Build swarm
Ticket intake routes through planning, architecture, implementation, testing, review, and delivery with dedicated bots for each role.
Open Swarm: secure ops, integrations, scale
Roger Murphy
I build working agentic software: multiple bots on multiple harnesses, one-click integrations, a secure operations hub, user and tenant isolation, dynamic node scaling, and real applications across finance, shopping, rides, support, GitHub, and Dropbox intelligence.
One-click integrationsFinance, GitHub, Dropbox, Google, home, commerce
Secure Ops HubQueues, health, logs, costs, approvals
Tenant isolationUser, single-tenant, and multi-tenant boundaries
K8s + GardenerDynamic node scaling and remote nodes
Flagship project
oshal is Open Swarm: any harness, API, or LLM.
oshal is Roger's Open Swarm implementation: a beta multi-tenant, multi-user, multi-agent runtime where specialized bots can run different agent harnesses, API providers, or LLMs, coordinate over a Redis Streams mesh, and register new apps, tools, agents, and workflows while the swarm is running.
The point is practical: turn agentic AI from one clever prompt into a governed operating layer with user and tenant boundaries, tickets, roles, tools, connector-backed data access, cost attribution, review gates, one-click integrations, secure operations, deployment options, and applications people can actually use.
TypeScript
Node.js
Redis Streams
Postgres
ChromaDB RAG
Docker Compose
Kubernetes
Gardener
OIDC / Keycloak
Single-tenant isolation
Multi-tenant
Multi-user
Connector token broker
A2A remote agents
Dynamic node scaling
Local LLMs
Proof points
Not a chatbot wrapper. A working swarm runtime.
The current oshal materials show a broad system: build automation, incident RCA, bot generation, education workflows, connector-backed communication intelligence, finance and commerce intelligence, support and GitHub workflows, remote node control, secure operations, and model optimization.
26registry bots in the current beta snapshot
68persona definitions for role-specific execution
18dynamic tools, agents, and containers launched in a validation pass
~68%lower SAP HANA RCA cost in one persona-gated path versus two-bot review
RCA swarm
Incident workflows produce root-cause analysis, impact assessment, remediation steps, rollback planning, and scripts.
Bot Forge
Codex-packer interviews an operator, emits a persona plus swarm app manifest, and injects a focused bot into the running swarm.
Self-optimization loop
oshal can replay agentic cycles, compare prompts and models, and tune per-call latency and cost so applications improve against real platform workloads.
Per-bot configuration
Every bot can choose its harness, provider, model, and tools.
oshal does not force the swarm into one provider or one runtime. Each bot can be configured for a working harness, API provider, and model under that provider while still using the same bot-to-bot communication layer and standard framework tools.
Harness per botA tutor, planner, incident responder, or build worker can run through the harness that fits its job.
Provider and model per botProvider choice and model choice are runtime configuration, not hard-coded architecture decisions.
Shared swarm communicationBots keep the same A2A and workflow messaging even when their underlying harnesses or models differ.
Tool policy per botStandard framework tools remain available, with auth and execution modes configured on a bot-by-bot basis.
Integration and operations layer
Agentic apps can connect to real systems without losing control.
The updated oshal feature set expands the showcase beyond assistant chat: finance, shopping, rides, support, GitHub, Dropbox, and secure operations can all run through scoped connectors, review gates, and isolated deployment boundaries.
Agentic Finance
Plaid and payment connector patterns let finance bots read approved account, transaction, invoice, and cash-flow context, then produce analysis or proposed actions with human approval.
Shopping and ride workflows
Shopping research, purchase lists, pickup flows, and rideshare or Uber-style requests can be represented as governed tasks instead of loose prompts.
Support, GitHub, and Dropbox intelligence
Support bots can combine tickets, runbooks, Dropbox files, and GitHub issues or repositories to draft fixes, summarize context, and prepare handoffs.
One-click integration paths
Google Workspace, Dropbox, GitHub, GCP, Plaid, payments, SmartThings, Nest, commerce, and ride actions can sit behind a connector broker scoped per user.
Secure operations hub
Operators get task queues, mesh visibility, bot health, logs, cost tracking, RAG surfaces, review state, and approval gates in one operating surface.
User, single-tenant, and multi-tenant isolation
User-scoped stores, token boundaries, OIDC/Keycloak patterns, single-tenant deployments, and distributed multi-tenant scoping keep work separated.
Gardener and Kubernetes support
Docker Compose remains useful for local work, while Kubernetes and Gardener-backed paths support distributed nodes, remote workers, and dynamic node scaling.
