Written for the solution architect, enterprise architect or security reviewer who has to sign off on a vendor that will read their source code. Every control here says whether it is in place today.
In placePrivate deploymentIn progress
At a glance
ARC runs on Amazon Web Services in US East (Ohio). Each organization's code, findings and search indexes are kept apart, every request is checked against the organization in the caller's signed token, and nothing you upload is used to train a model.
AreaWhat ARC doesStatus
Tenant isolationStorage prefixed per organization and project; a separate vector index per project; access checked on every request.In place
Model trainingYour code, findings and prompts are never used to train models, by ARC or by the model provider.In place
Code search embeddingsComputed inside ARC with open models. Your code is not sent to a third party to be indexed.In place
Multi-factor sign-inRequired for every account, with an authenticator app. It can't be switched off.In place
EncryptionAt rest with customer-managed KMS keys, rotated; in transit over TLS 1.2 and 1.3 only.In place
Audit trailEvery infrastructure action recorded, validated and write-locked for a year.In place
Dedicated environmentsSingle tenant, your own AWS VPC, on-premises or black box.Private deployment
Amazon BedrockClaude models called inside your own AWS account, so prompts never leave it.Private deployment
Single sign-onSAML and OIDC with your identity provider.In progress
SOC 2 Type IIControls implemented on AWS; audit underway.In progress
System architecture
How the hosted service is put together, from the people and agents who call it to the data it keeps. Every box inside the dashed line runs in ARC's own AWS account; nothing in it is shared with another vendor.
Watching all of itCloudTrail in every region, write-locked 365 daysGuardDuty threat detectionRequest logs with user and organization
Only the passages a task needs
Model providerClaude via Anthropic's API, no training on your data · Amazon Bedrock in your own account for private deployments
Tenancy architecture
The hosted service is multi-tenant, isolated by organization. Isolation is enforced at three layers, so a mistake in one does not expose another tenant's data.
One request, four checks
1Signed tokenThe organization comes from the token's claims, never the URL or body
2Membership checkNot a member? "Forbidden", before anything is looked up
3Team visibilityTeams can limit which projects a member sees
4Partitioned dataorg / project / … paths and the project's own index
1. Identity decides the tenant
The organization a request may touch comes from the caller's signed sign-in token, never from a URL parameter or a request body. Every organization endpoint checks membership before it looks anything up, and answers a stranger with "forbidden" rather than revealing whether a project exists. Inside an organization, teams can restrict which projects their members see.
2. Storage is partitioned by organization and project
Uploaded source, database schemas and every finding are stored under a path that starts with the organization and the project. Search indexes are not shared at all: each project gets its own vector index, created with the project and destroyed with it.
3. Only ARC's own operators cross tenants
One platform role can see more than one organization, for support. Every request it makes is logged with the person and the organization it touched.
What a project owns
Source and schema uploads, the findings and reports, a dedicated vector index, and its own credentials for calling the model provider. Deleting the project removes all of them together.
Deployment options
Choose how much of ARC runs in your boundary. The product, the phases and the deliverables are the same in all three.
Hosted, multi-tenant
The default. Fastest start.
ARC's AWS account · us-east-2
Shared application tierSigned-token tenant checks on every call
Claude via Anthropic APICommercial terms, no training
Included
Single tenant
Your own ARC environment, run by us.
Dedicated environment
Application tier for one organizationNo shared compute with other customers
Dedicated storage and keysYour own buckets and KMS keys
Model provider
Anthropic API or Amazon BedrockYour choice per environment
$1,000 / yr
Your VPC, on-premises or black box
ARC runs entirely inside your boundary.
Your AWS account or data center
ARC application and workersDeployed into your VPC, your network rules
Your storage, your keysCode never leaves your account
Amazon Bedrock over PrivateLinkModel calls stay on the AWS network, in your account
$75,000 / yr
Dedicated and private deployments are arranged with your account team rather than switched on in the app. Black box means ARC delivers the platform into an environment we can't reach after handover.
Model choice and Amazon Bedrock
ARC doesn't send your whole repository to one large model and hope. It routes each task to the model that fits it, and the routing is configuration, not code, so it can change per deployment without a release.
TaskModel
Indexing, rule extraction, chatClaude Haiku: fast and inexpensive for the many small reads a large estate needs.
