Atomic ClaudeCode graphs, wikis, and research-backed agentic coding.

Deterministic tools for nondeterministic workflows. An opinionated Claude Code config that compiles the knowledge, indexes the graph, and runs the loop. Local and free.

Get Started

How it works

GitHub

~/atomic-claude



A code wiki your agent compiles

Claude compiles a markdown wiki of your repo: framework, build and test commands, a map of what lives where. It reads that before your code, and ship commands keep it current.



A knowledge realm across repos

A folder of repos mapped as one knowledge base: per-repo summaries, the concerns they share, and capture buckets that fold in loose notes, research, and tickets.



A code graph that speaks SQL

One command parses 31 languages into a symbol graph: definitions, callers, blast radius, no compiler needed. SQL is first-class, including Snowflake, dbt, and T-SQL lineage down to the column.



See what Claude sees

`atomic serve` opens the same maps Claude reads as a browsable site on localhost. Wiki pages, a whole-system graph, federated code search. Read-only, nothing leaves your machine.



Autopilot, task to PR, hands-off

Hand it a description or an issue number. It plans, implements test-first, checks blast radius against the graph, and reviews its own diff in a fresh context. You choose how to merge.



Channels for your agents

`atomic bus` gives concurrent Claude sessions named rooms. Every message carries an addressee list, so a room of agents notes the news instead of answering it forever.

See it run

The full tour from atomic serve on localhost: wiki docs, command-K search, SQL schema, two agents on the bus, plans, the external-link registry, and the system and code graphs.

Run your first grounded loop

The next step is your own repo. Turn on the output style, index your code, and hand the loop a real task, from install to a first merged change.

Run your first task →