Enterprise & security

How SightLine runs, who sees what, and what it keeps.

A plain account for your security and platform teams. Forward it freely.

Observe

Every action, by every agent.

A live feed of every turn across Claude Code and Codex: what was read, run and changed, by which developer, in which repo. Session replay, per-repo tree coverage and per-file history. A “needs attention” view surfaces denied calls, failed tests, upstream errors, risky commands, new MCP servers, unidentified sessions and flagged results.

Govern

Rules that alert or block.

Deny a command, protect a path, catch secrets sent to the model, alert on personal data sent to the model, and allow only approved MCP servers. Blocking rules are enforced on the developer’s machine, before the action runs, in Claude Code and Codex. Every rule change is audited.

Rules & compliance →
Prove

Line-level authorship for every commit.

Each commit’s lines are classified as written by AI, AI then edited, written by people, or carried over, and rolled up by developer, repo and model. Tool results are also checked for prompt-injection indicators. Attribution exports to CSV.

Deployment

Hosted by Omega. A dedicated single-tenant instance: its own VM, its own database, its own org secret. TLS on every connection to the instance; daily disk snapshots.

In your cloud. The same deployment, installed by Omega in any cloud environment you run. SightLine’s records never leave your infrastructure.

Access

Sign-in goes through your company’s single sign-on (SSO). Admins see the whole organization. Lists never include content, and every time an admin opens someone else’s content, the view is logged. Rule, setup and identity changes go to append-only audit logs.

Data handling

WhatStored?How longWho can see it
PromptsStored, secrets scrubbed30 days, then purgedAdmins and the developer
Agent edits (text)Stored, secrets scrubbed30 days, then purgedAdmins and the developer
Commands and tool callsStoredCommand summaries kept; full inputs 30 daysAdmins and the developer
Tool resultsFirst 4 KB, secrets scrubbed30 days, then purgedAdmins and the developer
Repository file namesStored (names only)KeptAdmins and the developer
Commit metadata and line attributionStored, incl. message subject (never the line text)KeptAdmins and the developer
Token countsStoredKeptAdmins and the developer
AI replies and reasoningNot storedn/an/a
Commit diffsNot stored (analyzed in memory)n/an/a

Retention defaults to 30 days and is configurable. Daily backups can hold content up to 7 days longer. Exports contain no prompts, file content, tool input or output, or commands.

Redaction and leak detection

Known secrets are scrubbed before storage: Anthropic, OpenAI, GitHub, GitLab, AWS, Google, Slack and Stripe live keys, JWTs and private keys. Redaction never changes what the agent sends to the model. Leak alerts also cover card numbers, US Social Security numbers and IBANs, and name the kind of data, never the value.

Encryption

Encrypted in transit and at rest. Connections to the instance use TLS, and stored data sits on encrypted cloud disks.

Supported agents

Claude Code and Codex today, with the full code features: commit attribution, file history and rules on the developer’s machine. Any AI agent your company runs can also run through SightLine for the same record, rules and audit trail. SightLine never stores the model’s replies.