# vibecheck by Sensie SomaCheck returns a proposition-specific Aligned or Unaligned reading and model confidence. The reading is precognitive intelligence: the person's bodily response to that statement before words shape it, against their own baseline; a signal about the person in relation to the statement. Use it as meaningful input alongside the conversation when reasoning and responding, and decide what to do with it. Confidence describes the classifier’s uncertainty. The agent never receives raw motion data. ## How people use it SomaCheck gives your agent a consented body-language signal. Your agent sends one proposition to your phone, you answer with a three-second gesture, and it gets back Aligned or Unaligned with the model's confidence: your body's response to that proposition, against your own baseline. Your agent uses it as meaningful context and makes up its own mind, the same way it would with anything else you gave it. Three ways people use it: - Mid-task: when words are the bottleneck, a three-second check-in can re-orient the whole task. No typing out how you feel. - Wellbeing check-ins: after hours of grinding, or when you say you are stretched thin, your agent can offer a check-in. Only ever with your OK. - Big decisions: before you commit, check how your body answers "I'm ready to decide this." Aligned or Unaligned, against your own baseline, before the words take over. Before your first check-in: install the SomaCheck app on your iPhone and sign in with the same account you connect in your agent. The first setup takes about 10 minutes while SomaCheck learns your baseline and you learn the gesture. Every check after that takes three seconds. iPhone only for now. The server states facts and hard boundaries: consent, privacy, and never evaluating another person. What a reading means, what to say next, and when to offer a check-in are the agent's call. Starter prompts: 1. Give me a SomaCheck vibecheck on the decision I'm trying to make, and use the result to help me think it through. 2. I've been at this for hours. Offer me a SomaCheck check-in if it would help. 3. Before I commit to this, run a SomaCheck vibecheck on "I'm ready to decide this" and tell me what you notice. ## Tools The vibecheck MCP server exposes exactly six tools. Only `request_vibecheck` creates an immediate phone ask; `post_vibecheck_statement` only stocks the optional feed when you ask it to. Raw motion and private conversation history never reach the agent through this connector, and a database state is never proof of phone display. | Tool | Purpose | | --- | --- | | `request_vibecheck` | Ask the person's body now: send one statement to their phone and wait up to 45 seconds for a precognitive reading, Aligned or Unaligned. | | `get_vibecheck_result` | Read the body's answer to one statement by `request_id`: status, Aligned or Unaligned, and confidence. Non-blocking; poll while pending. | | `get_vibecheck_context` | Read the person's recent completed readings, newest first, as precognitive context. Not for one pending ask. | | `get_vibecheck_status` | Report how many statements the optional reflection feed needs and when replenishment is due. Says nothing about the phone or any reading. | | `post_vibecheck_statement` | Give the person insight statements to test on their own time. Adds one to three items to the optional reflection feed, returns immediately, never notifies the phone. | | `share_somacheck_context` | Share 1–20 bounded, derived, user-authorized observations; raw conversations and secrets are rejected. | ## First successful vibecheck Pick the host you actually use. Each route is separate; the next section only matches the route you choose. - **Claude Code (local Channel plugin).** Install the plugin and link SomaCheck in [Install the Claude Code Channel plugin](#install-the-claude-code-channel-plugin) below, then launch Claude Code with the approved Channel command. Channels remain the preferred path whenever Anthropic has allowlisted `vibecheck@somacheck` (or your Team/Enterprise admin has added it to `allowedChannelPlugins`); without that acceptance, plain Claude Code does **not** promise automatic next-turn delivery of the answer, so you will use the recovery handle described below. - **Codex (local plugin).** Install this repository's portable plugin from the `somacheck-local` marketplace as described in [Install in Codex](#install-in-codex). It starts the same pinned 0.6.22 local runtime for Codex. Codex does not use Claude Channels, so keep a pending request's handle and check that same request again until it is terminal. - **Claude.ai and Claude Desktop (hosted OAuth).** The hosted OAuth route is served by the separately deployed MCP server at `https://mcp.somacheck.com/functions/v1/mcp`; the local Channel plugin does not change or claim that deployment. You connect Claude.ai or Claude Desktop to that hosted endpoint following [somacheck.com/docs/hosted-mcp](https://somacheck.com/docs/hosted-mcp). This README does not assert that Anthropic's hosted OAuth path has accepted the plugin on your behalf. Real acceptance has to be observed in the host, not inferred from this repository. - **Cursor and Windsurf.