Local archive
Read-only repository collection and verifiable local evidence archives.
Security, privacy, and deployment
Local-first trust. Governed team and enterprise modes later.
Source code, issue discussions, decisions, incidents, and agent transcripts are sensitive. Privacy and access control are product features, not implementation details.
Local-first default
Validate usefulness without requiring hosted code transfer.
The first context and anchor implementation is intended to run locally against a developer’s repositories and KodeAtlas archive. This reduces deployment friction, aligns with sensitive-code requirements, and keeps the initial product operable without a centralized service.
Security requirements · Product direction
Controls have to follow the evidence into every delivery surface.
- read-only provider access by default, with optional write scopes separated;
- no tokens in archives, logs, command histories, stored URLs, or model prompts;
- repository, project, symbol, task, and transcript authorization in team deployments;
- content-addressed storage and integrity manifests for archived evidence;
- encryption in transit and at rest, with customer-managed keys where required;
- audit records for ingestion, agent queries, claims, decisions, corrections, exports, and deletions;
- retention controls for transcripts, issue attachments, incidents, and derived packets;
- model-provider policies that prevent sensitive evidence from reaching unapproved services;
- prompt-injection handling for untrusted issues, comments, documents, and transcripts;
- explicit provenance boundaries among human assertions, source facts, and model summaries.
Agent authorization · Planned
Bind claims to an authenticated execution context.
An agent should receive only the repositories, symbols, tasks, and evidence its user and execution context are allowed to access. Claims must bind to a human or service account, branch or worktree, base commit, task, heartbeat, and expiration policy.
An agent must not be able to impersonate another agent or silently extend a claim forever.
Transcript sensitivity · Future ingestion
Agent transcripts may contain more sensitive context than the repository.
Transcript ingestion will require secret scanning, configurable redaction, retention, deletion, and a choice between retaining raw text or only extracted evidence. Credentials, customer data, incident details, and uncommitted reasoning cannot be treated as ordinary search documents.
Deployment modes · Product direction
One product model, progressively governed deployments.
Local context
Embedded database, local repositories, local MCP server, and locally controlled artifacts.
Team self-hosted
Shared repositories, connectors, agent registry, PostgreSQL, object storage, web UI, policy, and audit.
Air-gapped enterprise
Customer-controlled models, no external code transfer, access policy, retention, and governance.
Take the next step
Start local. Add shared infrastructure only with explicit control.
Sensitive code intelligence should not force a team to surrender control of its source, history, or agent transcripts.