Dynamic node scaling
Bot lanes can scale independently as work arrives, with heartbeats and registry state helping the controller know which nodes are available.
Connector-backed intelligence
The useful pattern is not a bot with generic internet access; it is a bot with the right approved data, the right tool policy, and the right review point.
Feature inventory
Other capabilities surfaced in the oshal codebase.
These are the pieces that make the project interesting to hiring teams: not just AI calls, but the surrounding engineering needed to integrate, isolate, run, govern, and extend agentic systems.
Per-bot harness and model config
Each bot can choose its harness, API provider, model, selector skills, provider auth, and tool modes while still joining one coordinated workflow.
Loadable swarm apps
YAML app bundles declare bots, routes, tools, UI ribbon entries, migrations, voices, themes, and workflows.
Dynamic bot and tool creation
Tools, agents, and bot nodes can be registered into a running swarm, launched as containers, heartbeated, scaled, and made visible in the registry.
A2A and remote nodes
Remote clients and daemon-style nodes can register, receive commands, bridge local tools, join the swarm over private transports, and scale lanes independently.
Secure operations hub
Operator surfaces include task explorer, queues, mesh dashboard, ops, health, Redis visibility, logs, RAG center, cost views, and approval state.
Unified oshal assistant
Jarvis provides a front door into the swarm, routing requests across specialist apps, command center signals, voice, text, and workflow actions.
One-click data and action integrations
Connector patterns cover Google Workspace, social publishing, SmartThings and Nest, Dropbox, GitHub, GCP, Plaid finance, payments, shopping, and ride workflows, with token brokering scoped per user.
Agentic finance intelligence
Finance bots can use approved banking, transaction, invoice, payment, and budget context to explain spend, watch signals, and prepare governed actions.
Shopping and Uber-style tasks
Commerce bots can support shopping research, purchase planning, delivery or pickup handoffs, and rideshare-style requests through approval-gated workflows.
Support bot and GitHub intelligence
Support agents can combine tickets, runbooks, Dropbox files, GitHub issues, repository context, and operator notes into triage and remediation drafts.
Connected home control
Smart Home connects oshal to device state, SmartThings scenes, schedules, timers, and natural-language home commands.
Cost, tokens, and tool policy
Cost attribution is recorded per bot and call, with tool access controlled per agent through auto, ask, and off modes.
Application self-improvement
oshal's optimizer can replay real agentic cycles, tweak prompts and model choices, then feed cheaper and faster per-call paths back into applications.
RAG and memory surfaces
ChromaDB collections, provenance-aware citations, uploaded class materials, and SAP or infra runbooks can ground bot output.
Scheduling and self-healing
Cron-backed scheduling, heartbeats, stuck-agent watchdogs, re-registration, health monitoring, and stale-channel cleanup are part of the runtime.
User and tenant isolation
Tenant-aware apps, OIDC, production Keycloak patterns, user-scoped stores, connector token handling, auth-gated routes, single-tenant boundaries, and distributed multi-tenant scoping keep execution accountable.
Email and social intelligence
Google Workspace digests, Gmail triage, calendar context, LinkedIn draft and publish flows, and social signals run through a communications bot.
Career memory and tailored resumes
Career Hunter builds a structured bank from files, notes, and spoken context, then pulls the right experience into job-specific resumes and cover letters.
Presentation generation
Presentation Studio creates real PowerPoint decks from templates, topics, or outlines, with AI guidance and Dropbox, Git, local, or download storage paths.
Little Monsters app
A six-bot learning app handles lecture recording, transcription, flashcards, tutoring, textbooks, study plans, writing help, and presentations.
Runs close to the work
oshal supports Windows, Docker Compose, Kubernetes, Gardener-managed clusters, local models through Ollama or LM Studio, remote command execution, and dynamic node scaling.
How it works
A ticket becomes coordinated work.
oshal's architecture is built around clear ownership: the controller routes and observes; bot nodes execute; personas define role behavior; the mesh carries work between agents.
1. Intake
A ticket, alert, upload, or operator request enters the system. The workflow selects the right app, ticket type, and phase sequence.
2. Route
The controller dispatches envelopes over the mesh. Agents consume their own streams, so lanes can scale independently across local, Kubernetes, Gardener, or remote nodes.
3. Execute
Bot nodes run the LLM work. Each bot can use its harness, tools, persona, model, and connector context to read approved data or perform approved actions.
4. Govern
Results, tokens, cost, handovers, and review state are captured. The secure operations hub keeps human gates, logs, health, and approvals available where they matter.
Live application proof
Little Monsters turns the swarm into a student-facing learning product.
