APX v2 Agent Crew
Host your own team of AI agents on your machine: one super-agent that answers anywhere, project crews that do the work.
- Source repo
- agentprojectcontext/apx
- Stars
- ★ 35
- Last updated
- today
- License
- MIT
- Primary language
- TypeScript
- FA score
- 51/100 · Major gaps
At a glance
- Works with
- Universal · cross-platformChatGPT · Codex · Claude Code · Claude.ai · OpenAI API · Claude API
- You'll need
- Typical use
- A solo developer who messages one super-agent from WhatsApp or Telegram and has it hand research to another project agent, then open the source page in the embedded browser.
- Main limitation
- Installers are not signed yet, so macOS and Windows require extra manual approval — awkward for managed environments.
- Website
- apxagent.app ↗ set on GitHub
- Maintenance
- Commits in the last 30 dayslast commit 2026-11-20latest release v2.16.0 (2026-11-20)
- Source review
- 51/100 · Major gaps
What does this agent do, and when should you use it?
APX v2 is an open-source, local-first agent platform built around a single daemon (apps/core) that owns all state and exposes an HTTP + WebSocket protocol of roughly 260 typed methods, consumed through one contract in packages/protocol. On top of it sits a super-agent that replies in the app, on desktop, on Android, and across Telegram, WhatsApp, Discord, Slack and iMessage, delegating to per-project agents that carry a role, an area and their own notes. A task board with commitments and scheduled routines lets those agents work unattended and report back only when there is something to say. The Electron desktop app embeds a real browser the agents drive in watchable tabs, and can optionally let them use your own macOS and Windows apps while passwords and payments stay off-limits. Bring your own models — a ChatGPT plan, Claude, Gemini, OpenAI-compatible APIs or local models — and Claude Code or Codex sessions sit beside your chats. Install paths cover a desktop package, a terminal script (data in ~/.apx/v2), and a headless server deployment exposed over Tailscale or LAN and paired with devices.
Every agent turn runs through a single runTurn() path, and one tested context builder assembles the prompt, which measures about 8.2k tokens per call on a fresh install. MCP server tools are discovered through mcp_search rather than being sent one by one, and the prompt stays identical across calls so provider caching keeps working. Agent-to-agent communication is uniformly session-calling-session, and every user-facing channel goes through threads.send, tuned per agent, thread or channel. A conversation is stored as a thread plus a run plus steps. Agents can read and write project notes, claim and advance tasks on the board, fire scheduled routines, drive embedded browser tabs, operate your own apps when explicitly enabled, and build small widgets or apps (charts, checklists) that you pin to the board. Artifacts export and import as .apxart files and share links are served by your own APX. The apx CLI speaks only the protocol: apx desktop installs and opens the app, apx open opens the web panel, and apx setup --headless --expose tailscale plus apx pair wires a server to your devices.
- A solo developer who messages one super-agent from WhatsApp or Telegram and has it hand research to another project agent, then open the source page in the embedded browser.
- A small team running APX on a VPS or homelab, exposed over Tailscale, with members pairing the desktop app and Android app to their own project crews.
- Someone who needs a recurring weekly report: a scheduled routine runs unattended and only speaks up when there is something to report.
- Engineers using both Claude Code and Codex who want those sessions side by side with APX chats and tasks in one panel.
- Users who want local models or an OpenAI-compatible endpoint and prefer not to be locked into a single provider.
- People accumulating small internal tools: ask for a chart or checklist, let an agent build it, and pin the widget to the board.
Key setup facts
Pulled from the install and usage notes below. The project README has the current values.
- Install / run
curl -fsSL https://apxagent.app/install.sh | bashirm https://apxagent.app/install.ps1 | iex
- Ports
7440- Config files
~/.apx/v2- Install notes cover
- Windows · macOS · Linux
How do you install or deploy this agent?
Desktop app (macOS, Windows, Linux): download from the website or the latest v2 release and follow the setup; the apps are not signed yet, so the install guide shows the one extra click macOS and Windows ask for.
