Saguaro Agent

An open-source workflow, memory, and knowledge system for AI coding agents across Claude Code, OpenAI Codex, and Gemini CLI.

Published by 11 Mile Labs and presented here as public evidence of AB Whittington's agent-infrastructure work: workflow orchestration, durable memory, knowledge retrieval, cross-harness plugin packaging, and source you can inspect.

Public source liveGitHub repository

Why it exists

AI coding agents should not start from scratch every session.

Saguaro turns lost agent context into a product surface: durable memory for concise lessons, durable knowledge for document-scale retrieval, and repeatable workflows for phased engineering work.

The public repository frames the problem clearly: install once at user scope, initialize any existing project, and let agents retrieve prior decisions, search project knowledge, run structured workflows, and store lessons for the next session.

Design decisions

Three MCP servers, one public contract, no provider lock-in.

The public surface is intentionally small and stable: `saguaro-workflow`, `saguaro-memory`, and `saguaro-knowledge`, exposed through readable `workflow_*`, `memory_*`, and `knowledge_*` tool families.

The project is provider-neutral. It expects OpenAI-compatible embeddings and chat endpoints, while leaving hosted services, local model servers, and vector storage choices outside the core product contract.

The installer publishes generated artifacts for Claude Code, OpenAI Codex, and Gemini CLI from one source tree, including bundled MCP server builds so harness plugin caches do not need their own dependency install step.

Scope boundaries

Public, inspectable, and deliberately not private operating glue.

Project config lives at `.saguaro/config.yaml`; secrets stay in environment variables, and docs explicitly avoid reading shell profiles, user-home env files, or private machine-local setup.

Memory and knowledge are separate on purpose: memory stores short durable lessons, while knowledge stores document-scale material with ranked retrieval and optional synthesis.

The public repo does not claim adoption numbers, private workflow outcomes, or client-specific systems. Its proof is the shipped engine, CLI, docs, workflows, install artifacts, and source boundaries.

What is inside

Public surface.

These entries summarize the inspectable product surface: MCP servers, CLI/install paths, workflow contracts, and bundled skills. Links point to the public repository.

MCP servers

The core tool surface agents call from supported coding harnesses.

saguaro-workflowWorkflow orchestration, run state, dispatch logging, validation, phase artifacts, and lifecycle.
saguaro-memoryDurable run, project, and global lessons for future sessions.
saguaro-knowledgeDocument-scale knowledge ingestion, semantic search, retrieval, and synthesis.

Install and config

User-scope install, project-local configuration, provider-neutral runtime wiring, and safety boundaries.

saguaro CLI`saguaro init`, `saguaro doctor`, and `saguaro smoke` for project setup and verification.
Plugin installationGenerated Claude Code, Codex, and Gemini CLI install artifacts from one source tree.
.saguaro/config.yamlProject-local config, environment-variable mapping, storage backend selection, and redaction rules.

Workflows and skills

Portable workflow contracts and thin skills that teach agents how to use the system.

engineering-liteBundled engineering workflow with intake, plan, review gate, implementation, and review phases.
Workflow authoringGuidance for turning repeatable team habits into validated workflow YAML.
using-saguaroSkill guidance for choosing workflow, memory, and knowledge calls during agent work.

Compatibility

Harnesses targeted by the repository.

Harness support and runtime boundaries as stated by the public repository.
Item	Claude Code	OpenAI Codex	Gemini CLI
Plugin artifacts	Yes	Yes	Yes
Public tool families	workflow/memory/knowledge	workflow/memory/knowledge	workflow/memory/knowledge
Project config	.saguaro/config.yaml	.saguaro/config.yaml	.saguaro/config.yaml

Evidence trail

Source links.

GitHub repositorySource, history, public README, and repository context.
READMEProduct promise, public MCP surface, tool families, and documentation map.
Product and developer experienceOnboarding promise, primary users, core journey, DX principles, and success criteria.
Memory vs. knowledgeBoundary between short durable lessons and document-scale retrieval.
Workflow authoringWorkflow model, phase contracts, approval gates, and validation guidance.
AGENTS.mdPublic repository scope, included and excluded v1 behavior, and provenance rules.