VibeShare

Share your AI subscriptions with friends — peer-to-peer, no central server.

VibeShare is a native macOS menu-bar app. Run it and you get a single local OpenAI-compatible endpoint (http://127.0.0.1:8788/v1) that:

1. Runs a local OpenAI-spec server your tools can point at.
2. Connects your LLM providers (Claude, ChatGPT/Codex, Gemini, Kimi, …) via one-click OAuth — powered by the bundled CLIProxyAPI.
3. Shares usage with friends by generating a code they paste in. Sharing is one-directional: your code lets them use your models; you only get theirs if they send you a code.
4. Shows who's online and actively routing — discovered over public Nostr relays, connected directly over WebRTC (P2P).

When a friend requests a model you don't have, VibeShare auto-routes it to an online friend who shares it, streaming the response back over a direct, encrypted WebRTC channel. No VibeShare server exists — friends connect straight to each other.

See ARCHITECTURE.md for the full design.

---

How it works (30-second version)

your tools ──▶ 127.0.0.1:8788/v1  (VibeShare router)
                      │
        ┌─────────────┴──────────────┐
        ▼                            ▼
  local model?               a friend shares it?
  → cli-proxy-api            → WebRTC to their Mac → their cli-proxy-api
    (your subscription)        (their subscription)         │
                                       └──── streamed back ──┘

Connection setup (the WebRTC offer/answer) is signaled through public Nostr relays. Presence ("I'm online, here are my models") is broadcast as encrypted heartbeats only holders of the grant code can read. Once connected, traffic flows directly peer-to-peer over a DTLS-encrypted data channel.

A grant code is the only credential. From it both sides derive the same Nostr room and encryption keys. Treat codes like passwords.

Requirements

• macOS 13 (Ventura) or later
• To build: Go 1.23+ and Swift / Xcode command-line tools

Build & run

# 1. Download the provider engine (cli-proxy-api) into Resources
make fetch-provider

# 2a. Quick dev build + run (debug)
make swift
./src/.build/debug/VibeShare      # menu-bar icon appears

# 2b. Build, replace the installed app, and launch it (quit VibeShare first)
make run

# 2c. …or package a distributable disk image (VibeShare-<version>.dmg)
make dmg                          # this machine's arch
make universal-dmg               # Intel + Apple Silicon

make install copies it to /Applications. Without a Developer ID it is ad-hoc signed (fine locally; right-click → Open the first time). Quit a running VibeShare before installing, then launch the new app. Replacing its signed files while the old menu-bar process is still running can leave Little Snitch matching the old process identity against the new build and block all relay connections. Check the relay count and a borrowed connection after relaunching.

make dmg builds the app and wraps it in a drag-to-install disk image (app next to an Applications shortcut). With a Developer ID it signs the image; set NOTARIZE=1 (plus NOTARY_PROFILE, or APPLE_ID/TEAM_ID/APP_PASSWORD) to also notarize and staple it for distribution to other Macs.

Using it

1. Click the menu-bar icon → Providers → Connect your subscriptions (opens a browser OAuth flow; credentials are stored locally in ~/.cli-proxy-api). Disconnect deletes that stored sign-in again — reconnect whenever you like.
2. Point any OpenAI-compatible tool at http://127.0.0.1:8788/v1 (no API key needed locally).

Share with a friend

• Share tab → New → name it, pick what to share (everything, specific providers, or specific models) and an optional token allotment → Create code. Send them the VS-… code.
• They open VibeShare → Borrow → Enter code. Within seconds you appear online in their list, and your shared models show up in their /v1/models.
• Need your quota back for a while? Pause a friend instead of revoking — the code stays valid and you can resume any time. Revoke kills the code permanently and marks it revoked on their Borrow tab when their app sees the revoke notice.
• The Activity tab shows every request routed through your shares (both directions) as it happens.

Borrow from a friend

• Borrow tab → Enter code with the code they gave you. Their models now resolve through your local endpoint automatically.
• If a friend revokes a code while you are online, VibeShare marks that borrowed connection revoked, removes it from routing, and leaves the row so you can remove it yourself. Hosts also re-announce recent revokes while running, so borrowers who come online later can still learn about the revoke.
• Your local model normally wins when you and a friend both have the same model. For Claude/Codex models, if VibeShare knows your local 5-hour session is exhausted and a non-exhausted friend shares that exact model, it routes to the friend instead. If several friends share the same model, requests stick to the friend who served it last (keeps their provider-side prompt cache warm), prefer whoever has the most allotment left, and automatically fail over to the next friend if one is unreachable or out of budget.
• If this Mac only borrows a provider and its local sign-in is unusable, turn on Prefer a friend's models over local models in Settings (or run vibeshare config set prefer-borrowed on). A shared model then goes directly to the friend even when the local engine still lists it. The default is off.