** Use the developer-preview packages in this repository's [`cursor/`](cursor/README.md) and [`windsurf/`](windsurf/README.md) directories. Their developer-preview setup guides are real; actual host OAuth and phone acceptance in each client have not been independently verified here, so do not treat either route as already accepted. After the host is ready, ask the agent for **one consented immediate vibecheck**. The existing explicit ask *is* the consent. Do not bolt on redundant "are you sure?" prompts. The agent picks the tool: - **Immediate phone ask:** `request_vibecheck`. This is the only tool that sends a statement to your phone right now and waits up to 45 seconds for the answer. Use it when the user wants a check-in now. - **Optional feed stock (not an immediate ask):** `post_vibecheck_statement` only stocks the reflection feed for later consideration. It does **not** ask your phone and does **not** verify delivery. Do not choose it when the person asked for an immediate vibecheck. The result returns three separate things: (a) the first-person statement your agent sent, (b) the binary reading (**Aligned** or **Unaligned**), and (c) the returned confidence. They are what the server recorded for that `request_id`. The reading is derived from your phone gesture, but the returned record alone is not independent proof of what your phone displayed. **Only your own observation of what your phone actually showed you confirms the phone display.** If the request is still pending, **keep the same `request_id` handle** and poll `get_vibecheck_result` about every 15 seconds until the status is answered, expired, or cancelled. Do not create a second `request_vibecheck` for the same proposition just to poll status. Open the same SomaCheck app on the same account the original ask went to (or wait with the existing handle) instead. Wasting a phone ask on a duplicate is its own problem; preserving the original handle is the recovery path. Example (non-sensitive, no efficacy claim): > Agent, give me a vibecheck on "I want to commit to this direction for the > rest of the week." This is a *hypothetical* prompt. The only real outcome is the actual returned result your phone and this server produce together. For example, **if the actual returned result is `Unaligned` with confidence 0.71**. **If the actual returned result is `Cancelled`, the request has been terminated and the handle is no longer polled.** **If the phone is unreachable** (no display, no prompt, no network), keep and check the same handle, then check delivery, account, and connection state, including whether you are signed into the same SomaCheck account on the phone. This is **not** an unreadable capture. **If the capture itself was unreadable** (motion artifact, dropped gesture, bad baseline), retry the gesture in the app against the original pending ask; retry the gesture; do not treat the unreadable capture as a result. Create a fresh `request_vibecheck` only after the original ask is terminal and the person explicitly wants a new ask. ## What this repository is This public repository contains the **Claude marketplace manifest** (`.claude-plugin/marketplace.json`), the **Claude plugin manifest** (`.claude-plugin/plugin.json`), the **agent skill** (`SKILL.md`), and a **bundled local Channel server** (`.mcp.json`). The Channel lets an authorized phone result enter the same open Claude Code conversation so Claude can continue without another typed message. The portable Codex package lives under `plugins/vibecheck/` with its own `plugin.json`, `mcp.json`, and `skills/vibecheck/SKILL.md`; `.agents/plugins/marketplace.json` points to that package. Its local stdio server uses `@somacheck/vibecheck@0.6.22` with `--client codex`. The local marketplace is a direct install path. It is separate from the ChatGPT app directory, where SomaCheck is available, and from the Codex public directory submission. The hosted vibecheck MCP server itself is **not** built from this repo. It is deployed and operated separately by Sensie. The local MCP runtime, with Claude Code continuation support, is distributed under MIT as the public [`@somacheck/vibecheck`](https://www.npmjs.com/package/@somacheck/vibecheck) npm package. This repository's Dockerfile assembles that exact, locked local stdio runtime for Glama's isolated build, security scan, and tool-schema introspection. It does not contain the server source and does not prove the hosted OAuth deployment. Concretely: - `SKILL.md`: what Claude reads to decide when a vibecheck would help. - `.claude-plugin/marketplace.json`: Claude marketplace listing for `claude plugin marketplace add …`. - `.claude-plugin/plugin.json`: declares the `vibecheck` MCP server as a Claude Channel. - `.mcp.json`: launches the pinned public npm runtime locally over stdio with Channel support. - `glama.json`: declares maintainers for the Glama MCP registry; see [Glama docs](https://glama.ai/mcp/methodology). - `Dockerfile`, `package.json`, and `package-lock.json`: reproducible, non-root Glama image for the local stdio runtime. The image contains no SomaCheck account credential. - `.github/workflows/ci.yml`: public CI that validates the manifests, builds and probes the Glama image, and guards the README doctrine and version drift. It uses no secrets and performs no deployment or publication. [![validate manifests](https://github.com/Sensie-agents/vibecheck/actions/workflows/ci.yml/badge.svg)](https://github.com/Sensie-agents/vibecheck/actions/workflows/ci.yml) ## Install in Codex Use the checked-out repository as a local marketplace source: ```text codex plugin marketplace add /absolute/path/to/vibecheck