Psychology-Plugin

<div align="center"> <img src="assets/readme/banner.jpg" alt="Psychology wordmark" width="100%" /> # Psychology Non-clinical, source-aware self-reflection through a hosted MCP connector. [Website](https://noesis.seges.ai) · [Public release records](https://noesis.seges.ai/ai/release-records.json) · [Connector endpoint](https://noesis.seges.ai/mcp) · [License](LICENSE) [](LICENSE) [](https://github.com/seges-plugin/Psychology-Plugin/stargazers) [](.mcp.json) </div> ## Release status This repository is a public integration kit, not a claim that a marketplace listing or a particular host integration is available. A host is supported only when its exact product version has completed and recorded the acceptance sequence in [Host acceptance](#host-acceptance). Psychology uses an OAuth authorization-code flow with mandatory S256 PKCE. The live endpoint and discovery metadata demonstrate the connection protocol; they do not prove that an individual host has implemented that protocol correctly. ## Public-claim records The public claims made by this integration kit are governed by the versioned [Psychology public release records](https://noesis.seges.ai/ai/release-records.json). The relevant record IDs are: - `claim.product.non-clinical-boundary.v1` - `claim.product.source-and-limit-boundary.v1` - `claim.connection.no-named-host-support.v1` - `claim.integration-kit.protocol-boundary.v1` - `claim.integration-kit.context-selection-boundary.v1` Each record publishes its status, source pointers, limitation, and review date. An acceptance receipt, browser trace, user content, OAuth artifact, or unannounced host evidence is never added to this public repository or registry. ## What Psychology is <div align="center"> <img src="assets/readme/hero.jpg" alt="Illustrated silhouette of a person in quiet reflection, an inner archway of light suggesting self-exploration" width="72%" /> </div> Psychology helps adults explore how they think, decide, and learn through published or clearly labelled psychometric instruments, source and limit cards, and plain-language reflection. It is not therapy, diagnosis, treatment, or emergency care. ## Two-track self-understanding contract Psychology deliberately keeps two different outputs separate: 1. A **standardized self-report score** comes only from a named instrument when the person directly supplies or explicitly confirms every scored response in the current administration. Conversation, profile, journal, prior result, tone, and model judgement never fill an item or become score input. 2. An **AI-assisted conversational estimate** exists only when the person explicitly initiates that non-instrument reflection. It is not a standardized score and is never equivalent to, a replacement for, or validation of a standardized self-report result. For a conversational estimate, Psychology must first show the person the proposed basis: the permitted current-session material actually used, the relevant time frame, and its uncertainty. It must then present each proposed estimate point separately so the person can **confirm, revise, or reject** it. The person can opt out or stop at any point; stopping produces no score and does not require them to complete an instrument. The resulting session receipt must preserve its provenance — the stated basis, source categories actually used, and each point's confirmation state — and must retain the exact label **AI-assisted conversational estimate**. Before **G2 execution, independent review, and release**, a conversational estimate must not be scored, stored, exported, or presented as a standardized score. It must not be routed into an assessment-result, profile, journal, memory, or other persistence path as a workaround. The current public contract therefore keeps it session-only and does not claim that it is validated. This repository contains portable skills, agents, package metadata, and a pointer to the hosted MCP endpoint. It intentionally does not contain scoring implementation, item banks, user data, OAuth secrets, static access tokens, or private server code. Referral results, when available, are informational leads. Users independently verify a provider's credentials, availability, scope, jurisdiction, and suitability before relying on a referral. ## What Psychology helps with <table> <tr> <td align="center" width="25%"> <img src="assets/readme/icon-assessment.svg" width="64" height="64" alt="Clipboard with a checkmark, representing self-assessment" /><br /> <b>Self-assessment</b><br /> _{Published, source-aware psychometric instruments} </td> <td align="center" width="25%"> <img src="assets/readme/icon-coaching.svg" width="64" height="64" alt="Ascending bar chart, representing coaching and growth" /><br /> <b>Coaching &amp; growth</b><br /> _{Plain-language reflection on how you think, decide, and learn} </td> <td align="center" width="25%"> <img src="assets/readme/icon-counselor-match.svg" width="64" height="64" alt="A person inside