Export API

The export API returns the same datasets as the app's History → Export page, for scripts, spreadsheets and your own reporting. Scope: export:read (create an export key on History → Export). What a key can export follows its creator's current role — see Export your data. Metadata only: never chat text, messages, a routine's brief, keys or notification addresses. The exception is the work record of projects with Work details on: work and work_summary carry its instructions and the agent's summaries (first 600 characters, secrets masked) for a key whose person may approve the project.

#GET /v1/export

Lists the datasets this key may export, with their columns.

curl "$AGENTOWER_URL/v1/export" -H "authorization: Bearer $AGENTOWER_EXPORT_KEY"

{
  "datasets": [
    {
      "name": "agents",
      "label": "Agents and subagents",
      "description": "One row per agent or subagent: its short ID and name, project, person, tool, model, state and why, when it started, last signalled and ended, active time, tokens and ≈ cost.",
      "dateFilter": true,
      "columns": ["agent_id", "name", "kind", "parent_agent_id", "subagent_type", "project", "person", "tool", "model", "model_name", "state", "reason", "state_since_utc", "started_at_utc", "last_signal_at_utc", "ended_at_utc", "active_minutes", "tokens_in", "tokens_out", "cost_usd_estimate", "machine", "guardrail_state", "guardrail_note", "imported"],
      "url": "/v1/export/agents"
    }
  ],
  "usage": "GET /v1/export/{dataset}?from=2026-09-01&to=2026-09-30&format=json|csv&limit=1000&offset=0 (times in UTC)"
}

The datasets are agents, state_changes, work, work_summary, audit, usage_daily, projects, members, guardrails, channels, routines, routine_runs, computers, keys, my_data (what Agentower keeps about the key's own person, across their workspaces) and plan_accounts (plan limits per account and window). Not exported: agent chats (never — chat text is held only while the agent is open) and routines' briefs.

#GET /v1/export/{dataset}

from and to apply to datasets with "dateFilter": true.

curl "$AGENTOWER_URL/v1/export/agents?from=2026-09-01&limit=1" \
  -H "authorization: Bearer $AGENTOWER_EXPORT_KEY"

{
  "dataset": "agents",
  "columns": ["agent_id", "name", "kind", "parent_agent_id", "subagent_type", "project", "person", "tool", "model", "model_name", "state", "reason", "state_since_utc", "started_at_utc", "last_signal_at_utc", "ended_at_utc", "active_minutes", "tokens_in", "tokens_out", "cost_usd_estimate", "machine", "guardrail_state", "guardrail_note", "imported"],
  "rows": [
    {
      "agent_id": "DPP-01", "name": "Sarah", "kind": "agent", "parent_agent_id": null, "subagent_type": null,
      "project": "data-pipeline", "person": "Alex", "tool": "nightly-import", "model": null, "model_name": null,
      "state": "done", "reason": null,
      "state_since_utc": "2026-09-30T09:12:04.000Z", "started_at_utc": "2026-09-30T09:00:00.000Z",
      "last_signal_at_utc": "2026-09-30T09:12:04.000Z", "ended_at_utc": "2026-09-30T09:12:04.000Z",
      "active_minutes": 12, "tokens_in": 0, "tokens_out": 0, "cost_usd_estimate": 0,
      "machine": null, "guardrail_state": null, "guardrail_note": null, "imported": false
    }
  ],
  "count": 1,
  "nextOffset": 1
}

agent_id is the agent's short ID, which never changes; name is its first name — empty for subagents. Only open agents' names are unique, so a finished agent may share a name with another.

When there are more rows, nextOffset is the offset for the next page (and the x-next-offset header says the same); on the last page it's null.

curl "$AGENTOWER_URL/v1/export/state_changes?from=2026-09-01&to=2026-09-30&format=csv" \
  -H "authorization: Bearer $AGENTOWER_EXPORT_KEY"

at_utc,agent_id,project,person,model_name,from_state,to_state,reason

CSV comes as a download (content-disposition: attachment), with a header row. Cells that a spreadsheet could run as formulas are made safe.

#The work record dataset

work has one row per step of the work record — a file changed, a command, a test run, a commit, a push, a pull request, "and N more", or the end of a turn — in time order. Subagents' steps have parent_agent_id set; person is the top-level agent's person. A column is null when the step doesn't have it, or the tool couldn't measure it (never 0 in its place).

The turn_end row also carries the turn's totals (files, lines_added, lines_removed, commits, pushed, pull_requests and the last test result). Never code, diffs or command lines — Agentower doesn't keep them.

#The work summary dataset

work_summary is the work record rolled up by work stream and ISO week (Monday to Sunday, UTC) — the same numbers as History → Work. One row per stream, project and week, newest week first. The date filter picks the steps that count. One request covers at most 56 days (8 weeks): a wider range is cut to its newest 56 days, and the answer says so in an x-agentower-note header (and a note field in JSON). Ask for older weeks with their own from and to. Building the summary reads every step in the range, so those rows count against the workspace's daily export budget (see Errors).

The counts are for everyone who may export. commit_subjects, first_instruction and agent_summary_not_verified are filled in only for projects with Work details on and only for a key whose person may approve the project; otherwise null.

#Routines, computers, keys, my data and plan limits

approved_on_computer says whether the version the computer last said it approved is the current one; it's null until the computer has said anything, and approval_seen_at_utc is when it last did. profile_picture_url is an address in the app (/api/avatars/…), not part of the API: open it while signed in to download the picture; it's null when there's no picture.

#Errors

Exports through the API are recorded in the audit log, at most once per key and dataset every 10 minutes.