Education & Research

Scoutpost

by THOMAS THÉO ANDRÉ VAILLANT

Overview

Scoutpost lets journalists and researchers monitor public sources for news from ChatGPT. Create scouts that watch web pages, local news beats, public social profiles, council agendas and minutes, or ship and aircraft movements. Then search, verify and organize the facts they find, each linked to its source. Scouts run on a schedule, can send email notifications and use credits on your connected Scoutpost account, so ChatGPT asks before creating, running, changing or deleting anything. Findings stay unverified leads until you verify them.

Tools

create_project

ChatGPT

Create a new investigation project (a workspace for grouping scouts + units).

create_reflection

ChatGPT

Create a reflection (durable editorial note) over scouts, units, or entities. Embedded at write time for semantic search.

create_scout

ChatGPT

Create a new scout. Required: name and type (web|beat|social|civic|transport). Probe first: web scouts run test_web_scout, civic scouts run discover_civic_sources then preview_civic_items, and pass one of the returned candidate URLs as tracked_urls. The server re-runs step 1 before inserting and answers HTTP 422 with the probe envelope { ok:false, stage, error_code, error } when it fails: civic → no_meetings_detected (the body also carries system, validated, invalid and candidates — read candidates and retry with one of them); web → unreachable|blocked|empty_content|outside_configured_page|page_too_long. The 422 body is returned in this tool's error text. web/beat/social/civic also need either location or topic; transport needs config (not location/topic) — see below. Topic is 1-3 short comma-separated tags for organization, not long instructions. Put long human context in description and filtering/notification rules in criteria. Web scouts require url. Beat scouts should pass criteria and optionally location/source_mode/priority_sources. Civic scouts require root_domain and tracked_urls. Social scouts require platform and profile_handle; MCP defaults them to monitor_mode=criteria, which requires criteria text. Pass monitor_mode=summarize explicitly to collect all substantive new posts. Transport scouts (Fleet Scout) are Pro/Team on hosted Scoutpost and require config: { mode: aircraft|vessel, watch_ids, geofence: { center: {lat,lon}, radius_km, display_name?, maptiler_id? }, categories?, criteria? }. Every mode needs the circular area and alerts when a watched object enters it. criteria is optional and only filters those entry alerts; it never replaces the area. watch_ids is required (max 20); categories only narrow it. Run test_transport_config first and pass its baseline_ids as transport_baseline_ids here. Transport supports 3h/6h/12h/daily regularity. Scheduling: pass schedule_cron OR regularity + time (+ day_number for weekly/monthly). Scheduled creation establishes web/beat/social/civic baselines immediately; Fleet creation seeds the tested baseline when transport_baseline_ids is supplied, while omission preserves the legacy first-run baseline. Web/Page scouts can turn on evidence archiving with archive_enabled (Pro/Team); captured snapshots are then retrievable via list_snapshots.

delete_project

ChatGPT

Delete a project by id.

delete_scout

ChatGPT

Delete a scout and unschedule its cron job.

delete_unit

ChatGPT

Soft-delete an information unit by id.

discover_civic_sources

ChatGPT

Step 1 (detect) for Civic Scout creation: discover_civic_sources → preview_civic_items → create_scout. Given root_domain, maps the council site, fingerprints the committee system (moderngov|generic) and returns only listing pages behind which at least one meeting document is already visible (documents_visible ≥ 1), best first with recommended:true. Response: { ok, stage:'detect', error_code?, error?, system, candidates:[{url, description, confidence, system, documents_visible, recommended}], unverified? }. ok:false with error_code no_meetings_detected is an outcome, not a transport error: candidates is empty and unverified lists ranked pages a human can inspect by hand. Alternatively pass tracked_urls (1-2 URLs you already chose) instead of root_domain to validate them: the response then carries validated, invalid and replacement candidates, and create_scout will reject any URL listed in invalid.

get_civic_item

ChatGPT

Read one Civic accountability item and its canonical provenance summary. Use get_unit_evidence for the owner-scoped exact evidence expressions.

get_civic_promise

ChatGPT

Read one owner-scoped Civic promise, including its human lifecycle status and audited status history.

get_civic_run

ChatGPT

Read safe aggregate diagnostics for one owner-scoped Civic run.

get_project

ChatGPT

Fetch a single project by id.

get_scout

ChatGPT

Fetch a single scout by id.

get_snapshot_url

ChatGPT

Get a short-lived (5 min) signed download URL for one artifact of a snapshot. Artifacts: mhtml (full captured page, opens in Chrome/Edge), screenshot (png), rawhtml, markdown, manifest (JSON of per-artifact SHA-256 hashes), tsr (RFC 3161 timestamp token). Every URL downloads as an attachment. Returns 404 if the snapshot lacks that artifact (e.g. mhtml on a markdown_only capture) — call list_snapshots first to see each snapshot's artifacts.

get_unit

ChatGPT

Fetch a single information unit by id.