Terminal install (detects your system, no root, data in ~/.apx/v2):
curl -fsSL https://apxagent.app/install.sh | bashirm https://apxagent.app/install.ps1 | iexOn a server (VPS, homelab, no screen), pass --server after the pipe (PowerShell: -Server), then run headless setup and pairing:
bash -s -- --server
apx setup --headless --expose tailscale
apx pairBuilding from source requires Node.js and pnpm:
pnpm install
pnpm app # starts the core on 127.0.0.1:7440 and opens the desktop app
pnpm preflight # privacy + schema checks, lint, typecheck, testsHow do you use this agent?
After first launch, the setup asks whether this machine is a server or a client, then you connect a model (ChatGPT plan, Claude, Gemini, OpenAI-compatible API, or a local model). Day to day you have three entry points: the desktop app's inbox and chats, the web panel (apx open), and the phone app (APX-<version>-android.apk in each v2 release, paired with the code the desktop shows).
From a terminal checkout you can also drive the CLI directly:
apx desktop # installs and opens the desktop app
apx open # opens the web panelalias apx="node $PWD/apps/terminal/src/main.ts"To stand up a project crew, give each agent a role, an area and notes, put tasks on the board with an owner and a due date, and configure routines where it should run unattended. Letting agents use your own apps is an explicit opt-in setting; passwords and payments stay outside that scope.
What are this agent's strengths and limitations?
- A ground-up rewrite with one way to do each thing: a single
runTurn(), a singlethreads.send, one context builder and one thread+run+steps store, which keeps debugging and extension costs low. - Prompt size is budgeted and tested at about 8.2k tokens per call, the lowest among the agents it measured; MCP tools are found via
mcp_searchto keep the prompt cacheable. - Provider-agnostic: a ChatGPT plan, Claude, Gemini, OpenAI-compatible APIs and local models are all documented, with Claude Code and Codex sessions shown beside chats.
- Reaches desktop, web, Android and several messaging channels, and a self-hosted server can be exposed over Tailscale or LAN and paired with your devices.
- A documented migration path from v1: v2 takes over the
apxcommand and imports v1 projects, agents, skills, chats and channels.
- Installers are not signed yet, so macOS and Windows require extra manual approval — awkward for managed environments.
- Desktop, server and phone are separate surfaces; a headless server needs Tailscale/LAN exposure and device pairing configured by you.
- Several capabilities are explicitly foundations, not finished products:
.apxartexport is not a marketplace, and self-hosted server access is not a cloud account or login. - APX Enterprise is planned but its scope, licensing, pricing and availability are unannounced, so the long-term commercial direction is open.
- Running locally does not mean everything runs offline: cloud models and services you connect still receive what each request needs.
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| APX v2 Agent Crew This agent | 51 · Major gaps | — | ★ 35 | today | TypeScript | ChatGPT · Codex · Claude Code · Claude.ai · OpenAI API · Claude API |
| NextClaw | 61 · Some gaps | Desktop appFree + model costs | ★ 264 | 1d ago | TypeScript | Codex · Claude Code |
| somora | 58 · Major gaps | CLIFree | ★ 35 | 4d ago | TypeScript | ChatGPT · Codex · Claude Code |
| Brigade | 53 · Major gaps | Self-hosted serviceFreemium | ★ 16 | today | TypeScript | ChatGPT · Codex · Claude Code · Claude.ai |
How does FollowAgents rate this agent?
Why each dimension lost points
AGENTS.md rules 7 (loopback default, APX_TOKEN required for non-loopback), 8 (idempotency keys for side-effect methods), 10 (confirmation for strong actions, no password/payment typing, temporary consent not global), 11 (secret:// references), plus apx-mcp scoped client tokens and apx-extensions install-off with strict schema, support least privilege and confirmation (2 each). Data-flow transparency: README honestly states cloud models receive what each request needs and local does not mean offline, but no per-flow inventory (2). Sensitive data: secret:// and no-secrets-in-.apc rules, but no storage implementation shown (2). Dependency security: only devDependencies and pnpm pinning visible, no audit/lockfile policy evidence (1). External effects: curl|bash install, taking over v1's apx command and stopping v1 daemon are disclosed but no rollback path, so external_effects and rollback each 1. Source attribution: MIT and repo links present, but publisher unverified and no signing (1).
