English · 简体中文
Claude Code hit its usage limit mid-task. Codex picked it up in the same conversation. Nothing was re-explained.
The project agent
Is your context actually yours?
The plan it took three rounds with Claude Code to settle, the file Codex left half-edited, the trade-off you decided in Cursor: each one is locked inside its own session.
Close the session, hit the usage limit, switch tools, and that context is out of reach. Next task, you explain it all again.
You are the agent's intern: copying and pasting, ferrying context, remembering where all the odd files live, while the agent gets the thinking.
Orbital takes the context out of the session and puts it back in your local folder. Any agent can pick it up at any time, with all of it intact. Claude Code, Codex and Cursor all work.
The context is yours. The intelligence is interchangeable.
Your context, usable by any agent
Set up in under 5 minutes. No Python or Node required. Bring your own API key.
One-minute demo: one task, two agents, nothing re-explained
demo-en-web.mp4
You're three rounds into a design discussion with Claude Code when it stops with "You're out of usage credits." Your Codex quota is sitting right there, but Codex knows nothing about this project: the goal, the two decisions you just made, the half-edited files. You'd have to explain all of it again.
People already use several capable agents at work — for the newest model, the leftover quota, or because a particular tool is better at the job.
But each agent works inside its own session, with its own context and history. When you move between sessions or tools, you become responsible for carrying the project between them: restating goals, explaining previous decisions, locating artifacts, and checking what was left unfinished. You end up working as their intern, ferrying context between them to keep your own project moving.
Orbital changes the unit of work from the session to the project.
A project agent stays responsible for the project across tasks, sessions, and worker agents. It maintains the shared context, decides what needs to happen next, delegates when useful, and records every outcome back into the project.
Individual agents complete tasks. Orbital keeps the project moving.
Orbital maintains five things that normally disappear or fragment between agent sessions — all of them as plain files in your project folder:
- State — what is true about the project now (PROJECT_STATE.md)
- Decisions — what was decided and why (DECISIONS.md)
- Lessons — what the project has learned (LESSONS.md)
- Work — what is running, completed, or blocked (queue.json). Every queued task ends Completed or Blocked — an agent that stops without a verdict gets re-prompted, then force-blocked with a reason. Nothing drifts away silently.
- Artifacts — what the agents researched, wrote, or built (the workspace itself, plus orbital/output/)
These stay in the local project and become context for future work. On a cold start, Orbital assembles them into its own system prompt before it acts.
When Orbital delegates, the worker is pointed at those same files and told they are authoritative — that briefing is rendered fresh on every dispatch, not something you paste in. Each worker also keeps its own memory file inside the project, so it accumulates its own experience across dispatches. When the task finishes, Orbital reads the result and records what matters back into the project.
The worker can change. The project continues.
-
Launch Orbital — the setup wizard guides you through two steps: Step 1 — LLM Provider: In mainland China, tap Sign in with TokenDance, authorize, and start on free tokens. Otherwise pick a provider from the preset cards, follow the key-console link to grab an API key, and paste it in. Supports DeepSeek, Anthropic, OpenAI, Moonshot, and a dozen other providers. Step 2 — Connect Your Accounts: Link API connectors (Google Calendar, Drive) and sign in to sites your agents will need (Google, GitHub, etc.) so they can browse without getting blocked by CAPTCHAs. Everything here is optional and can be done later in Settings.
-
Create a project — give it a name, pick a workspace directory, set an autonomy level
- Chat — type a task in the chat bar and the project agent handles it
- Walk away — queue the next tasks; each finished one becomes context the next builds on
The project agent keeps the project's state, decisions, and lessons current across sessions.
It delegates to Claude Code, Codex, or Gemini CLI against the same project context, then records the result.
Memory, scheduling, and sub-agents are table stakes now — every tool below has them. These are the three questions where the answers still differ.
The short version: the difference isn't any one capability — it's that the project, not the session, is the unit that owns state, workers, and governance.
Orbital Is / Is Not
Project & Workspace Model
Each project maps to a workspace directory and maintains its own sessions, queue, triggers, and configuration.
