Preview: Fleet Visibility is available in early access via The Station. Some features described here are in active development. See the Roadmap for current status.

Fleet Visibility gives engineering, security, and platform teams a unified view of every AI agent operating across your repositories — what they did, what it cost, and whether it was within policy.

What ChooChoo captures per trace

Every agent action produces a trace record. Fleet Visibility aggregates these records across agents and repositories. Each trace record includes:

Field Description
agent_id The registered ID of the agent (e.g. claude-code, gemini-cli)
tool.name The harness that fired the trace hook (e.g. claude_code, cursor)
action The action type (file.edit, file.read, validate, context.generate)
artifact The file or schema path the action targeted
file_scope List of files touched in this session
cost_signal Token count or estimated compute cost (where available from the harness)
risk_score Governance risk score (0.0–10.0) calculated from confidence, scope, and history
outcome approved, blocked, pending_approval
timestamp ISO 8601 timestamp

Fleet dashboard in The Station

The Station's Fleet tab provides a live view of agent activity. It surfaces anomalies and usage patterns that are invisible at the individual trace level.

Available now (preview):

  • Agent activity feed — timestamped list of all traces, filterable by agent, action type, and time range
  • Risk score heatmap — which agents and artifact types are generating high-risk changes
  • Agent leaderboard — traces per agent, error rate, approval rate
  • Lineage graph — interactive visualization of artifact → agent → artifact provenance

Coming in Phase 2:

  • Approval inbox — review and approve/reject blocked changes directly from the UI
  • Cross-repository view — fleet visibility across multiple repositories in one dashboard
  • Anomaly alerting — notify when an agent's behavior deviates from its historical baseline
  • Context Graph — full cross-agent context visualization showing shared skills and artifacts

Cost governance

ChooChoo tracks agent spend at trace granularity. For harnesses that expose token counts (Claude Code, OpenCode), each trace record includes a cost_signal field.

From the CLI, you can break down cost by agent and time period:

# Approximate cost by agent over the last 30 days
choochoo trace list --since 30d --format json \
  | jq 'group_by(.agent_id) | map({agent: .[0].agent_id, traces: length})'

In The Station, the Cost tab shows:

  • Token consumption by team, agent, and change type
  • Expensive operations flagged for review
  • Trend charts to catch cost spikes before they accumulate

Use [governance.approval] in choochoo.toml to gate high-cost operations:

[governance.risk]
require_approval_above = 7.0   # High-risk + high-cost changes need sign-off

Cross-team visibility

The same trace data supports different views depending on your role:

Role What they see Why it matters
Engineering Which agents touched which files, risk scores, test outcomes Catch regressions, track agent quality over time
Security Actions on sensitive paths (auth/**, contracts/**), PII field access Verify boundary enforcement is working
Legal / Compliance All agent-authored changes with attribution, approval records Evidence for EU AI Act, SOC 2, GDPR audits
Platform Cost by team, agent adoption, harness distribution Capacity planning, ROI measurement

Querying traces

CLI:

# All traces from a specific agent in the last 7 days
choochoo trace list --agent claude-code --since 7d

# High-risk traces across all agents
choochoo trace list --min-risk 7.0

# Traces touching a specific file
choochoo trace list --artifact src/auth/session.py

# Output as JSON for further processing
choochoo trace list --since 7d --format json

The Station filters (UI equivalent):

  • Agent dropdown — filter by one or more registered agents
  • Action type — file.edit, validate, context.generate, etc.
  • Time range — presets (1h, 24h, 7d, 30d) or custom date range
  • Risk score — slider to surface only high-risk activity
  • Outcome — approved, blocked, pending_approval
  • Agent Trace — Trace record schema and how traces are emitted.
  • Governance (Preview) — Risk scoring, approval workflows, and audit trail.
  • Agents — Registering agents and declaring boundaries enforced at trace time.