📁 Loading... 🧠 Agent 📋 Plan
Agent (Goal off) — turn by turn. Each message is its own request; when the answer is done, the agent stops.
Agent + Goal on — the chat carries an objective. omp's hidden
It does not run by itself. Every turn still starts when you send something — omp only auto-continues goals in its own terminal UI, not here.
Plan is the third mode and cannot be on at the same time: it writes a plan without executing it.
Agent + Goal on — the chat carries an objective. omp's hidden
goal tool is switched on for this chat, so the objective, its status (active / paused / complete) and its spend are remembered between messages and steer them, instead of every message starting fresh.It does not run by itself. Every turn still starts when you send something — omp only auto-continues goals in its own terminal UI, not here.
Plan is the third mode and cannot be on at the same time: it writes a plan without executing it.
Limits for this chat only — this chat
Every chat tab has its own goal, its own counters and its own limits. Stop this goal after:
This goal: —
👥
⚙️ One quick install and you're ready
Agent Pi needs its AI engine - pick one below and it installs itself automatically (no terminal, nothing to configure). Not sure? Take the recommended one. You can add or switch engines later in Settings.
oh-my-pi (OMP) recommended
Native binary · bundled with the app · installs offline · no Node.js required
Pi CLI
Node.js · sessions · tools · extensions · requires Node.js + npm
Little Coder
Pi + curated skills, tuned for small local models · requires Node.js ≥ 22.19 + npm
📬 Optional: get setup tips & release news by email
No spam - just releases and tips. Unsubscribe anytime.
⚙️ Preferences
Agent Pi server v…
An agent is a saved chat: its own system prompt, model, tools and @handle. Open one in its own tab, or type its @handle in any chat to call it.
No custom agents yet
Prebuilt roles. Adding one gives you an editable copy.
Configure default orchestration when multi-agent @mentions or channel conversations run.
This chat's next turn runs on …
Resolving…
How engines are chosen
Little Coder — local models (LM Studio, Ollama, llama.cpp, vLLM).
Pi CLI — multi-agent chats, where agents talk to each other.
oh-my-pi — everything else.
Runtime - v?
🧠 Default Models & Planning
🧭 Reasoning & Quality ?
1 Start of every turn
?
Checking…
Active Providers: ?
2 Every file edit
3 End of every turn
💾 Context & Token Optimization ?
1 As each tool result comes back
Checking...
2 As each request is sent
Checking...
Engine chats (omp) never go through this proxy; they rely on Compaction (step 3).
3 When the chat fills up
Compact at: 75%
🧠 Memory & Auto-Actions ?
Automated Triggers (after each assistant turn):
🔔 Notifications & Sounds Checking…
50%
Per-Event Notification & Sound Configuration:
✅ Response Complete
Agent finished responding
❓ Decision Needed
Interactive question or escalation
⚠️ Errors & Failures
Tool error or process failure
⏳ Still Working / Stalls
Silent step running > 45s
🐢 Tool taking too long
One command/tool still running after
min
⏱️ Long-running turn
Agent has been working on one turn for
min
🤝 Long AI-to-AI conversation
Agent-to-agent channel run lasting over
min
Desktop notification permission is not granted yet in this browser.
🐝 Multi-Agent & Swarm ?
🧰 Tool Calls ?
🛡️ Safety & Command Approval ?
⚡ Agent Pi API
One key for everything — no provider signups, no keys to manage.Not available yet
Agent Pi API is still being built. When it is ready you will be able to top up a balance here and use the latest models without signing up to any provider.
Until then, use Subscriptions for plan logins you already have, API Keys for your own provider keys, or Local Models to run offline.
Until then, use Subscriptions for plan logins you already have, API Keys for your own provider keys, or Local Models to run offline.
🔐 Subscription Logins
Sign in to your coding subscriptions via browser OAuth. Tokens are stored in omp's credential vault and used automatically across all engines.🔷 Claude (Pro / Max)
Your Anthropic Claude Pro or Max subscription — no API key. oh-my-pi and Pi CLI both use it directly.
Checking…
🚀 Speed increase: Up to ~2.5× faster token generation output via Anthropic's priority tier.
⚠️ Cost note: Premium token pricing applies (~1.5×–2× rate on API / rapid quota burn on high-volume runs). Off by default.
🤖 Codex / OpenAI
Use your ChatGPT Plus / Pro / Business plan through Codex — no API key. Sign in once and omp draws on your plan quota.
