Project knowledge first
Topchester builds and syncs a repo-local topchester-kb/ so the agent starts from a
project map instead of rediscovering everything from raw files.
Terminal-native agent for real codebases
Topchester builds a local map of your repo first, then uses it while planning, editing, checking work, and keeping project knowledge current.
Quickstart
Start with the npm package, add a small model config, build the project knowledge base, then open the terminal agent.
npm install -g topchester-ai
export OPENROUTER_API_KEY=...
topchester kb init
topchester kb sync
topchester
Config
Create topchester.jsonc in the project
Topchester should work on. Keep API keys in environment variables, not committed files.
{
"$schema": "https://topchester.com/schemas/config.v1.json",
"models": {
"default": "openrouter/google/gemini-3.1-flash-lite"
}
}
{
"$schema": "https://topchester.com/schemas/config.v1.json",
"models": {
"default": "openrouter/anthropic/claude-sonnet-4.5",
"kb.summarize": "openrouter/google/gemini-3.1-flash-lite"
}
}
Topchester builds and syncs a repo-local topchester-kb/ so the agent starts from a
project map instead of rediscovering everything from raw files.
It shows when project knowledge is stale and lets you sync changed files back into the KB.
Chat, inspect tool output, review diffs, resume sessions, and keep working from the terminal without switching into a heavyweight IDE workflow.
Features
Topchester combines a repo knowledge base with the practical tools expected from a coding agent: file edits, tests, Git, sessions, and safe command execution.
Sync a local project map before work starts, then keep it current as files change.
Use kb status, kb sync, and kb sync --full to keep knowledge fresh.
Chat, review tool output, follow progress, and resume work without leaving the terminal.
Read, search, create, and edit workspace files with hash-aware safeguards around writes.
Run test, lint, typecheck, build, format-check, and smoke commands through a strict validator path.
Inspect status, diffs, and logs, then stage or commit only explicitly acknowledged files.
Keep a visible session plan for multi-step changes without burying the thread in checklist prose.
Delegate focused exploration, persist session logs, and resume the latest project-local conversation.
Use one model for everything, or route chat and KB summaries to different configured models.
Store project knowledge, caches, logs, and sessions in predictable repo-local folders.
Workflow