Browse, preview, and upload files in each project's workspace — and watch the agent's output accumulate
{workspace}/
+-- AGENTS.md # Onboarding signpost for external agents (seeded at creation, user-owned)
+-- orbital/ # Operational metadata
+-- sessions/
| +-- {session_id}.jsonl # Append-only session log
+-- instructions/
| +-- project_goals.md
| +-- user_directives.md
+-- skills/ # Project skills
+-- sub_agents/ # Sub-agent transcripts + per-worker MEMORY.md
+-- tool-results/ # Tool output artifacts
+-- output/ # Agent work artifacts
| +-- screenshots/ # Browser screenshots
| +-- pdfs/ # Saved PDFs
| +-- shell-output/ # Shell command output
+-- queue.json # Task queue (queued / running / completed / blocked)
+-- PROJECT_STATE.md # Current-state scratchpad (overwrite)
+-- DECISIONS.md # Durable decisions + reasoning
+-- LESSONS.md # Durable heuristics / playbooks
+-- INDEX.md # Navigation map: file tree + one line per file
+-- DECISIONS_ARCHIVE.md # Demoted decisions (read-on-demand)
+-- LESSONS_ARCHIVE.md # Demoted lessons (read-on-demand)
~/orbital/ # Home global (daemon infrastructure)
+-- daemon.pid # Singleton enforcement
+-- device.json # Device identity
+-- browser-profile/ # Shared browser profile
+-- credential-meta.json # Credential metadata
Session format: One JSON line per message (role, source, content, timestamp, tool_calls). Append-only with file locks. Never modified except during compaction.
Context Management & Compaction
This is how the project agent keeps context available across sessions. The agent-maintained Layer-1 files are injected every turn (bounded per file) and consolidated at session boundaries:
Each entry carries system-managed metadata (id / created / touched / tag) so dedup runs on recency. Per-turn injection bounds each file to a budget derived from the active model's context window. Session-end runs a deterministic size backstop (demote/trim, never an LLM call) plus a best-effort LLM dedup/merge that fixes contradictions. (SESSION_LOG.md was retired; the Layer-1 files are injected every turn, so a separate session history is redundant.)
Cold resume: On session start, these files are assembled into the system prompt so the agent can reorient before it acts.
DECISIONS.md and LESSONS.md — written by the agent as it works, and carried into every future session
Compaction (when context usage exceeds 80%): memory flush, LLM-driven summarization of older messages, recent messages kept intact, post-compaction reorientation with project goals and current state.
Prefix caching (v0.4.2): the system prompt is split into static, semi-stable, and truly-dynamic sections so up to ~95% of input tokens hit the provider's prefix cache on follow-up turns. See the v0.4.2 release notes for benchmark numbers.
Sub-Agent Delegation
Orbital is not tied to a single AI tool. The project agent plans and delegates, while specialized workers execute — each reading the same accumulated project context. Any CLI-based agent can be registered via a manifest file; Claude Code, Codex, Cursor, Gemini CLI, Aider, Cline, Goose, Copilot CLI, and Continue ship with one.
Each dispatch renders a fresh inheritance prompt that points the worker at PROJECT_STATE.md, DECISIONS.md, LESSONS.md, INDEX.md, the project's instructions, and its skills — declaring them authoritative and off-limits for writes. Workers read them on demand rather than receiving a pasted copy, so the brief never goes stale.
Each sub-agent keeps its own long-term memory across dispatches — curate what it remembers...
...then delegates a task to @claudecode, reviews the result, and writes it back into the project
Transport types:
Note: Codex uses its native app-server path, not PTY or ACP. Orbital launches
codex app-serverand speaks JSON-RPC directly. ACP is available for any ACP-compliant worker; PTY is the default for other interactive CLI agents.
Task Queue
Each project has a queue, stored with the project at orbital/queue.json. Add tasks — pin urgent ones to the front — and your agent works through them one at a time, in order, without you watching.
The agent must declare an outcome on every item; it can't silently drift to the next one:
Pause to steer. Pause the queue mid-item to chat freely — your clarifications land in the same session, so the agent sees them when you resume.
Continuity by design. Each completed item's artifacts are already in the project when the next item starts, so the agent can use them when supervising later tasks. The project's triggers (schedules and file watchers) are listed in the queue's Automations section alongside your tasks.
Queue tasks and walk away — each finished one becomes context the next builds on
Workbench — what only you can decide (beta)
Some things an agent genuinely cannot finish for you: a spend decision, a message that has to come from your account, a judgment call between three options it already researched. As the agent works, it flags those in the project's state file — and the Workbench collects them from every project into one list.
