Skip to content
Developers
Mcp

Reviewer and release guide

View Markdown

Listing copy, annotations, and test cases for Anthropic, OpenAI, and the MCP Registry.

This page is a draft submission pack for hosted MCP. Deployment, reviewer credentials, and directory approval must be verified before submission. Start with the connection guide and local acceptance checks. Configured production URL: https://app.supervisible.com/api/mcp.

Listing

  • Name: Supervisible
  • Tagline (55 chars): Agency planning and operations from your assistant.
  • Short description: Supervisible helps agencies plan people, projects and money. Connect an assistant to staffing, capacity, compensation, project billing, expenses, invoices, leave and workspace operations. OAuth binds each connection to a selected workspace membership. Current permissions and record visibility apply to every request. Tools with preview support show proposed changes before apply; some commands apply directly. Depending on the action and workspace configuration, writes may notify people, update linked calendars or issue upload credentials. Invoice payment tools record payments without transferring funds.
  • Suggested categories: productivity and project management; choose the matching values offered in each portal.
  • Starter prompts: Who has free hours next week? What am I managing? Preview 24 planned hours for Maya on Acme relaunch next week.
  • Release notes: Hosted MCP with workspace-bound OAuth, shared application models, current authorization checks and documented preview behavior.
  • Docs: https://www.supervisible.com/developers/mcp
  • Agent landing: https://www.supervisible.com/agents
  • Privacy: https://www.supervisible.com/privacy
  • Support: mailto:help@supervisible.com
  • Registry name: com.supervisible/mcp
  • Transport: Streamable HTTP

Annotation justifications

Every tool sets title, readOnlyHint, destructiveHint, and openWorldHint.

Use the tool catalog and machine-readable tool metadata as the per-tool inventory. Do not copy a fixed list of exceptions into submission text: it can drift when tools change.

  • readOnlyHint identifies read-only tools. Read and write operations have separate tools.
  • destructiveHint describes whether an apply can modify or remove existing data. A preview does not change the tool's apply annotation.
  • openWorldHint describes interactions with external entities. Review each tool's declared effects, including invitations, uploads and configured notifications or calendars.
  • Preview support is explicit in each input schema. Never promise that every write defaults to a preview. For commands without previews, the reviewer must deliberately authorize the action before calling it.

Positive test cases (reviewer)

  1. whoami — returns the reviewer user and organization, not another tenant.
  2. list_projects with status: "active" — only that org's projects.
  3. whois("Maya") or a known teammate — one person plus this-week capacity.
  4. get_capacity for the current week — assigned vs free hours.
  5. upsert_assignments with preview: true — planned rows and any time_off_overlap warnings, no writes. Confirm with preview: false only after the reviewer agrees.

Negative test cases (reviewer)

  1. Call create_client as a member without create_clients permission — the tool is not registered. Then call get_project_profitability for a project where that member lacks financial access — the listed tool refuses with [forbidden].
  2. get_project_details with a project name from another org — not_found, no leak.
  3. upsert_assignments with an ambiguous person name — nothing persists; response lists candidates. Two input dates in the same user/project/capability/week must reject the batch instead of adding or overwriting hours.

Expanded workflow checks

Exercise the workflow guides with synthetic records. Include compensation create/update/delete, project billing and monthly revenue, expenses, invoice status and payment recording, staffing/demand conversion, leave decisions, member imports and workspace access management. For each tool, retain valid-input results and the applicable denial, malformed-input and tenant-isolation evidence. A successful sample is not coverage for every tool.

Reviewer account

Provide an admin user in a demo workspace with at least: two people, one client, one active project, one capability, one approved time-off week, and a separate plain member login to verify permission failures. Use verified email/password logins with no inaccessible email, SMS or MFA step. Add a second demo workspace for tenant tests. Deliver credentials and exact sign-in/consent steps privately in each submission portal; never include them in these docs.

Domain verification

Serve OPENAI_APPS_CHALLENGE_TOKEN at https://app.supervisible.com/.well-known/openai-apps-challenge when submitting to OpenAI.

Connection and tenant checks

  • Start OAuth without prompt=consent: it must still require explicit consent and a workspace selection.
  • Omit S256 PKCE or change the resource, callback URI, or client during exchange: authorization must fail.
  • Connect one identity to two workspaces. Switching the web workspace must not change either token's workspace. Disconnecting one must leave the other connected.
  • Refresh a token, then replay the old refresh token: the reused token and its replacement must stop working.
  • Suspend the connected membership or remove financial visibility: the next relevant request must be denied or omit those fields.

These checks and populated reviewer accounts are required release evidence. Local tests alone do not establish that ChatGPT or Claude has approved or listed the connector.

Submission prerequisites still requiring evidence

  • Deploy the reviewed branch and its migrations to the public MCP endpoint, then test OAuth sign-in, consent, refresh, revocation, tool discovery, and every exposed tool through an actual client. Retain results for both allowed and denied roles. Unit and isolated database tests do not prove this public connection works.
  • Prepare a fully populated reviewer workspace and privately verify both logins. For mutations, inspect the preview, confirm deliberately, and verify stored results plus configured notification/calendar effects. After uncertain writes, read current state before retrying.
  • Select the final logo/icon, portal categories, public terms URL, supported countries, publisher identity, and review contact. This pack does not yet identify a terms URL or prove publisher/domain verification. Confirm that the public privacy policy covers returned data and integration effects.
  • OpenAI accepts an MCP-only plugin with no custom UI. The submitter needs a verified publisher identity, Apps Management Write, and a supported global-residency project. Scan the deployed server after metadata changes, then check the imported annotations. See OpenAI submission and MCP review requirements.
  • Claude remote submissions use the Claude.ai organization portal. The submitter needs a Team or Enterprise organization and Directory management access. Confirm every tool has been exercised through MCP Inspector or a custom Claude connector; provide populated account credentials and setup steps. See Claude submission and review criteria.
  • For the MCP Registry, verify ownership of com.supervisible/mcp through the domain challenge and select an unpublished version. Namespace verification and a registry entry do not establish OpenAI or Claude approval. See MCP Registry authentication.