Checking…
On a headless box use “ChatGPT Plus/Pro (Codex, headless/device)” in the picker below — it swaps the browser callback for a device code.🪐 Antigravity / Gemini
Google Antigravity plan (Plus / Pro / Ultra) via Google OAuth. Exposes Gemini 3 plus the Claude and GPT-OSS models Antigravity carries.
Checking…
Signing in refreshes the model catalog, so the plan's models appear in the picker without a reload (v2.4.300).🔐 Other Supported Subscriptions (66 Providers)
Sign in to any supported coding subscription (Copilot, Grok, Z.AI, Mistral, Moonshot, Qwen, Minimax, etc.) via browser OAuth.
The link is auto-copied to your clipboard. If it didn't open, paste it into your browser.
🔑 Bring Your Own Keys (BYOK)
Enter your personal or organization API keys for direct, unmetered access to providers. Keys are stored encrypted on this machine only.🔐 Key storage security
How your API keys are encrypted on this machine. Keys never leave your computer.
Checking…
🖥️ Local Model Runtimes
Connect local AI servers running on your machine via their standard OpenAI-compatible/v1 endpoint. No API key or account needed.Anything on this machine that serves the OpenAI
/v1 API — LM Studio (port 1234), Ollama (port 11434), llama.cpp / llama-server (Unsloth GGUFs, port 8000), vLLM, LocalAI. Give it a port; there is no account and nothing to sign into.Checking…
🧪 Free WebApp Proxy Gateway · AI WebApp Harness
Unified Web2API proxy utilizing browser session tokens & cookies for free web tiers (DeepSeek, ChatGPT, Claude, Gemini, Grok, Qwen/Kimi/GLM).○ Stopped
⚠️
High-Risk Reverse-Engineered Automation — Terms of Service Notice
This feature connects to unofficial reverse-engineered web endpoints using session cookies and browser tokens.
• Account Ban Risk: Providers actively enforce Terms of Service and may apply temporary rate limits or permanent account bans.
• Best Practice: ALWAYS use dedicated throwaway / burner accounts. Never use primary personal, work, or billing accounts.
• Breaking Changes: Upstream web frontends change without notice. Use the "Sync Upstream Repositories" button below to pull fixes when a provider breaks.
• Account Ban Risk: Providers actively enforce Terms of Service and may apply temporary rate limits or permanent account bans.
• Best Practice: ALWAYS use dedicated throwaway / burner accounts. Never use primary personal, work, or billing accounts.
• Breaking Changes: Upstream web frontends change without notice. Use the "Sync Upstream Repositories" button below to pull fixes when a provider breaks.
Local Gateway Service http://127.0.0.1:5050/v1
Checking gateway status…
👤 Connected Session Accounts
Add New Account Credentials
Loading accounts…
🔄 Upstream Web2API Repositories (14 Clones)
Tracks 14 upstream reverse-engineered repos (Chat2API, DeepSeeker, grok2api, etc.)
🖥️ Tools & shells the AI can use
These are the apps and shells Agent Pi found on your computer - the AI uses them to run things for you. Just installed something? Click Rescan. Missing something you want? Click Install.
⚡ Performance
Changes require app restart to take full effect.
Allocated to Node.js V8 heap. Higher = faster but uses more RAM. Default: 2048 MB. Set to ~50% of system RAM.
UV thread pool size for async I/O. Default: 4. Match your CPU core count for best performance.
Text-based app doesn't need GPU. Saves ~100MB RAM. Recommended ON.
Removes CSS animations and smooth scrolling. Reduces CPU usage.
💾 Conversation cache
Cached conversations open instantly. Old entries are evicted automatically once the cache passes the size cap.
When the cache exceeds this, the least-recently-used conversations are dropped. Default: 1 GB.
…
🚀 Server auto-start
When enabled, the Agent Pi server will start automatically when you log into your computer. The server registers the login entry itself - no browser extension required.
🧠 Agent memory & folder search
Gives every saved agent a private long-term memory and its own searchable index of the folder it works in. Nothing leaves your computer - it runs in three local Docker containers. Agents work fine without it; this only adds recall.
🧠● checking…
Memory stack
Checking…
📊 Usage & Efficiency
Measured across your recent conversations - tokens, prompt-cache hit rate, how often context is compacted, what tool output costs, and flags for anything that looks wrong.
Open this panel to measure.
📡 Integrations
Connect external automation services. Agent Pi holds the credentials server-side (encrypted at rest) so they never appear in prompts.