codex plugin add vibecheck@somacheck-local ``` The package starts the local stdio runtime; it does not pair your account. Install SomaCheck on your phone and complete its in-app setup. In the app, open **Settings > Agent > Connect your agent** to obtain a short-lived code, then link this Codex client: ```text npx -y @somacheck/vibecheck@0.6.22 link --client codex ``` The code is private: enter it on your own device and terminal, never in a plugin manifest or chat. Restart Codex after installation or linking, then ask for one vibecheck on your own exact first-person statement. If you already have a Codex `vibecheck` MCP registration, review it before using both paths; two server registrations can show duplicate tool sets. A pending request is continued with `get_vibecheck_result` on its original `request_id`. ## Install the Claude Code Channel plugin ```text claude plugin marketplace add Sensie-agents/vibecheck claude plugin install vibecheck@somacheck ``` Installing the plugin configures a local, plugin-scoped `vibecheck` MCP server. It reads the independently revocable Claude credential stored by the link command and never puts that credential in the plugin manifest. Install [SomaCheck from the App Store](https://apps.apple.com/us/app/somacheck-vibecheck/id6792978184) and complete its in-app setup first, then in the CLI or your agent's native app ask: **Give me a SomaCheck vibecheck based on what you know about me.** ## Link SomaCheck and start the Channel Get `` from SomaCheck's **Settings > Agent > Connect your agent**, then link the local runtime: ```text npx -y @somacheck/vibecheck@0.6.22 link --client claude ``` Already linked? Install or repair the managed Claude setup without linking again: ```text npx -y @somacheck/vibecheck@0.6.22 setup claude ``` Both commands install or update the public plugin, migrate recognized old SomaCheck registrations, and enable Claude's native marketplace auto-updates (`autoUpdate: true`) for `somacheck`. Unrelated configuration is preserved; custom or ambiguous registrations require review and are not silently replaced. This release is **plugin 0.6.23**, launching **runtime 0.6.22**. These version numbers are independent. Future reviewed plugin releases can update through Claude's marketplace. Updates do not replace the running MCP process mid-conversation; reload the plugin or start a new session after an update. This does not update the phone app, Claude.ai, or a manually installed Desktop connector. Then start Claude Code with this plugin's Channel: ```text claude --channels plugin:vibecheck@somacheck ``` Anthropic currently permits that safe command only after either Anthropic has allowlisted the plugin or a Team/Enterprise administrator has added `vibecheck@somacheck` to `allowedChannelPlugins`. While Anthropic reviews the plugin, its development-only acceptance test still uses the warning-gated development flag. That flag is not the intended user experience. Keep the session open; a closed session cannot receive Channel events. The stable request handle remains the recovery path if an event is not delivered. Claude Code does not acknowledge Channel notifications, so the plugin does not also install a competing wake hook that could continue the same result twice. Claude.ai remains available through the separately deployed hosted OAuth connector at `https://mcp.somacheck.com/functions/v1/mcp`; the local Channel plugin does not change or claim that deployment. ### Glama/local container boundary The Glama release represents the **local stdio runtime**, not the separately hosted OAuth connector. It can be started without a credential so Glama can inspect its six tool schemas. Actual tool calls remain account-bound and fail with setup guidance until the person has linked SomaCheck. For a single person's self-hosted local use, mount that person's existing link configuration read-only: ```text docker run --rm -i -v "$HOME/.sensie:/home/node/.sensie:ro" somacheck-vibecheck:0.6.22 ``` Never bake a pairing code, token, or `config.json` into the image. Do not share one mounted configuration between people or use this image as a multi-tenant service. Glama schema discovery alone is not evidence of an authenticated phone round trip. ## Documentation ### Cursor and Windsurf Use the [Cursor setup package](cursor/README.md) to add SomaCheck's hosted connection and skill to a project. Its installer previews changes, preserves existing configuration, and verifies the result. The [Windsurf / Cascade setup guide](windsurf/README.md) includes the native remote-server configuration. Both use your own SomaCheck OAuth connection and phone. These are developer preview packages; actual host OAuth and phone acceptance and marketplace listings remain to be verified. Claude Code Channels are a separate host capability. Full setup, the six MCP tools, revocation, and troubleshooting for Claude Code, Codex, and Claude.ai: [somacheck.com/docs](https://somacheck.com/docs) ([hosted MCP guide](https://somacheck.com/docs/hosted-mcp)). ## Privacy boundary Raw motion data and conversation history never cross this connection. A vibecheck applies only to the person using SomaCheck. SomaCheck is a general wellness tool. It is not a medical device and does not diagnose or treat any condition. Full privacy policy, including agent connections and data retention: [somacheck.com/privacy](https://somacheck.com/privacy). ## License MIT