI opens a URL matching this hostname, these shortcuts are injected into its context so it can use them instead of hunting for DOM elements. Shortcuts for navigation (like j/k on X) work with the keyboard focus tracker so the AI knows exactly which item it’s on.
Known Shortcuts
Remote access
Off
Off by default — local Wi-Fi pairing still works exactly as before. Turn this on to make the QR code encode a public HTTPS URL (e.g. a Tailscale Funnel address) so phones can pair from cellular or any other network. Desktop approval is still required for every new device.
Mode
Public URL
Paste the HTTPS address Tailscale assigned your machine.
Tailscale status
Click Detect Tailscale to check installation, login, and Funnel state.
One-time setup:
1. Install Tailscale from tailscale.com and sign in.
2. In a terminal, run:
1. Install Tailscale from tailscale.com and sign in.
2. In a terminal, run:
tailscale funnel 18789
3. Click Detect Tailscale, accept the suggested URL, save, then generate a new QR.
Pair a phone
Loading QR…
Scan this QR with your phone's camera. After it opens, you'll see an approval prompt below.
QR codes expire in 5 minutes. Generate a new one anytime.
QR codes expire in 5 minutes. Generate a new one anytime.
Awaiting approval
0
No incoming pairing requests.
Trusted devices
0
No devices paired yet.
Disable instantly revokes a device's access without forgetting it — re-enable anytime.
Remove deletes it; the user will have to re-scan a QR.
Teams
0 teams
Subagents
0 agents
Loading subagents...
Create Team