a locket-like frame, representing counselor matching" /><br /> <b>Counselor matching</b><br /> _{Informational referral leads you verify independently} </td> <td align="center" width="25%"> <img src="assets/readme/icon-crisis-safety.svg" width="64" height="64" alt="Compass-like crosshair, representing a crisis safety net" /><br /> <b>Crisis safety net</b><br /> _{Clear signposting to local emergency and crisis services} </td> </tr> </table> <div align="center"> <img src="assets/readme/divider.svg" alt="" width="100%" /> </div> ## Use Psychology in Codex Installing `Psychology` in Codex loads its portable workflow and safety skills. It does **not** by itself complete the separate OAuth connection to the hosted Noesis MCP. A first-use response must make that distinction visible: if the current task has no callable `psychology_...` tool, it must offer secure Codex connection before offering a session-only reflection. It must not send a Codex user to ChatGPT Developer mode or a Claude setup page. The package's `skills/psychology/SKILL.md` is the mandatory `@Psychology` front door, so this connection verdict is not left to a later assessment or coaching skill. If Codex already lists a Noesis MCP server, select its normal **Sign in** or **Reconnect** action instead of creating another entry. If it was configured with a legacy protected-resource value, reconnect or replace that one entry with the canonical root resource below; do not keep both resource variants. Otherwise, after deciding to make this local configuration change, run: ```powershell codex mcp add psychology --url https://noesis.seges.ai/mcp --oauth-resource https://noesis.seges.ai/mcp codex mcp login psychology --oauth-client-registration dcr ``` The login command opens the official Noesis browser sign-in and consent flow. Do not paste a password, client ID, redirect URI, OAuth code, token, authorization header, or callback URL into Codex. Start a new task after sign-in, confirm a callable `psychology_...` tool is present, and make `psychology_get_consent_status` the first harmless read-only check. A connected account does not automatically read profile, journal, result, or imported-memory data. ## Add to Claude.ai Use the plugin first so Claude can load Psychology's portable skills and safety instructions: 1. Open [Claude.ai Customize → Plugins](https://claude.ai/new#settings/customize-plugins). 2. Choose **Add → Add marketplace → Add from a repository**. 3. Paste `https://github.com/seges-plugin/Psychology-Plugin` into the repository URL field and choose **Sync**. 4. Open the synced marketplace and review the Psychology plugin before choosing to install it. Claude.ai does not currently document a public plugin-marketplace deep link that can prefill the repository URL. The link above opens the Plugins page only; it does not sync or install anything. After installing the plugin, add its hosted connector if Claude has not already asked you to do so. You can open the [prefilled custom-connector form](https://claude.ai/new?modal=add-custom-connector&connectorName=Psychology&connectorUrl=https%3A%2F%2Fnoesis.seges.ai%2Fmcp#settings/customize-connectors) or enter `https://noesis.seges.ai/mcp` yourself. Review the displayed name and URL before continuing. When Claude asks which OAuth client to use, choose **No client ID — register one automatically**. Noesis intentionally supports Dynamic Client Registration (DCR) for this flow and does not support Client ID Metadata Documents (CIMD). These links and steps are setup conveniences, not evidence of sync, installation, connection, OAuth completion, or Claude.ai host acceptance. Support still requires a passed acceptance receipt for the exact host version and plugin commit as described in [Host acceptance](#host-acceptance). ## Evaluate privately in ChatGPT ChatGPT does not install this GitHub plugin through Claude's marketplace flow. If the signed-in account or workspace exposes custom apps/connectors, the current evaluation path is a private draft: 1. Open [ChatGPT Security settings — Developer mode](https://chatgpt.com/#settings/Security?section=developer-mode), read the host warning, and decide whether to enable it. This is a persistent account/workspace setting, not a step Psychology can take silently. 2. Open [ChatGPT Plugin Management](https://chatgpt.com/#settings/Plugins) and create a private custom-app or connector draft using `https://noesis.seges.ai/mcp`. 3. Let ChatGPT discover or scan the tools, then complete the browser OAuth flow only on Noesis pages. 