Little Monsters is the education app riding on oshal: a multi-bot student workspace for lectures, study loops, tutoring, flashcards, quizzes, class management, and presentation generation.
Student dashboardProgress, streaks, lectures, class actions, and study status live in one cockpit surface.
Lecture to study loopRecordings can be captured, transcribed, replayed, and turned into review material.
Practice surfacesFlashcards, quiz practice, tutor chat, and class bank flows make the agents useful to students instead of only operators.
Shared platform capabilityThe same Presentation Studio deck generator is available inside the learning app and as its own oshal app.
Live application proof
Career Hunter turns your experience into a living career bank.
Career Hunter is not just a job board. It builds a database of your career history from files, notes, and conversations, then uses that experience bank to score roles, surface relevant stories, and write tailored resumes and cover letters for the specific job.
Career memoryUpload files, speak into the system, and add notes so your accomplishments, project stories, skills, and context become reusable structured data.
Job-specific resumesFor each role, it pulls the most relevant experience from your bank and drafts a resume and cover letter matched to the posting.
Experience extractionWhen a job exposes a gap, the app asks targeted questions and turns your answers into future resume material.
Pipeline intelligenceScored jobs, recruiters, approvals, fit thresholds, salary signal, company breakdown, and funnel analytics stay tied to your profile.
Live application proof
Presentation Studio turns an AI conversation into a real PowerPoint deck.
The deck generator runs as a standalone oshal app and inside Little Monsters. You can start from a template or topic, work with the deck-builder agent to shape the outline, then generate a real .pptx for storage or download.
AI-guided deck craftThe agent helps move from a rough idea to title, outline, slide sections, and bullets before generation.
Template or topicPitch, strategy review, project update, teaching, blank, and AI-drafted deck starts are all supported.
Real outputThe workflow generates actual PowerPoint .pptx files rather than a static preview or mock slide outline.
Storage choicesDecks can flow through Dropbox, Git-backed storage, local files, oshal storage, or direct download.
Applications
Working demos and signed-in surfaces.
The portfolio is strongest when people can open the work. Some apps are public prototypes; oshal cockpit surfaces are intentionally auth-gated, including the full cockpit with all tools loaded.
Full oshal Cockpit
All tools
The complete signed-in cockpit surface with the assistant, app routing, tools, task views, queues, mesh visibility, logs, RAG, and operations controls loaded together.
Open full cockpit ->Career Intelligence
oshal app
A career-history bank that accepts files, notes, and voice context, then scores jobs and drafts tailored resumes from the most relevant experience.
Open cockpit ->Little Monsters
oshal app
A learning workflow app for classes, recorded lectures, flashcards, tutor chat, OCR, retrieval, and presentation generation.
Open cockpit ->Presentation Studio
oshal app
AI-assisted deck creation with templates, topic-to-outline drafting, real .pptx generation, deck storage, and direct download.
Open cockpit ->Smart Home Control
oshal app
Connected-home control for SmartThings devices, scenes, schedules, timers, and natural-language commands like make it cozy.
Open cockpit ->oshal assistant
oshal app
The Jarvis cockpit surface: voice and text entry into the swarm, command center signals, app routing, and assistant-led workflow starts.
Open cockpit ->oshal apps dashboard
oshal
The manifest-driven app dashboard for loading swarm apps, importing YAML, focusing a cockpit surface, and toggling active application bundles.
Open gallery ->Intelligent Instructions
Sign-in
Turns goals and instructions into structured phases, tasks, dependencies, and agent-readable execution plans.
Open demo ->oshal optimization engine
oshal engine
A behind-the-scenes view of how oshal tunes applications by replaying agentic cycles, testing prompts and models, and reducing per-call time and cost.
Open engine ->Why this matters
This is a portfolio of engineering judgment.
oshal shows platform thinking, but the surrounding apps show product delivery: user interfaces, authentication, databases, queues, connectors, LLM economics, deployment, observability, and operational safety.
1
Systems builderConnects agents, queues, data, UI, tools, and infrastructure into working products.
2
AI pragmatistOptimizes for cost, correctness, reviewability, and model flexibility instead of vendor lock-in.
3
Full-stack operatorMoves between TypeScript, Python, Docker, Kubernetes, auth, databases, and frontend UX.
4
Product-minded engineerTurns platform capability into demos people can test, critique, and reuse.
Looking for senior AI platform, full-stack, or automation roles.
Agentic Federal is the proof bench. oshal is the flagship. Roger is available to talk about roles where practical AI systems, workflow automation, platform engineering, and product delivery all meet.