AI coding agent that knows your codebase and rules

An AI coding agent that knows your codebase should retrieve current repository evidence, stable architecture decisions, commands, and your reviewed preferences—while the coding client still edits code and tests every change.

Developers describe returning to a project and rebuilding the mental model, or correcting an agent on Tuesday only to see the same mistake on Friday. Giant context files also go stale. The durable answer is a small hierarchy of current repository instructions, decisions, examples, and verification commands, each with an owner and review date.

The Agent should retrieve from an owned, reviewable source rather than turn a chat transcript into truth. Give every durable fact an identity, scope, source, freshness signal, and correction path. Keep the external system of record authoritative when a calendar, repository, catalog, or reference manager already owns the data.

Treat the repository as the primary source: tracked AGENTS files, architecture notes, schemas, test commands, and examples. The personal Agent can retrieve and explain that context across supported coding clients, but it should not pretend memory is newer than Git. Project rules must be updated with the code change that makes them true.

Use Knowledge for durable approved material, User Memory for small private preferences or context, and tools for live reads or explicit actions. A remembered preference cannot grant access, publish, buy, delete, or override server rules. Sensitive writes need the real user, narrow permissions, confirmation, history, and revocation.

Publish one owned Agent through its remote MCP endpoint, then connect it only where the chosen client, account, transport, and OAuth flow support MCP. ChatGPT, Claude, Cursor, Codex, or another client remains the interface; the Agent keeps the same identity, instructions, knowledge, tools, and access policy instead of being rebuilt inside every ecosystem.

Pilot with representative questions, stale or missing data, denied actions, and one recovery path. Measure successful jobs, corrections, safe refusals, time saved, and maintenance work. Expand only after the narrow version stays accurate; remove data or tools that do not earn their risk and upkeep.

Prostir can host the owned context-and-tool Agent; it is not automatically a local coding executor. Cursor, Codex, Claude Code, or another coding client performs edits under its own permissions. Require diffs, tests, exact expected versus actual output, and human review before merge or deployment.