Architecture and capability mapsClaude Opus with a long context window, for the few decisions that need the whole system in view.
Code and text searchOpen embedding models (Jina code, MiniLM text) running inside ARC. No third party sees your code to index it.
Where your prompts go
Your codeUploaded or connected to your project
ARC, in your tenantIndexes locally, sends only the passage a task needs
Model endpointAnthropic API, or Amazon Bedrock in your AWS account
Cited resultStored back in your project, linked to file and line
Why Bedrock matters to a security review
On Amazon Bedrock the model runs inside AWS and your prompts and outputs are not shared with the model's maker. In a private deployment ARC calls Bedrock from your own account over AWS PrivateLink, so the call stays on the AWS network under your IAM policies, your CloudTrail and your region choice. Your security team audits one boundary, not two vendors.
No training, anywhere. Neither ARC nor the model provider trains on your code, findings or prompts.
Least context. Retrieval sends the passages a task needs, not whole repositories.
Swap without re-platforming. Because routing is per task, a deployment can move a task to another approved model when your policy changes.
Identity and access
ControlDetailStatus
Sign-inAmazon Cognito with threat protection enforced: compromised-credential and risky sign-in checks on every login.In place
MFAAuthenticator-app codes required for every user, enrolled at first sign-in.In place
PasswordsAt least 12 characters with upper and lower case, a number and a symbol.In place
RolesOrg admin, Operator, Contributor and Viewer, with an editable permission matrix. Teams limit which projects members can see.In place
Single sign-onSAML 2.0 and OIDC with Okta, Entra ID and other identity providers. Enterprise plans, priced by active users.In progress
Encryption and monitoring
ControlDetail
Encryption at restObject storage encrypted with customer-managed AWS KMS keys, rotated automatically; public access blocked at the account level. Job records encrypted with their own key, with point-in-time recovery.
Encryption in transitHTTPS only, under a TLS 1.3 and 1.2 policy; plain HTTP is redirected. Certificates managed by AWS Certificate Manager.
SecretsHeld in AWS Secrets Manager, injected at runtime, never in code or images.
Audit trailAWS CloudTrail across all regions with log-file validation, stored in a write-once bucket for 365 days.
Threat detectionAmazon GuardDuty watches the AWS account; its findings are routed to alerting automatically.
Application logsEvery request carries a request id, the user and the organization, so any action can be traced to a person.
Retention and deletion
Your code and findings are kept for the life of the project, so discovery results stay searchable while you plan and build. You decide when that ends.
Your code, from connection to deletion
ConnectRead-only credential, deleted after the copy unless you keep it for re-runs
IndexParsed and embedded inside ARC; nothing sent out to be indexed
Analyse, plan, buildKept encrypted and isolated for the life of the project
Delete the projectEvery stored version, its index and its credentials go together
Application logs 14 to 30 daysAudit trail 365 days, write-lockedModel training never
DataHow long, and how it goes
Source, schemas, findingsUntil you delete the project. Deletion removes every stored version, the project's vector index and its model-provider credentials, together.
OrganizationArchived first, then permanently deleted on request.
Application logs14 to 30 days.
Infrastructure audit trail365 days, write-locked.
Model trainingNever. Not by ARC, not by the model provider.
Deletion on request and the DPA
Ask for a full deletion, or for our Data Processing Agreement, at contact@paracorp.io. Engagements run under mutual NDA.
Compliance
ARC's controls are designed to the SOC 2 Trust Services Criteria and implemented in our AWS infrastructure today: mandatory MFA, managed encryption keys, immutable audit logs, threat detection and least-privilege roles. Our SOC 2 Type II audit is underway.
FrameworkPositionStatus
SOC 2 Type IIControls in place; audit underway. Control summary available under NDA.In progress
GDPRData Processing Agreement available. Data processed in the United States; EU-only processing via private deployment.In place
Security questionnairesYour format, answered by an engineer rather than a sales team.In place
MCP for your agents
ARC's MCP server gives your coding agents read-only access to what discovery found: business rules, the code index, the legacy source and the spec pack. It runs as its own service, accepts the same signed sign-in tokens as the app, and enforces the same organization checks, so an agent sees exactly what its user may see and nothing else. Responses are size-bounded.
Need the full security pack?
Architecture diagrams, the control list, our DPA and answers to your questionnaire. Book a call with an engineer and bring your security reviewer.