Security and data handling
Private code, résumés and email deserve a private system
Conductor handles exactly the data a services company will not put in a shared SaaS. It is single-tenant by design, role-gated on the backend, audited end to end, and able to run its model inside your network.
Access
- RBAC on every endpoint
- Admin, executive and developer roles are enforced on the backend for every route. The console's guards are convenience only, never the security boundary.
- Developers see only their own data
- Task queries are scoped on the server; a task that is not yours returns not found, not forbidden.
- Tokens that expire and rotate
- Access tokens live in memory for thirty minutes; refresh tokens rotate with reuse detection and are revoked on password change or account disable.
- Rate-limited, audited sign-in
- Login, refresh and SSO exchange are rate-limited per client, and failed logins are audited with uniform timing.
Integrity
- Grounded, never narrated
- Task completion is a fixed rule over merge state and review outcomes. A model is never asked whether work is done.
- Third-party text is data
- Every prompt treats pull-request descriptions, emails and documents as data: a PR claiming the criteria are met is not evidence.
- Verified webhooks
- GitHub signatures are checked in constant time on the raw body; the payload is size-capped before parsing and every write is idempotent.
- Everything audited and overridable
- 77 event types record each autonomous action and each manual override with actor, timestamp and reasoning.
Data
- Single-tenant by design
- Each subscription is an isolated stack: its own database, encryption keys, credentials, mailbox and model endpoint. Nothing is pooled.
- Secrets encrypted at rest
- Integration secrets are Fernet-encrypted in the database and never serialised back out; a production boot refuses default or weak secrets and hides the API docs.
- Local inference available
- Run the model through LM Studio and no prompt, résumé, diff or email leaves your network.
- Assist, never surveil
- Profiles and emails inform assignment and communication; no raw prompt or response is stored in telemetry, and uploads are size-capped.
Every action, with its reason
77 audit event types. Agent actions are blue-pencil entries; overrides are red. Each carries the actor, the timestamp and the reasoning, so a decision can always be explained later.
ActivityThe audit log
- Agenttask.auto_completedTASK-142
PR #87 merged into main; review 9f3c2a1 approved; acceptance met.
- Userassignment.approvedTASK-152
Aiko owns the mobile capture flow.
- Agentrisk.detectedMeridian field app
TASK-151 blocked 6 days; milestone in 5 days.
- Overridetask.force_completedTASK-138
Delivered via hotfix branch outside the PR flow.
Ask us the hard questions
Bring your security questionnaire. We answer from the code, not from a brochure.