Skip to content

Documentation / Plugin pilot

Bring your account.
Work from the record.

A guide to the Starglass research workflow for ChatGPT and Claude Code.

Before the first pilot session

Invited testers receive the endpoint, https://aculeus.ai/api/mcp, and instructions for their supported host. Installing the client does not grant access on its own: each pilot account is enabled one person at a time. The first exercise uses public or synthetic material.

  1. Use your own host account. Its subscription and usage limits remain separate.
  2. Connect Starglass through the host's OAuth consent flow using an enabled Starglass pilot account. No host password or subscription token belongs in a research prompt.
  3. List the available sources and check their coverage. No research workspace or Case is required.

Personal ChatGPT Pro and the Claude Code plugin both completed sign-in, source listing, search, record retrieval and revocation in operator testing on 2 October 2026. There is no approved directory listing yet.

The three tools

Starglass exposes three read-only tools. sources_list shows source coverage, availability and limits. source_search searches one selected source and returns typed record identifiers. source_fetch returns a bounded passage of a returned record with its canonical URL, publisher, retrieval time, line range and integrity hash. There is no arbitrary URL fetch, no write tool and no paid research or model fallback.

Sources in the pilot: USAspending, FEC committee registrations, Congress, GovInfo, Senate lobbying disclosures and the NPPES provider registry. SEC filings are listed but not configured yet. Configured is not a health check, and an empty result does not prove absence.

The first research exercise

Ask a narrow public-record question. Select a relevant source, search for records and retrieve the record text. Check entity identifiers, dates and the record's meaning.

Build an evidence ledger in your host session. Cite the public URL and returned line range, retain the representation hash, separate supported facts from interpretation and look for contradictions. Search snippets and an empty result set are insufficient evidence.

Pause and continue

Your host manages the conversation and any local notes. Record open questions and next actions there before stopping. Host retention and continuation features vary; Starglass does not supply a shared research archive or promise unattended parallel research.

Permissions and data

The core tools only read public records. Starglass keeps account access, client connections and request counters. It does not save research queries or source bodies in a Starglass research workspace. Source systems receive the queries needed to answer a request; your host controls its own conversation retention.

Revoke a client in Research connections to block new Starglass operations, and disconnect it in your host to clear host credentials. In operator testing, a revoked client was refused and a restored one reconnected.

Allowances and price

The private pilot starts with a bounded allowance of 1,000 source operations per person per month and 60 requests per minute across connected clients. Failed acquisitions and retries count; no paid overage or fallback starts automatically. Some sources require an operator-provisioned API key.

$49 per user / month is an unactivated pilot price hypothesis. Final pricing follows measured retrieval, hosting and support costs. The core path has no Starglass-paid model or Parallel run; infrastructure and acquisition still have costs.

Existing hosted Deep Research plans remain a separate offering.