Use with Claude Code

Claude Code speaks the Anthropic Messages API (/v1/messages), which VibeShare routes the same way as the OpenAI path — local first unless the local Claude session is known exhausted, otherwise to an online friend who shares the model. Point Claude Code at the gateway:

ANTHROPIC_BASE_URL=http://127.0.0.1:8788 ANTHROPIC_API_KEY=vibeshare claude

ANTHROPIC_API_KEY only needs to be non-empty (the loopback endpoint doesn't check it). Tip: wrap it in a shell function so your normal claude is unchanged:

vibeclaude() { ANTHROPIC_BASE_URL=http://127.0.0.1:8788 ANTHROPIC_API_KEY=vibeshare claude "$@"; }

Agents and the CLI

vibeshare drives the running menu-bar app over the loopback control API. Claude Code, Codex, Hermes, and Grok can share, borrow, pause, and change settings without clicking the menu. vibeshare help lists every command. Add --json (or set VIBESHARE_JSON=1) when a program is reading the result.

vibeshare status
vibeshare doctor             # app, router, and provider engine health
vibeshare models             # all models exposed at the local endpoint
vibeshare usage              # every available subscription limit window
vibeshare connections          # friends whose models you are borrowing
vibeshare grants               # codes you have given out
vibeshare grants pause <id>    # stop lending to one friend; borrows stay up
vibeshare env                  # ANTHROPIC_BASE_URL / OPENAI_BASE_URL exports
vibeshare logs router          # last 50 router log lines
vibeshare app restart --yes    # full restart after relay changes

Revoking a code, dropping a borrow, and disconnecting a provider require --yes. Provider sign-in (vibeshare providers login claude) opens the same browser flow as the menu and does not run unless you ask for it. vibeshare --json returns the router's JSON for reads and writes. CLI-only commands return an object (doctor, logs, app) or model list (models). Errors exit nonzero and write {"ok":false,"error":"..."} to stderr. The CLI covers sharing, grants, connections, provider sign-in, usage, activity, models, logs, and router settings. macOS launch-at-login and native notification preferences are still managed in the app's Settings tab.

Subscription bars show every limit window returned by a connected provider, including separate model-family windows such as Claude's weekly Fable limit. Share cards show the host's windows for the providers in that grant. Borrow cards show the same windows only when the host opts into the windows usage-sharing level. Off shares no subscription details; resets shares only reset times for providers paused by the reserve gate. A friend still running VibeShare 1.0 will not send these bars until they upgrade. These percentages describe the host's subscription, separate from the token allotment assigned to a particular friend.

Version 1.1 speaks the same offer, answer, and data-channel messages as 1.0. A Mac that has installed 1.1 keeps working with a friend who is still on 1.0. The newer side is more patient when the path blips and puts ICE candidates in the SDP, which 1.0 already applies. Install 1.1 on both sides when you can: both ends then wait out a short disconnect instead of hanging up immediately.

Ports & state

Port	What
8788	VibeShare OpenAI endpoint (point your tools here)
8799	Router control API (used by the app UI; loopback only)
8317	cli-proxy-api provider engine (internal; loopback only)

State lives in ~/.vibeshare/ (config + grants + connections + logs). Grant files contain secret codes — never commit or share them.

Project layout

router/     Go engine: OpenAI front server + P2P routing (pion/webrtc + go-nostr)
src/        Swift menu-bar app (manages the two Go binaries, drives the UI)
scripts/    fetch-cliproxyapi.sh

Status

This is an initial, working version (P2P routing of streaming completions is verified end-to-end). Rough edges / planned next steps:

• Provider connection currently relies on cli-proxy-api's OAuth flows; the "connected" indicator is inferred from the live model list.
• WebRTC uses public STUN only — very restrictive NATs may need a TURN server.
• Grant codes are bearer credentials; pausing a grant suspends access (the code survives), revoking ends it permanently.
• Relays can be edited in Settings; saving restarts the router automatically (relays only apply at router start).

Credits

Provider authentication and the OpenAI-compatible provider API are handled by the bundled CLIProxyAPI.

VibeShare was inspired initially by VibeProxy, and its peer-to-peer signaling draws on ideas from SeedShell. The code here is its own.

Disclaimer

VibeShare lets you share access to your LLM provider accounts with other people. Doing so may violate the terms of service of the providers involved (such as Anthropic, OpenAI, Google, and others) — many subscriptions prohibit sharing accounts, credentials, or access with third parties.

You alone are responsible for ensuring your use of VibeShare complies with the terms of every provider and account you connect or share. The authors and contributors take no responsibility and accept no liability for any consequences of using it — including but not limited to account suspension, termination, lockout, rate-limiting, loss of access, or any other action a provider may take against you or anyone you share with.

VibeShare is provided "as is", without warranty of any kind. Use it at your own risk.

License

MIT