Skip to content

Repository files navigation

VibeGo

VibeGo

Your local coding agent, available from any browser.
Run the agent next to your workspace and continue the same conversation from a laptop, phone, tablet, or foldable device.

Get started · Security · 中文说明

Early access · Single-user · Local-first · Building toward the first public release

VibeGo is a lightweight agent app for developers who want remote control without moving their coding workspace into an unbounded cloud shell. The daemon stays close to your files, while the Web console gives you a conversation-first way to start, observe, approve, cancel, and recover work.

Why VibeGo?

  • Your workspace stays nearby. The agent runs on the computer that owns the project instead of requiring a second remote coding environment.
  • One conversation, any screen. Use the same responsive console from a desktop, portrait monitor, phone, tablet, or foldable browser.
  • Safe defaults for uncertain work. Approval, path guards, sandbox rules, workspace boundaries, and output limits are part of the run—not an afterthought.
  • Settings instead of config archaeology. Model access, workspace choices, permissions, and run limits are configured from the authenticated Web UI.

Architecture at a glance

The user sees one Web console. Behind it, VibeGo keeps the execution boundary on the local development computer and sends only bounded, resumable run events back to the browser.

flowchart LR
    B["Browser<br/>Desktop · Phone · Tablet · Foldable"]
    H["VibeGo Host<br/>Launcher + same-origin Web"]
    D["Local daemon<br/>Auth · Run manager · API"]
    L["Agent loop"]
    C["Context manager"]
    M["Model provider"]
    P["Approval + policy"]
    S["Sandbox resolver"]
    T["Guarded tools<br/>Files · Git read-only · optional container shell"]
    E[("SQLite event store")]

    B -->|"Pairing + HTTPS/SSE"| H
    H --> D
    D --> L
    L --> C
    L --> M
    L --> P
    P --> S
    S --> T
    D --> E
    E -->|"Replay by sequence"| B
Loading

In practical terms:

  1. The Host starts the local daemon and the compiled Web console.
  2. You pair a browser and choose a workspace in Settings.
  3. The Agent loop calls the selected model within the context and run limits.
  4. Tools pass through policy, approval, workspace, and sandbox boundaries.
  5. The browser receives resumable progress and terminal events; it never needs to run the agent itself.

Can I use it now?

VibeGo is evolving toward its first installable release. You can already try the local workflow, the browser conversation experience, and its safety boundaries while release packaging and broader deployment validation continue in parallel.

Status What it means
Available now Local daemon, React Web console, pairing, model setup, workspace selector, conversation runs, streaming/replay, cancellation, approval cards, recovery retry, LAN/TLS gate, guarded filesystem/Git-read-only paths, and bounded container-shell wiring.
In progress DeepSeek provider-specific search/reasoning compatibility, complete Goal Control workflows, TencentDB memory sidecar promotion, and cross-platform external-sandbox evidence.
On the roadmap Signed installers, upgrade/rollback, ACME and OS certificate managers, Tailscale/SSH adapters, native mobile clients, and broader device/accessibility validation.

Get started

One-click start (Windows)

Double-click start-vibego.bat in the repository root. It finds a Node.js runtime — a portable copy under .ready4vibe/runtime, one already installed on PATH, or an official Node.js LTS zip it downloads into .ready4vibe/runtime — then installs dependencies, builds the workspace when needed, starts the Host and opens the console. Nothing is installed system-wide and all state stays inside the repository. Ctrl+C stops the Host. The same flow is available on any platform as pnpm launch.

Source checkout (manual path)

Requirements: Node.js >=22.12.0 and pnpm 11.9.0.

pnpm install
pnpm build
node scripts/host-launcher.mjs --open

The Host opens the local console at http://127.0.0.1:8787 unless an explicit port is selected. This is the source-based path while the signed, one-click installer is being prepared.

Start your first conversation

  1. Open the URL shown by the Host.
  2. Click Connect in one click — on loopback the browser pairs with the daemon automatically; manual code entry remains available as a fallback.
  3. The first-run setup wizard opens automatically when no model is configured: step 1 verifies and saves your provider key (DeepSeek or any OpenAI-compatible endpoint), step 2 selects or adds the workspace directory on the daemon computer, step 3 confirms everything is ready. It can be dismissed and re-opened later from the setup banner.
  4. Choose New conversation, describe the task, and send it.
  5. Review approval cards before a guarded tool runs. You can cancel, reconnect, or explicitly retry a recovered run from the same conversation surface.

Normal setup does not require editing .env, YAML, or JSON files. Provider keys are sent through the authenticated setup action, kept in daemon process memory, and are not placed in browser storage, URLs, events, or logs.

Security by default

  • Loopback first: the daemon binds to the local computer by default.
  • LAN is opt-in: remote LAN access requires an explicit setting and TLS by default; pairing and request protections remain enabled.
  • Untrusted work fails closed: untrusted content cannot silently select a host tool path or bypass the external-sandbox requirement.
  • Every tool is bounded: paths, arguments, environment propagation, workspace roots, approval decisions, timeouts, and output sizes are checked.
  • Secrets stay out of the UI ledger: credentials, private keys, raw model responses, and full tool output are not written to browser storage, run/Goal events, or release artifacts.

See the security defaults and LAN access decisions for the detailed boundaries.

Remote access

Connection Default Current boundary
Same computer Enabled Loopback HTTP/HTTPS with pairing
Local network Disabled Explicit opt-in, TLS required by default, pairing still required
Public Internet Follows the public-deployment milestones ACME, reverse-proxy, and operational hardening are being delivered as dedicated adapters
Tailscale / SSH Planned Reserved adapter boundary; no second agent runtime

Do not expose the daemon directly to the Internet until the public-deployment and certificate gates are complete. The advanced LAN/TLS guide contains the operator-only environment and certificate details.

What VibeGo is—and is not

VibeGo is a local-first, single-user browser console for a coding agent. It is not a hosted multi-tenant service, not a replacement for your editor, and not an unrestricted remote shell. The browser controls a daemon that keeps the workspace, approvals, sandbox decisions, and durable run history on the host computer.

For contributors

The project is developed module by module. Each substantive change is expected to have a spec boundary, focused unit tests, synchronized documentation, and an independent Git commit.

# Fast inner loop for one affected module
pnpm check:module -- @ready4vibe/model-openai

# Repository verification before a larger handoff
pnpm verify

Use the contributing guide, architecture, implementation status, and roadmap for engineering details. Release readiness is tracked separately from feature implementation; see the release publishing spec.

Documentation

VibeGo is still being built in public. Treat the status section above as the source of truth for what is usable today, and treat the detailed specs as the source of truth for constraints and future milestones.

About

Minimal, local-first agent harness with secure LAN access and a responsive Vibe Coding web console.

Topics

Resources

Contributing

Stars

2 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages