—— OFFICE OF THE CTO · APR 2026

DocsClaw
A Lightweight Agentic Runtime
for OpenShift

ConfigMap-driven AI agents that fit in 5 MiB.
Same binary, different config = different agent.

A2A Protocol ConfigMap-Driven Zero-Trust Ready Go 1.25

Pavel Anni · Office of the CTO · Red Hat

DocsClaw mascot — a lobster reviewing documents

—— THE PROBLEM

Existing Runtimes Are Too Heavy for Simple Agent Tasks

1,577

MiB per pod · OpenClaw

vs

5

MiB per pod · DocsClaw

A 300x reduction in memory footprint. Same agentic capabilities: tool use, LLM integration, A2A protocol.

Source: oc adm top pod on OpenShift demo cluster, April 2026

—— THE JOURNEY

We Didn't Need a Framework —
We Needed a Thin Runtime

Started with simple Go agents (summarizer, reviewer) for our Zero Trust demo
Wanted them universal — same binary, any personality via ConfigMap
Evaluated NanoClaw, ZeroClaw, PicoClaw — good, but include extras we didn't need (gateway, channels, etc.)
Realized our existing A2A-compliant Go code was 80% there
Analyzed PicoClaw architecture + Claude Code's harness pattern
Added the missing 20%: tool registry, agentic loop, SKILL.md loading

Go Binary A2A Native Kagenti Compatible Built in a Day

—— ARCHITECTURE

Same Binary, Different ConfigMap = Different Agent

CONFIGMAP

system-prompt.txt

agent-config.yaml

skills/

↓

DocsClaw Binary

↓

A2A Endpoint

→

Agentic Loop

→

LLM Provider

↓

BUILT-IN TOOLS

exec

web_fetch

read_file

write_file

load_skill

—— IN ACTION

Tested on OpenShift, Works End-to-End

Standalone Mode

Binary + API key + ConfigMap
No SPIRE, no OPA, no sidecars
Plain HTTP or TLS termination
Deploy in any namespace
5 MiB memory footprint

Zero Trust Mode

SPIFFE workload identity
OPA policy evaluation
Permission intersection:
User ∩ Agent = Effective
Delegation context propagation
Same binary, same protocol

Deploy: oc apply -f deploy/standalone-agent.yaml

—— WHAT'S NEXT

DocsClaw Is Just Getting Started

Generic fetch_resource Tool

Configurable backend URL — fetch from any service, not just documents

MCP Tool Integration

Connect to external tool servers via Model Context Protocol

Multi-Agent Orchestration

Agent-to-agent delegation via A2A protocol

OCI-Based Skill Distribution

Pull skills from a registry like container images

Memory & Context Persistence

Evaluating Letta, Mem0, Graphiti for OpenShift

github.com/redhat-et/docsclaw — Contributions welcome

Thank You