The 'one way to do each thing' principle, single runTurn()/threads.send contract, protocol method registry, and test support that validates method names and param schemas (transport.test-support) show good internal consistency (2). Dependency availability: only engines node>=24.11 and [email protected] visible, no lockfile or dependency manifest evidence (1). Failure messages: apx-mcp gives step-by-step troubleshooting for missing tools and apx status code_sha verification (2).
README and skills cover desktop, server, phone, messaging channels, Claude Code/Codex scenarios (audience_and_scenarios 2). Capability boundaries are explicit via the Works today / Foundations / Roadmap table and the 'Foundations are not finished services' caveat (2). Trigger precision: apc-context's description demands 'ALWAYS activate' and forces a migration offer, which is over-broad and coercive (1). Environment fit: macOS/Windows/Linux, no root, ~/.apx/v2 documented (2).
AGENTS.md provides a glossary, repo map and numbered rules — clear information architecture (2). Install notes cover desktop, terminal, server, phone (2). Naming is stable (core/terminal/protocol terms fixed) (2). Examples exist as command snippets but no consolidated FAQ (1). Known limitations are scattered ('apps are not signed', 'Foundations not finished') (1). LICENSE is full MIT (2). Versioning has v1/v2 branches and release badges but no CHANGELOG (1). Maintenance responsibility: publisher unverified, only Discord and issues channels (1).
Output usability: concrete CLI commands, skills and protocol methods (2). Marginal value: multi-agent crew, messaging channels, browser driving and Claude Code/Codex integration are clearly differentiated (2). Cost-benefit: README claims minimal prompt size (8.2k tokens) citing rules/PROMPT_SIZE.md, which is not in the packet, and install/takeover costs are unquantified (1).
Claim traceability: key claims (prompt size, v1/v2 comparison) cite rules/PROMPT_SIZE.md and asset charts not present in the packet, so unverifiable (1). Cross-source corroboration: AGENTS.md, SKILL.md and README agree on protocol, secret:// and confirmation rules, but performance and download claims lack independent support (1). Fact/inference separation: README separates Works today from Roadmap, yet marketing claims like 'least tokens' sit alongside facts (1).
- Install uses curl|bash and irm|iex remote code execution, stops the v1 daemon and takes over the apx command; review the script and back up v1 data first.
- Desktop apps are unsigned, requiring extra clicks on macOS/Windows, which raises supply-chain and impersonation risk.
- The apc-context skill demands 'ALWAYS activate' and forces a migration offer, risking misfires in unrelated contexts and deleting .apc/migrate.md.
- Publisher identity is unverified with no signing or provenance; performance and download claims cannot be checked within the packet.
- No evidence of dependency audit or lockfile policy in the sample; audit dependencies before production use.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
v2.16.0 APX 2.16.0
Release notes excerpt
Full Changelog: https://github.com/agentprojectcontext/apx/compare/v2.15.3...v2.16.0View on GitHub ↗ -
v2.15.3 APX 2.15.3
Release notes excerpt
Full Changelog: https://github.com/agentprojectcontext/apx/compare/v2.15.2...v2.15.3View on GitHub ↗ -
v2.15.2 APX 2.15.2
Release notes excerpt
Full Changelog: https://github.com/agentprojectcontext/apx/compare/v2.15.1...v2.15.2View on GitHub ↗
FAQ
Which port does APX v2 Agent Crew use?
Does APX v2 Agent Crew work on Windows?
irm https://apxagent.app/install.ps1 | iex.Is APX v2 Agent Crew open source?
Is APX v2 Agent Crew safe to use?
Is APX v2 Agent Crew still maintained?
Who makes APX v2 Agent Crew?
What is APX v2 Agent Crew's official website?
Thanks — we use this to improve the page.