get_unit_evidence

ChatGPT

Read owner-scoped bounded exact evidence expressions for one canonical unit. Use this to verify any AI-extracted Civic lead against its official source before publication.

ingest_content

ChatGPT

Ingest a URL or raw text into the knowledge base. Creates a raw_capture row and extracts atomic information_units with Gemini through OpenRouter's Google Vertex route.

list_civic_items

ChatGPT

List owner-scoped Civic accountability items. Promises have a human lifecycle status and due date; decisions are completed adopted actions and never enter the promise tracker.

list_civic_runs

ChatGPT

List safe aggregate diagnostics for Civic runs, including policy version, semantic-zero state, counts, and rejection aggregates without exposing rejected source text or private criteria.

list_projects

ChatGPT

List investigation projects owned by the caller.

list_reflections

ChatGPT

List editorial reflections (agent-written synthesized summaries) owned by the caller.

list_scouts

ChatGPT

List all scouts owned by the caller (id, name, type, schedule, is_active).

list_snapshots

ChatGPT

List evidence-archive snapshots for the caller's Web/Page scouts, newest first. Each row: id, scout_id, captured_at, capture_kind (baseline|change), fidelity (full|rendered_thirdparty|markdown_only), sizes, trust (RFC 3161 timestamp status + Wayback status/url), and the artifacts available to download. Filter by scout_id. Snapshots exist only for scouts with archiving enabled (create_scout/update_scout archive_enabled).

list_units

ChatGPT

List information units owned by the caller. Supports project, scout, verification, usage, and deleted-state filters.

mark_unit_used

ChatGPT

Flag a unit as used in a published article so it leaves the inbox. Optionally record the URL and timestamp.

merge_entities

ChatGPT

Collapse duplicate entities into a single keeper. Use after search_entities surfaces near-duplicates.

pause_scout

ChatGPT

Pause a scout: set is_active=false and unschedule its cron job.

preview_civic_items

ChatGPT

Step 2 (sample) for Civic Scout creation, after discover_civic_sources and before create_scout. Read-only preview of 1-2 listing URLs: resolves meeting documents, parses at most two, runs one extraction pass and returns sample_items plus preview_snapshot_token (pass it to create_scout with import_current_items). Returns dated promises to follow and adopted material decisions; a valid empty result means no accountable item was found. Response carries the probe envelope { ok, stage:'sample', error_code?, error? } with valid and documents_found; ok:false codes are no_documents, unreachable (a linked meeting could not be fetched), parse_failed or model_failed. Every result is an AI-extracted lead that must be checked against the cited official source.

reject_unit

ChatGPT

Reject a unit (not wanted editorially). Sets verified=false and records the reason in verification_notes.

resume_scout

ChatGPT

Resume a paused scout: set is_active=true and re-schedule its cron job.

run_scout

ChatGPT

Trigger an on-demand scout run. Spends credits. Returns 202 + run_id.

search_entities

ChatGPT

Find canonical entities (people, orgs, places, policies) across the knowledge base. Returns entity rows with type + canonical_name.

search_reflections

ChatGPT

Semantic search over the caller's reflections.

search_units

ChatGPT

Search the caller's information units. Modes: semantic, keyword, or hybrid. Supports project, scout, verification, usage, and deleted-state filters.

set_civic_promise_status

ChatGPT

Apply a human editorial Civic-promise transition. Civic extraction never makes this judgment. Read first and pass its exact updated_at as expected_updated_at.

test_transport_config

ChatGPT

Step 1 for Fleet Scout creation. Test a transport config against current live data without creating a scout or spending credits. Returns baseline_ids and a compact preview. Pass those baseline_ids as transport_baseline_ids to create_scout in Step 2 so objects already present do not alert immediately. Hosted Scoutpost requires Pro/Team.

test_web_scout

ChatGPT

Step 1 for Page (web) Scout creation. Probe a URL the way create_scout will: one fresh scrape plus an optional criteria check, no scout created, no baseline stored. Returns the shared probe envelope { ok, stage:'reach', error_code?, error? } alongside summary, scraper_status and criteria_status. ok:false carries error_code unreachable|blocked|empty_content|outside_configured_page|page_too_long and a human-readable error — create_scout would reject the same URL with HTTP 422, so fix the URL first. criteria_not_met is advisory: ok stays true and the scout can still be created. Run this before create_scout for type=web.

update_project

ChatGPT

Patch a project — name, description, visibility, or tags.

update_scout

ChatGPT

Patch an existing scout. All fields optional; only sent keys change. Keep topic as 1-3 short comma-separated tags; put longer context in description and filtering/notification rules in criteria. A scout must retain either location or topic (transport scouts are scoped by config instead). For Fleet Scouts, pass a full replacement config with the required circle area; it alerts when watched objects enter that area, with criteria only as an optional post-entry filter. Fleet Scout configuration is Pro/Team-only on hosted Scoutpost. For web/page scouts, toggle evidence archiving with archive_enabled / wayback_enabled (archive_enabled is Pro/Team-gated).