# September 11th Documents Toolkit

Independent research tools for New York City's [September 11th Document Portal](https://sept11documents.cityofnewyork.us/). The local MCP server searches catalog metadata, finds the folders filed under a building, reads captured pages, checks and formats citations, and follows the City's commitments and dates. The watchdog compares accepted catalog captures and reports observed changes. The site is at [pranava0x0.github.io/sept11documents-mcp](https://pranava0x0.github.io/sept11documents-mcp/).

**Purpose.** This toolkit was built so that New Yorkers, first responders, survivors and their families can read what the City's records show about the air in Lower Manhattan after September 11, 2001. It also lets them check what the City has committed to in releasing those records. Each quote is verbatim from a captured source, a Bates-numbered portal page or an official page, and the build refuses one it cannot locate there.

This is an independent project. It is not an official City of New York website, is not affiliated with 9/11 Health Watch or any official, and provides no legal advice.

## The site

`docs/` is the GitHub Pages site. `make site` serves a local preview at [localhost:8000](http://localhost:8000). The pages, in menu order:

- **Overview** (`index.html`): the portal, the purpose, the commitments quoted from the stipulation, Resolution 560-A, the September 8 transcript and release and DOI's letters, the three announced amounts, the dated obligations and the watchdog capture.
- **The MCP** (`toolkit.html`): the tools, recorded exchanges with the local server, the commands, how to watch a call happen, and the portal's API.
- **Examples** (`examples.html`): demos built on the MCP, in four groups. *Find records*: a cited record, the folder-label search, the folders filed under a building, a portal search builder, collections and research routes. *Read and cite*: public statements beside City records in date order, sampling results as printed, three citation formats for a page, and a reading list with a CSV download. *Help and money*: the official proof-of-presence directory, each announced commitment through five evidenced stages, and the dated obligations with a calendar file. *Releases*: the saved capture, release observations, the catalog footprint and a dated context scan. Each panel names its tools, the command that reproduces it and the recorded response.
- **Records** (`records.html`): the documents located so far, each with its quote and Bates page.
- **Community support** (`community.html`): links to organizations serving 9/11 responders, survivors, and families of fallen responders.

The demos read dated files in `docs/data/` and recorded server responses, and call no live server. A guide that routes a person to a proof-of-presence path waits on practitioner review.

## The MCP server

```bash
python3 scripts/sept11_mcp.py doctor                    # what this machine holds; no network
python3 scripts/watchdog.py run                         # one catalog export, stored as the local snapshot
python3 scripts/sept11_mcp.py call catalog_search --args '{"folder": "John Street", "count": 5}'
python3 scripts/sept11_mcp.py configure claude-code --dry-run   # also claude, cursor, vscode, codex
python3 scripts/portal_api.py search '"Clean Up Initiative" extension:pdf' --count 5
```

Eighteen tools, eight prompts and eleven resources. The newer tools answer the questions people ask most often: `building_lookup` (what the City filed under an address), `timeline_lookup` (statements beside records), `readings_lookup` (sampling results as printed), `presence_evidence` (what each program accepts), `upcoming_dates` (with an iCalendar file), `portal_query_draft`, `citations_format` and `records_manifest`.

`skills/` holds six instructions an assistant can load for one kind of question each: verify a quote, building records, statements and records, proof of presence, follow the money, and release watch.

The server uses the Python standard library only. `uv pip install .` installs a `sept11-mcp` command; point it at a checkout with `SEPT11_ROOT`. `server.json` is the registry entry.

The server answers from what this machine has captured: the accepted catalog snapshot, captured page text and the files on the publication allowlist. A fresh clone holds no catalog snapshot until `python3 scripts/watchdog.py run` makes one export. A document whose page text has not been captured returns `not_cached`. `portal_search` reaches the City only when enabled with `SEPT11_ALLOW_LIVE=1`; its response identifies the recipient and the query.

## Layout

| Path | Contents |
|---|---|
| `sept11/` | The package: `core/` evidence and citation models, `adapters/portal.py`, `storage/` (immutable snapshots, cache, budgets, publication allowlist), `mcp/` (stdio server) |
| `scripts/` | Thin CLI entrypoints, listed in `scripts/README.md` |
| `skills/` | Six assistant skills, one folder each, that name the tool calls for a kind of question and the rules for the answer |
| `schemas/` | `envelope.schema.json`, generated from the models; every tool result validates against it |
| `docs/` | The GitHub Pages site: five pages, one stylesheet, one script, and the published data in `docs/data/` |
| `review/` | `anchors.json` (located documents) and `publication.json` (the allowlist of files any public surface may serve) |
| `evals/` | `golden.json`, the golden evals seeded from `review/anchors.json` |
| `research/claims/claims.json` | The claim registry: each quote and number on the site with its source |
| `research/sources_manifest.json` | The primary sources the server can cite by id, with their URLs |

## Checks

`make check-public` runs every gate this repository can satisfy on its own; CI runs it on each push. `make check` adds claim verification, the golden evals and the freshness gates for generated pages. Those read captured evidence, including the page text of City documents, which stays out of the repository because it can contain personal information about private individuals.

## Sources

Every number on the site traces to a record in `research/claims/claims.json`, and each primary source is listed with its URL in `research/sources_manifest.json`. Records on the portal are public records of the City of New York; this project links to the City's copies and does not re-host them.