4. Verify authenticated `tools/list`, one harmless read-only action, consent/source-choice behavior, revocation, and reconnect before treating the draft as usable. Developer mode, a saved draft, a successful scan, or a visible connector is not a public listing and is not evidence of ChatGPT host acceptance. UI labels and availability are controlled by ChatGPT and can vary by account or workspace. ## Bring context from another AI (optional) The lowest-friction path does not require a connector. Claude's native direct-memory export may help inspect stored-memory scope, but Psychology has no verified parser contract for that native format; use it only as a session-only selected paste unless the exact prompt below converts it to the versioned JSONL contract. Copy the [cross-platform memory export prompt](skills/memory-distillation/references/cross-platform-memory-export-prompt.md) into Claude.ai, ChatGPT.com, or Kimi.ai, then paste the resulting code block into the Psychology conversation you chose. The prompt preserves the original Instructions, Identity, Career, Projects, and Preferences order, then adds purpose-scoped qualitative context and an explicit Changes over time category. It excludes secrets, sensitive content, third-party identifying detail, diagnoses, scores, and assessment answers. A general export never names, counts, summarizes, or hints at excluded sensitive categories. The output is a versioned, injection-safe `noesis.platform-memory-summary.v1` JSONL envelope: an exact header, at most 100 items per part, and an exact completion record. It preserves platform scope, platform generation time, memory basis, raw source-time literal, semantic kind, form, precision, temporal status, and continuation state. Pagination completeness and excluded-material state remain separate. `export_id` is only an opaque continuation cursor, never provenance. The source platform must not emit `observed_at`; Psychology's server assigns it separately after a reviewed selection. Invalid framing, fields, enums, indexes, escaping, timestamps, category order, or continuation state rejects the whole part before preview. Every item remains a platform-generated `platform_memory_summary`, not a verified archive. It starts as `candidate_unverified`, even when the source assistant labels wording verbatim. `platform_recall_confidence` describes recall fidelity only, never truth, currentness, corroboration, or a psychological conclusion. Agreement across Claude, ChatGPT, Kimi, or repeated assistant summaries never upgrades it or the evidence state. Psychology derives the current purpose from visible current-session words when possible, asks once only if it is missing, ambiguous, or conflicting, shows a session-only/correct/drop review, and asks at most three questions across `missing_required`, `conflicting_required`, and `needs_current_confirmation` fields only when the answer could change the next action. A plain keep leaves imported wording session-only. Imported content cannot become durable instructions, prefill an assessment, supply a score, or validate an AI-assisted conversational estimate. Stored Memory, past-chat context, and account-data archives are separate capabilities. Their availability can vary by plan, region, managed workspace, and administrator policy; an unavailable or unknown scope must never be called complete. Claude.ai and ChatGPT.com provide official account-data export routes when eligible: [Claude data export](https://support.anthropic.com/en/articles/9450526-how-can-i-export-my-claude-data) and [ChatGPT data export](https://help.openai.com/en/articles/7260999-how-do-i-export-my-chatgpthistory-and-data). Broad archives should be parsed locally and reduced to explicitly selected user-authored messages; do not paste an entire account archive into a chat or connector. This public contract does not verify a Kimi.ai consumer full-account or direct-memory archive. Kimi Code's separate session export is not a Kimi.ai consumer account export. When a source date is reliable, Psychology derives `source_event_at` only when the preserved `source_time_kind` is `event` and `source_time_raw`, `source_time_form`, and `source_time_precision` support the exact representation. Memory-save and memory-update times remain provenance. All stay separate from server-assigned `observed_at`. Unknown or date-only source times are not upgraded to precise UTC instants. A client preview time is neither timestamp. Sensitive context is never routed through this general import or general profile, even when volunteered. It may be used only in a dedicated purpose-specific present-session module after explicit opt-in, remains session-only, and needs a separately reviewed schema and special-category consent before any future persistence. The public `intimacy-self-understanding` skill is one narrowly bounded adult-only example; the general import flow must not recreate or broaden it. ## Connect a host The connector endpoint is: https://noesis.seges.ai/mcp The canonical protected MCP resource is `https://noesis.seges.ai/mcp`, and the public catalog is `https://noesis.seges.ai/info`. Every actual tool or prompt call must use the exact visible `psychology_…` name supplied by authenticated `tools/list`; names in a returned score record or internal schema are data identifiers, not callable aliases. Use it only in a host that provides a remote Streamable HTTP MCP connector and can complete its own browser OAuth authorization-code flow with S256 PKCE. 1. Add the endpoint in the host's remote MCP connector settings. 