Privacy Policy

1. Summary

Agent Burndown is hosted observability for AI coding agents. Our collector runs on your machines, reads local agent telemetry, and forwards metadata — counts, durations, tokens, cost, and privacy-safe labels — to our service so your organization can see what its agents cost.

• By default we do not collect prompts, completions, source code, tool inputs or outputs, secrets, environment values, unredacted file paths, or raw remote URLs.
• Two features collect additional metadata only after an organization administrator explicitly opts in: a sanitized skills/plugin/MCP inventory, and exact identifiers for unrecognized models.
• Telemetry is scoped to your organization. Any member of your organization can see telemetry attributed to any other member.
• The one free-text field that can pass through is a truncated error_message (see §7).

2. Who we are

Agent Burndown ("Agent Burndown", "we", "us") provides the hosted service at app.agentburndown.com and this website at www.agentburndown.com. For telemetry that an organization enables about its own workspaces and members, the organization is the data controller and Agent Burndown acts as its processor. For account, billing, and website data described below, Agent Burndown is the controller.

3. Account, organization, authentication & billing data

To run your account we store, in AWS DynamoDB:

Category	Fields	Purpose
Organization	id, name, slug, feature flags (auto-enroll, inventory, model diagnostics), created date	Tenancy and configuration
User	email, role (admin/member), active flag, billing-seat flag, created date, PBKDF2 password hash	Authentication and access control
Multi-factor auth	TOTP secret, hashed recovery codes, email one-time-code hashes	Account security (MFA is mandatory)
Collector credentials	SHA-256 hash and prefix of each collector key, label, device name, expiry (default 90 days)	Authenticating your collectors; the full key is shown once and never stored
Collectors	machine label, reporting user email, version, status, last-seen, pipeline counters	Fleet health
Billing	Stripe customer ID, subscription ID, plan, status, seat count, interval, period end	Subscription management
Early access	email, name, company, team size, plan interest, use case	Waitlist (does not create an account)
Audit & alerts	admin action log, alert rules and events	Security and operations

Payment cards: card and bank details are handled by Stripe through Stripe Checkout and the Stripe Customer Portal. Agent Burndown never receives or stores your full card number.

4. Default operational telemetry

With a collector installed, the following metadata is collected by default. It is emitted per event, summarized per session, and aggregated into rollups. No opt-in is required for this tier.

Field	Meaning
event_id, event_name, timestamp, session_id	Event identity, type, time, and the agent session it belongs to
model	Model identifier reported by the agent (see §5)
tool_name, tool_success, tool_duration_ms	Which tool ran, whether it succeeded, and how long it took
mcp_server, mcp_tool, mcp_server_tool_count, mcp_schema_tokens	MCP server/tool names and schema sizes
skill_name	Skill attribution when the agent supplies it
cost_usd	Reported cost
input_tokens, output_tokens, cache_read_tokens, cache_create_tokens	Token counts
repo	Privacy-safe repository/workspace label (see §6)
error_message	Truncated error text — the only free-text field (see §7)
Attribution	user_id, collector_id attach each event to a reporting user and machine
collector_version	The collector's own build version, attached once per uploaded batch (not per event) and recorded on the resulting rollups. Used to facet usage and data quality by which collector build produced it — for example, spotting degraded data tied to an older build. Optional; batches from collectors that omit it roll up under an unknown-version bucket. Carries no information about your machine, code, or agent activity.
Session summaries	Per-session totals: timestamps, repo, model breakdown, outcome, token totals, cost, tool-call and error counts
Rollups & metrics	Additive counters by hour, repo, model, source, user, and collector; an allow-listed set of Claude Code metrics (token usage, cost, commit count, lines-of-code count)

5. Model identifiers & privacy-safe mapping

The model identifier reported by your agent is stored with each event. Separately, the collector sends a model-mapping health report that, by default, contains only normalized model-family status and counts (which providers/families were seen, event and token counts) — not raw identifiers for models it does not recognize. Exact identifiers for unrecognized or custom models are collected only under the opt-in model diagnostics feature described in §8.

6. Repository & workspace labels

The repo label is derived on your machine from the working directory: it is the project's directory name only. Full filesystem paths and Git remote URLs are never read or sent, and values that look like paths or credentials are rejected. Note that this label is not hashed or anonymized — a distinctive project folder name is transmitted as plain text, so avoid encoding sensitive information in directory names.

7. What we exclude by default

By default, the collector does not collect any of the following:

• Prompts and model completions
• Source code and file contents
• Tool inputs, arguments, and outputs
• Secrets, credentials, and environment variable values
• Unredacted local file paths
• Raw Git remote URLs

One documented exception: the error_message field carries agent-reported error text, truncated to roughly 2 KB. A misbehaving agent could place fragments of other content into an error string, so we treat this as the single free-text field and truncate it. We do not make an unconditional promise that no free text ever transits this field.