Each card carries its provenance, so you are never asked to act on a bare instruction. Expand Why I believe this and you see the evidence the agent recorded and the session it came from. Cards sort overdue-first, then oldest, and each one exits in a single tap — Done once you've handled it, Delete when it stopped mattering. Tapping the card itself opens that project's chat with the decision pre-filled, so answering is one message instead of a hunt for context.
The Today strip along the top lists the day's automation slots, including the ones that already fired — so a single glance covers both what needs you and what ran without you.
Every project's open decisions in one list — with the evidence behind each one a click away
Calendar (beta)
Automations you set up months ago shouldn't fire invisibly. Every enabled schedule trigger projects its upcoming runs onto a week grid, so the rhythm of the project is something you can see rather than remember. Dated commitments the agent recorded in the project state land on the same grid and drop off once they're resolved, and connecting Google Calendar brings those events into the same view.
The project agent can read this calendar too, so "what's already on the schedule" is context it plans around instead of something you have to restate.
The week ahead, as your automations will actually run it
Quick Tasks
The sidebar includes a Quick Task section for fire-and-forget interactions. Scratch projects skip the full project creation flow — useful for one-off tasks that don't need a dedicated workspace.
Self-Improving Skills
Agents create reusable skills from multi-step workflows and consult matching skills before starting similar tasks. Skills are stored as SKILL.md files in the workspace and managed through the Settings UI — another way the project gets more capable the longer it runs.
Skills like Efficient Execution, Learning Capture, and Task Planning shape how your agent works — and the agent adds its own
Built-in Tool Suite
The project agent has access to these tool categories:
Browser Automation
Built on Patchright (a Playwright fork with anti-bot-detection):
- Stealth mode: Anti-automation detection scripts injected into every browser context
- Shared profile: One browser profile across all projects — log into services once, all agents share cookies
- Accessibility-first: snapshotreturns an accessibility tree with[ref=eN]element references for reliable interaction
- 26 browser actions: navigate, click, type, fill, press, hover, select, drag, upload, snapshot, screenshot, extract, search (page), evaluate, tab management, go back/forward, reload, wait, PDF export, web search, URL fetch, batch
Your agent browsing arxiv.org — scanning for AI reasoning papers on a daily schedule
Continuous Operation & Triggers
Agents run continuously via triggers — no manual intervention needed. Create triggers through natural language in the chat:
"Watch the uploads/ folder for new .jpg files and analyze them" "Run a research scan every morning at 6 AM"
The project agent translates this into a create_trigger tool call with the appropriate type and parameters.
Trigger types:
File watch trigger: watches uploads/ for new photos and analyzes each one on arrival
Schedule trigger: a daily competitor watch dispatched every day at 2 PM — 19 runs so far
Real-world example — Health Tracker with file watch:
Left: "Watch uploads/ for meal photos and track calories." Right: Drop a photo, get instant nutritional analysis.
LLM Provider Routing & BYOK
14 providers supported out of the box:
Anthropic, OpenAI, DeepSeek, Moonshot (Kimi), Groq, Google Gemini, xAI, Mistral, Together, OpenRouter, Zhipu, Qwen, TokenDance (词元跳动 — China-mainland model router), plus a custom entry for any OpenAI-compatible endpoint (e.g., Ollama, Azure OpenAI, self-hosted models).
- SDK routing: Anthropic SDK for Anthropic, OpenAI SDK for OpenAI-compatible providers
- Per-model metadata: Display name, tier, context window, max output, capabilities (vision, tool use, streaming), pricing
- Fallback rotation: When the primary provider fails, the loop rotates to fallback providers with error classification (transient, rate limit, abort)
Autonomy & Approval System
Three autonomy presets control how much supervision agents receive:
Pick an autonomy level and set budget limits per project
Approval flow:
- Interceptor catches tool call based on autonomy rules
- Frontend shows an Approval Card with tool name, arguments, and context
- User can Approve, Deny, or Auto-approve for 10 minutes
- Per-action bypass: same tool+args auto-approved for 60 seconds
Approve agent actions from your phone — with full context and optional guidance
Cost Controls & Budget Limits
Per-project budget limits prevent runaway spending:
The agent loop tracks cumulative token usage and computes cost using per-model pricing from the provider registry. When the budget threshold is reached, the configured action fires (ask pauses the session; stop halts the agent). Budget events do not currently trigger push notifications.
