NeoForge treats agents like Unix processes. Single-file scripts, JSON over pipes, a kubectl-like CLI, and pluggable discovery — so you can bolt multi-agent workflows onto the codebase you already have.
pip install neoforge
Requires Python 3.11+ & uv
$ agentctl get agents NAME STATUS DISCOVERY RUNTIME hotspot-finder running file uv drift-detector idle dns-srv uv test-runner running file uv doc-writer idle etcd uv $ agentctl describe hotspot-finder agent: hotspot-finder file: agents/hotspot_finder.py input: {"repo": "string", "limit": "int"} deps: gitpython>=3.1, rich>=13 # PEP-723 inline $ agentctl run hotspot-finder \ --input '{"repo": ".", "limit": 10}' \ | agentctl run doc-writer hotspot-finder → doc-writer piped 847 tokens done docs/hotspots.md written
Inspired by the Unix philosophy: small tools that do one thing and pipe cleanly into the next.
Every agent is a standalone Python script with an inline PEP-723 header declaring its name, version, and dependencies. No separate requirements.txt, no package to publish. Drop the file in agents/ and it's discoverable.
Agents find each other through a tiered resolver: file registry for local dev, DNS SRV for service-mesh environments, etcd for production clusters. Switch tiers per context without changing agent code.
Agents communicate over JSON on stdin/stdout. Bulk payloads spill to temp files. No shared database, no schema migrations, no ORM. Compose workflows with the pipe operator, just like shell commands.
The agentctl CLI speaks kubectl verbs: get, describe, run, logs, exec, start, stop. If you've managed a Kubernetes cluster, the mental model is already in your head.
Code Atlas converts Claude Code sessions into a searchable knowledge graph. It runs entirely on NeoForge agents — hotspot finder, drift detector, session indexer — wired together with agentctl pipelines and YAML orchestration.
Get notified for releases, RFCs, and community calls. One email per milestone — no drip campaigns.
No spam. Unsubscribe anytime.