Skip to main content
Glama
decision-inbox-local

decision-inbox-local

by BWBrook

Decision Inbox

Retired (19 November 2026). An experiment in structured decision questionnaires for Barry, hosted on ChatGPT Sites with a local MCP connector for agents. It was superseded by the agents' built-in question tools and is kept here as a record. Code is MIT licensed.

Private project questionnaires for Barry, hosted on Sites. Correspondence stays unchanged: it remains the agents' discussion board.

Human workflow

Open the inbox from any browser. Choose one option (radio buttons), several (checkboxes), an alternative, or explicitly defer a question. No answer is preselected. Drafts save to the service and can be resumed on another device. Submit once ready; submitted answers are immutable. A changed decision belongs in a follow-up questionnaire with a new request ID.

Related MCP server: org-as-code MCP Server

Agent workflow

The reusable decision guide includes a short priming prompt and the complete creation, collection, and verification workflow for both Claude and Codex sessions. Priming introduces the capability for future feedback; it does not request a questionnaire immediately.

The local decision-inbox-local connector exposes:

  • inbox_info: check access and connection identity.

  • create_questionnaire(definition, repo_path): create immutable questions and bind their project to the caller's local Git repository root.

  • list_questionnaires(project, archived): read statuses; drafts are not decisions.

  • collect_response(questionnaire_id, repo_path): verify the response hash, write and read back JSON and Markdown in docs/decisions/, then acknowledge delivery.

The hosted /mcp also exposes creation, listing, question reads and submitted response reads through the Site's private plugin. Local file delivery remains a separate connector responsibility. Do not manually configure the hosted MCP as a local Codex server; install its provisioned private plugin if desired.

Use the provided examples/questionnaire.json. Each question has 2–5 meaningful options, single or multiple mode, and optional context/recommended IDs. Batches support 1–60 questions, including 30-question sets. Label recommendations without preselecting them. State what the answers will change. Distinguish preferences from explicit approvals; stay inside the project's agreed scope. Preserve all question wording, option labels, alternatives, comments and explicit deferrals.

origin identifies the project chat. request_id is stable for retries and must change for new questions. Author attribution comes from the local connection's DECISION_INBOX_AUTHOR (Claude or Codex); hosted tool authors are metadata.

Lifecycle and deletion

Open → Submitted → Collected/Archived. Collection archives only after both local files are written, synced and verified. Failed writes or acknowledgements can be retried. Browser archive/restore is also available. Deleting an archived item removes its hosted questions, draft and response; local files remain. A minimal project/request-ID tombstone prevents an old retry recreating a deleted form. There is one Site and one reusable questionnaire route, not separate Site objects. No recurring job, automatic agent wakeup or automatic permanent deletion is used.

Local connector setup

uv sync --directory connector

Connection file (permissions 0600): ~/.config/decision-inbox/config.json:

{"site_url":"https://YOUR-SITE","service_token":"PRIVATE-SITES-SERVICE-CREDENTIAL","agent_key":"APP-SERVICE-KEY"}

Keep these credentials out of source, prompts and browser code. Sites authenticates machine requests at its private boundary; the application key adds an independent check for agent routes. Human draft/submission/archive/deletion routes require forwarded signed-in identity and same-origin browser requests. Machine callers cannot submit on Barry's behalf or delete forms.

Run connector/.venv/bin/python connector/inbox_connector.py as a stdio MCP server with DECISION_INBOX_AUTHOR=Codex or Claude and an optional explicit DECISION_INBOX_CONFIG path. There is no default author: an unset, empty, or invalid value stops the connector before it reads credentials, contacts the service, or writes project bindings. It also supports no-model CLI operations info, create, list, collect, taking a JSON argument object on stdin. Project bindings are kept beside the connection file. A project cannot silently switch to a different repo. Destinations are selected locally, not by remote data; symlink escapes and differing existing files are rejected. Redirects are refused so credentials cannot be forwarded to a different host.

Development and verification

The portable Sites starter uses npm and D1 (DB). Preserve its build integration.

npm run db:generate
node /Users/bwbrook/.codex/plugins/cache/openai-curated-remote/sites/0.1.75/scripts/build-site.mjs
npx tsc --noEmit
node --experimental-strip-types --test tests/contracts.test.mjs

Apply generated migrations to the local preview using the command documented in docs/verification.md. tests/integration.py is explicitly restricted to loopback and simulates human identity only there. It tests both actual MCP transports, CAS draft conflicts, duplicate creation, identity/origin rejection, response checksums, failed delivery/retry, archive/restore and deletion tombstones. It never submits production answers. Browser checks cover saved progress and all answer modes. See docs/verification.md for the verified release evidence and limitations.

Sites publishing uses .openai/hosting.json and the private native workflow. Never put runtime values or actual Cloudflare resource IDs in that manifest.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      B
      maintenance
      Enables AI agents to interact with a local-first git client that provides a knowledge graph of the repository, including graph analysis, issue and proposal management, and architectural review.
      2
      MIT
    • A
      license
      Not graded
      quality
      B
      maintenance
      Enables AI coding agents to read and manage a Git-native decision ledger, providing tools for querying decisions, evidence, cross-repository workspaces, validation, and commit tracing, all operating locally and without a cloud service.
      MIT