Set a limit and a reset period; watch the live per-model spend and cost breakdown
Mobile Remote Control
Control agents from your phone on the local network or via a cloud relay.
Left: Project dashboard on phone. Right: Your agent completes its research after you approve from anywhere.
Local network: Scan the QR code in Settings to open Orbital on your phone via LAN.
Scan to open Orbital on your phone — same Wi-Fi network required
Cloud relay (optional): Deploy a relay server for access outside your home network. Push notifications for approval requests and agent status changes.
Credential Management
Website credentials stored in your system keychain. Your agent always asks permission before using them.
- API keys: Stored in OS keychain (keyring), masked in API responses, per-project BYOK override
- Website credentials: Metadata in credential-meta.json, values in OS keychain. Therequest_credentialtool lets agents request credentials mid-session via a secure modal — credentials never appear in chat history.
Loop Safety Guards
The agent loop includes multiple safety mechanisms to prevent runaway execution:
Desktop App & System Tray
Orbital ships as a desktop application bundled with PyInstaller:
- System tray: Agent activity status, quick access menu, running port in tooltip
- Native window: Embeds the React frontend via pywebview— no browser needed
- Daemon lifecycle: Desktop app spawns the daemon on launch, manages port allocation, cleans up on exit
- Sleep prevention: Blocks system sleep while agents are active (Windows SetThreadExecutionState), re-allows when idle
Orbital is one persistent agent bound to a project — not a chat session. It acts as a local control plane for the project's workspace, instructions, state, queue, budget, and approval rules. It plans, delegates, supervises, and records outcomes; worker agents execute against the same project context. You supervise from anywhere.
flowchart TB
UI["<b>Frontend (React SPA)</b><br/>Chat UI · Approval Cards · Settings · Files"]
subgraph daemon["Daemon (FastAPI + uvicorn)"]
direction TB
AM["AgentManager<br/><i>lifecycle</i>"]
SAM["SubAgentManager<br/><i>delegation</i>"]
TM["TriggerManager<br/><i>cron · file watch</i>"]
Loop["Agent Loop<br/><i>streaming · safety guards</i>"]
TR["Worker Transports<br/>Codex app-server · SDK · PTY · ACP · Pipe"]
LLM["LLM Provider<br/><i>OpenAI + Anthropic SDK</i>"]
Tools["Tool Registry<br/><i>shell · file · browser · triggers</i>"]
Auto["Autonomy Interceptor<br/><i>approve · deny · bypass</i>"]
AM --> Loop
SAM --> TR
TM --> AM
Loop --> LLM
Loop --> Tools
Loop --> Auto
end
Platform["<b>Platform Layer</b><br/>Windows sandbox user · macOS Seatbelt · Linux bubblewrap (planned)"]
Relay["<b>Cloud Relay (Node.js, optional)</b><br/>REST proxy · Event forwarding · Push notifications · Pairing"]
Phone["Phone"]
UI <-->|REST + WS| AM
UI <-->|REST + WS| SAM
Tools --> Platform
AM -.WebSocket tunnel.-> Relay
Relay -.WebSocket.-> Phone
Key design decisions:
- The agent owns the project: it maintains structured state, decisions, lessons, and session history so planning and accountability remain in one place
- Isolation: OS-level sandboxing (Windows sandbox user, macOS Seatbelt, Linux bubblewrap planned)
- Fail-closed interceptor: Any approval system error results in DENY, never ALLOW
- Single daemon: PID file enforcement prevents multiple instances
- Local-first: Your files and project state live on your disk. The cloud relay, when enabled, proxies approvals and events — not your files.
- Download the Orbital-Setup-*.exefrom Releases (latest Windows build)
- Run the installer and follow the prompts
- Launch Orbital from the Start Menu or desktop shortcut
Windows SmartScreen Warning
Orbital is not yet code-signed, so Windows will show a security warning:
Windows protected your PC — Microsoft Defender SmartScreen prevented an unrecognized app from starting.
Click "More info" then "Run anyway". Code signing will be added in a future release.