🐙
GitHub ● checking…
Sign in with GitHub for clone/push, PRs and issues. Managed in the Git panel.Checking…
Loading…
No workflows yet
Tools
Everything this agent can call, wherever it comes from — the engine's own tools, MCP servers you configured, extensions you installed. The badge on each row says which. Switching something off here means this agent may not use it; it stays installed.
⚛ Extensions & MCPs
Extensions and MCP servers you can add to the agent. The badge on each row says how it is delivered; the trigger says when it runs. Skills have their own panel (🧩 Skills), and native engine tools are in 🔨 Tools.
Installed and switched on. The trigger on each row decides when it actually runs.
Click Refresh to search GitHub for community extensions & MCP servers.
Loading the Pi package catalog…
🔗 MCP Servers
MCP (Model Context Protocol) servers give the agent access to external tools - databases, browsers, APIs, and more. Enable a server here to make its tools available in all chats. MCPs that require configuration must be set up and validated before they work.
Configure
📝 Custom Prompt
Role & custom instructions are injected at the top of every message the agent receives. (Skills now have their own panel - the 🧩 Skills button.)
📈 Capability usage
Nativeomp's own tools · Skillinstructions the agent opened · Extensionomp plugin tools · BridgedPi CLI plugins repacked for omp · MCPtools from MCP servers
Loading…
🧩 Skills
Install a skill to make it available to the agent. Use the ⚡ Skills menu in the header to activate skills for a specific chat.
Installed & available
Click Refresh to search GitHub for community skills & prompts.
Loading the Pi package catalog…
⚙ Node
⛓ Run
🤖 Save this chat as an agent
A model wrote this from your transcript - read it before saving. It can miss or misstate a decision, and you are editing the agent's system prompt.
🤖 Edit Agent
0. Identity & @Handle — how to call this agent in any chat
Tagged agents show a 🔑 badge. Set the key in Settings → MCP Servers, on the server that provides the agent's tools — never in a chat message.
1. System Prompt — the agent's core identity, directives, and rules
2. Folder(s) & Workspace — where this agent operates
3. Triggers & Autonomy — when and how far the agent runs
4. Credentials, Tools & MCPs — tools and external systems this agent can reach
Loading…
5. Skills Suite — active capabilities enabled for this agent
Click skill pills to toggle them on/off for this agent:
👥 New Agent Team
Describe the team in plain words and the AI sets it up. Every setting stays yours to change.
Start from a sample
Every line is optional. Leave something out and the AI picks a sensible default; name a model, a schedule or a limit and it is set exactly. Ctrl+Enter builds.
👥 HOW MANY AGENTS? 2 agents
💬 SESSION TYPE
🎯 ORCHESTRATION & TOPOLOGY
How the Team Runsmode, schedule, limits, lead, stop rule
👤 Prompts only (0 prompts in chat)
Jump to Date in Chat
inactive
Click the ＋ under a node to add a step.
Drag nodes to arrange · drag the canvas to pan · scroll to zoom.
Drag nodes to arrange · drag the canvas to pan · scroll to zoom.
🛠 Builder - tell it what to change
🤖 Agents
Saved Agents
Loading…
Agent Templates read-only · adding one gives you an editable copy
Loading…
Loading conversation…
🔀
🩺 Connection doctor
Running checks…
Sign in to Agent Pi
or with email
🧑🤝🧑 Run in parallel
Two ways to work in parallel: let the assistant split big jobs on its own, or list specific tasks to run at the same time. Up to 8 at once.
🤖 Let the assistant use helpers on its own
For big jobs, it splits the work across background helpers automatically.
🐝 Multi-agent mode - split big requests
When one message holds several independent jobs, run one agent per job (up to 4).
Or run specific tasks at the same time:
⚠️ Helpers can run commands on their own in this folder.
🧑🤝🧑 Helpers - 0 of 0 done
🧑🤝🧑 Looks like 2 separate tasks - run them in parallel?
🔍 Search
📅 Date:
–
tags:
Ctrl+F = current conversation | Ctrl+Shift+F = all conversations
Tags
Enter adds | Esc closes | click a chip to remove
>
Esc close | ↓↑ navigate | Enter open | Type to filter
@
Esc close | ↓↑ navigate | Enter switch | Type to filter open chats + history
Select working folder
Sort:
/
⚡ Workspace
Type or paste a folder path, or pick from recent.
Clone a GitHub repo as the workspace for this chat. The clone lives at
~/.pi/repos/ and is reused across sessions.⚠ GitHub isn't connected, so your repositories can't be listed. Connect GitHub →
You can still type
You can still type
owner/repo above to clone a public repo.