8. Optional, opt-in data

Two features collect extra metadata only after an organization administrator explicitly enables them. Members can view but cannot enable, export, or delete these.

Sanitized inventory (skills, plugins, MCP, context)

• Fields: kind, name, source, optional description, version, script/tool counts, schema token counts, presence flags — allow-listed metadata only.
• Excluded: credentials, paths, prompts, environment values, and schema contents are never collected.
• Retention: bounded by your plan's history window; disabling consent or deleting the inventory removes stored snapshots immediately.
• Controls: admin-only enable, export, and delete; every action is audit-logged.

Model diagnostics (unrecognized identifiers)

• Fields: exact identifiers of models the mapping did not recognize, with event counts, so mapping coverage can improve.
• Retention: 30 days, and deleted immediately when consent is revoked.
• Controls: admin-only consent; identifiers are sanitized against paths, URLs, and credentials.

9. This website

www.agentburndown.com uses Google Analytics 4 (measurement ID G-GPJ9N8HBDT) to understand aggregate traffic, and loads Google Fonts. Both cause your browser to share your IP address and request data with Google. The site does not require an account and does not collect telemetry through the collector.

10. How we use data

• Provide the dashboards, reports, and alerts you request
• Attribute cost and usage by user, repo, model, tool, and session
• Operate, secure, and troubleshoot the service
• Bill your subscription and manage seats
• Communicate service and account messages

We do not sell personal data and do not use your telemetry to train models.

11. Who can access your data

• Within your organization: there are two roles, admin and member. Data is isolated per organization, but there is no per-user privacy boundary inside an organization — any member can see telemetry attributed to any other member through dashboards and reports. Admin-only actions include inviting/removing members, managing keys and billing, exports, and the opt-in features above.
• Auto-enrollment: if an admin enables auto-enroll, telemetry reported for an email that has no account creates a login-less member record, and that activity becomes visible to the organization.
• Agent Burndown staff: access is limited to operating the service; a small platform-admin allowlist can view the cross-organization early-access waitlist only.

12. Retention

Dashboards and reports return data within your plan's history window — 30 days on Free, 90 days on Solo Pro, and 365 days on Team and Team Plus. Important: for raw events, session summaries, and rollups, this window is enforced by limiting what queries return; the underlying records currently remain stored beyond the window so that plan changes are safe. They are not automatically purged. Short-lived data (dedupe markers, device-pairing codes, one-time codes) expires automatically, and the opt-in inventory and model-diagnostics data have the deletion behavior described in §8.

13. Your choices, export & deletion

• Stop collecting: revoke collector keys or uninstall the collector at any time.
• Opt-in data: admins can export or delete the sanitized inventory, and delete model-diagnostics data, from the app.
• Reports: admins can export aggregated CSV reports (which never include prompts).
• Removing a member revokes their access and collectors but preserves historical telemetry already attributed to them.
• Full account or raw-data deletion / export: self-service deletion of an entire organization or a full export of raw telemetry is not yet available in-app. Contact us at support@agentburndown.com and we will handle these requests.

14. Subprocessors

Provider	Purpose
Amazon Web Services	Hosting and storage — DynamoDB (data), Lambda, API Gateway, S3 (static site), CloudFront (CDN), SES (email), SSM (secrets), Route 53, ACM
Stripe	Billing, subscriptions, and payment processing
Google	Analytics (GA4) and Fonts on this website

15. Security

• TLS 1.2+ enforced for all traffic in transit
• Encryption at rest using AWS-managed keys
• Passwords hashed with PBKDF2-HMAC-SHA256 (200,000 iterations); mandatory MFA (TOTP or email one-time code)
• Collector keys stored only as SHA-256 hashes, shown once, revocable, and expiring by default
• Per-organization tenant isolation; session tokens are HMAC-signed and short-lived
• Ingest hardening (compressed-body and batch-size limits) and audit logging of sensitive admin actions

No system is perfectly secure; we work to protect your data but cannot guarantee absolute security.

16. Sharing & disclosure

We share data only with the subprocessors above to run the service, and when required by law or to protect rights and safety. We do not sell personal data. If Agent Burndown is involved in a merger or acquisition, we will provide notice before your data becomes subject to a different privacy policy.

17. Data location

Data is processed and stored in Amazon Web Services in the United States (region us-east-1).

18. Changes to this policy

Each version carries a version identifier and effective date. When we make material changes, we will update the effective date, communicate the change, and, where consent is required, ask you to acknowledge the new version before continued use — the app records which policy version each user accepted. You can withdraw consent for the opt-in features at any time as described in §8.

19. Contact

Questions or requests: support@agentburndown.com. See also our Terms of Service.