- Download the Orbital-*-macOS.dmgfrom Releases
- Open the DMG and drag Orbital to your Applications folder
- Launch Orbital from Applications or Spotlight
Requires macOS 13 (Ventura) or later, Apple Silicon (M1 or newer). Intel Macs are not supported by this build (the bundle is arm64-only).
Release builds are Developer-ID signed and notarized by Apple, so the app opens normally on first launch — no Gatekeeper warning or "Open Anyway" workaround needed. (If you built Orbital from source or grabbed a CI branch artifact, that build is ad-hoc signed and macOS will still ask you to approve it once via right-click → Open.)
# Clone the repository
git clone https://github.com/zqiren/Orbital.git && cd Orbital
# Install Python dependencies (Python 3.11+)
pip install -e ".[desktop]"
# Install frontend dependencies (Node.js 18+)
cd web && npm install && cd ..
# Start the daemon
python -m uvicorn agent_os.api.app:create_app --factory --port 8000
# Start the frontend dev server (separate terminal)
cd web && npx vite --host 127.0.0.1 --port 5173Open http://localhost:5173 in your browser. The setup wizard runs on first launch.
Orbital prevents system sleep while agents are actively working (via OS-level sleep inhibition on Windows and macOS). When all agents are idle, sleep is re-allowed. The system tray icon shows current agent activity status.
# Start daemon
python -m uvicorn agent_os.api.app:create_app --factory --port 8000
# Restart with fresh code
bash scripts/restart-daemon.shcd web
npm install
npx vite --host 127.0.0.1 --port 5173# Unit + platform tests
python -m pytest tests/unit/ tests/platform/ -q
# TypeScript check (zero errors expected)
cd web && npx tsc -b
# Daemon integration test
bash scripts/restart-daemon.sh
curl http://localhost:8000/api/v2/projectsKnown pre-existing test notes:
- test_e2e.py,- test_user_stories.py— require a real LLM API key set via- AGENT_OS_TEST_API_KEY
- Multi-provider LLM routing with fallback rotation
- Three autonomy presets with cascade to sub-agents
- Streaming chat with real-time WebSocket events
- Browser automation with anti-detection (Patchright)
- Continuous operation via schedule and file-watch triggers
- Natural language trigger creation
- Cloud relay with push notifications and device pairing
- Context compaction with pre-compaction memory flush
- Prefix-cache-optimized prompt assembly (v0.4.2)
- Per-project budget limits and cost tracking
- Credential management (API keys + website credentials)
- Desktop app with system tray and native window
- Agent loop safety guards (iteration cap, repetition, ping-pong, circuit breaker)
- OS-level sleep prevention during agent activity
- Sub-agent delegation with @mention routing
- Webhook triggers — HTTP endpoint that fires agent tasks on incoming webhooks
- Pipeline triggers — Chain project outputs as inputs to other projects
- Network isolation — Per-project domain allowlists enforced at OS level
- Linux sandboxing — bubblewrap enforcement
- Code signing — Eliminate SmartScreen warnings on Windows
- Auto-resume on daemon restart — Restore in-progress sessions
I loved Claude Projects. I hated that I couldn't let an agent update the project, and that it didn't live on my machine.
I loved OpenClaw. I hated the lack of control — no budget, no sandbox, no way to supervise from my phone when I stepped away.
Orbital is the thing I wanted. One agent accountable for the whole project: the plan, the decisions, the queue, the budget, and the approvals. The phone to check in when I'm not at my desk. Claude Code, Codex, and Gemini CLI as workers it can choose for the job without handing away the project's context.
Built nights and weekends while working full-time. Still very early. Feedback and issues welcome.
Orbital is proudly sponsored by Watcha (观猹) — the team behind TokenDance (词元跳动), the China-mainland LLM router built into Orbital. Thanks to this sponsorship, new users in mainland China can sign in to TokenDance with one tap during onboarding, claim free tokens, and start working immediately — no manual API-key setup.
Orbital sends one anonymous aggregate per day — counters, enums, and booleans only. Never prompts, files, paths, model output, or any project/session identifier. The exact outbound JSON is inspectable verbatim in Settings → Data & privacy, where a single toggle turns it off. The full published schema is in docs/TELEMETRY.md.
Orbital is licensed under the GNU General Public License v3.0.
Orbital — Every agent owns a session. Orbital owns the project.
Copyright (C) 2026 Orbital Contributors
This program is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation, either version 3 of the License, or
(at your option) any later version.