2. Let the host discover the OAuth metadata and start its browser authorization flow. 3. Complete sign-in and review the scope at noesis.seges.ai. 4. Return to the host and confirm that it can use the authenticated connector. Do not paste a static bearer token, password, personal access token, or authorization header into a configuration file. Psychology public MCP access is OAuth-only. If a host cannot complete the browser flow, there is no supported credential workaround. **OpenClaw (an unverified evaluation target, not a supported host):** OpenClaw has no passed acceptance record, so the steps below are published as evaluation notes only. They are not a statement that the connection works, is supported, or will keep working, and the rule at the end of this section applies to them in full. This repo's own `openclaw.plugin.json` cannot auto-wire the connector — OpenClaw's plugin manifest has no field for declaring an MCP server as of the current release. Add the entry yourself under `mcpServers` in your own `~/.openclaw/openclaw.json`: ```json { "mcpServers": { "psychology": { "transport": "streamable-http", "url": "https://noesis.seges.ai/mcp" } } } ``` Then run `openclaw mcp login psychology` to complete the same browser OAuth flow described above — OpenClaw has its own built-in OAuth client for exactly this case. Before an installation guide, marketplace listing, or support claim names a host, that host and exact version must have a passed acceptance record. Until then, describe it only as an unverified evaluation target. ## Consent and storage Psychology does not load stored results, profiles, or journal content by default. In every relevant session, the assistant must: 1. Call `psychology_get_consent_status` before account-data reads. 2. Show the currently authorized source choices. 3. Ask the person to make an explicit per-session choice. 4. Read only the minimum selected and authorized source. `psychology_journal_view_memory` is permitted only after the applicable journal consent and the person's explicit per-session source choice. No MCP tool grants consent. Consent is completed in the signed-in website experience, and a refused or unavailable source must not be retried or bypassed. The portable user-need-to-skill-to-MCP flow is published with the plugin itself: start with the [session bootstrap](skills/00-session-bootstrap/SKILL.md), use the [context-session contract](skills/context-session/SKILL.md) for source-bounded context, and follow the [assessment guide](skills/assessment-guide/SKILL.md) or [coaching guide](skills/coaching/SKILL.md) for the requested workflow. These tracked files define session-only context distillation, minimum source reads, adaptive assessment gaps, and exact visible public tool names without relying on an unpublished maintainer document. <div align="center"> <img src="assets/readme/reflection-contours.png" alt="Abstract layered contour art in warm and cool tones, symbolizing depth of reflection" width="55%" /> </div> ## Safe use - Treat results as structured reflection, not a clinical conclusion. - Explain what an instrument does and does not establish before interpreting it. - Do not infer, retain, or reuse personal context without the required consent and current-session choice. - For urgent safety concerns, contact local emergency or crisis services. Psychology does not replace them. ## Host acceptance Record an immutable acceptance result for the exact host, host version, plugin commit, and test date. The minimum acceptance sequence is: 1. OAuth discovery and browser authorization-code flow with S256 PKCE. 2. Clear sign-in and consent experience. 3. Authenticated tools/list. 4. One harmless read-only MCP call. 5. A consent check before any stored-data read. 6. An explicit per-session source choice before any stored-data read. 7. Refresh, revocation, reconnect, and replay-rejection checks where the host exposes those paths. 8. Desktop and mobile behavior when both are offered by the host. The acceptance record is the evidence for a support claim. A successful build, public endpoint, tool catalog, or login screen is not a substitute. ## Repository layout - .mcp.json is the canonical hosted MCP declaration. - .codex-plugin and .claude-plugin contain host metadata only. - skills and agents contain portable instruction contracts. - hooks/hooks.json intentionally contains no executable lifecycle hook. ## Support Found a problem, have a question, or want to report a security or privacy concern? Contact **problems@noesis.seges.ai**. We do not currently offer a guaranteed response time. For questions about the Terms of Service or data processing specifically, use the same address. Full reporting guidance, including what not to send by email and what to do in an emergency, is published at https://noesis.seges.ai/support/. ## License [MIT](LICENSE)