Your idea. A working agent flow. A native macOS agent studio powered by your locally installed Codex runtime.
I built AgentFoundry out of curiosity. I wanted to visualise agent flows: what each agent is responsible for, how the steps connect, and what happens when the work actually runs. That led to a second question: could I explain an outcome in ordinary language and let Codex create the setup for me?
AgentFoundry is my exploration of that idea. This is an evolving MVP, shared to show the experience and what is possible.
Describe what you want to accomplish. Choose Create professional workflow proposal, and Codex turns the rough idea into a structured, readable proposal covering the approach, agent responsibilities, outputs and checks.
The proposal remains editable. Change the text directly or ask Codex to refine it. When it matches your intent, choose Implement. AgentFoundry creates the agents, workflow steps and connections automatically. You do not need to configure every agent manually or wire the nodes together. The canvas is there to help you understand the flow and refine it when useful.
Implementation saves the setup. Review the run inputs and start execution separately; it does not silently run the task or activate a schedule.
AgentFoundry combines guided workflow creation with a visual canvas, reusable agents, a catalogue of 20 specialist roles, runtime capability discovery, MCP integrations and real execution history. Agent-level controls are available when you want to adjust a model, instructions or tool access, while the guided path handles the initial setup.
Runs distinguish technical execution from declared outcome checks. Where configured, file checks independently open and inspect generated outputs. Service evidence reflects what the real tool or provider confirms. Saved definitions, artifacts, version history, recoverable library deletion and optional verified checkpoints support an iterative workflow.
Recurring schedules are available while AgentFoundry is open and the Mac is awake. Schedule activation is a separate step.
A daily news briefing. “Read the five leading stories on HotNews.ro, analyse what matters, include the sources, and create a PDF on my desktop.” Codex proposes the research and document steps; AgentFoundry creates the flow and runs it with the configured output checks. Enable a recurring schedule separately when ready.
A clearer project update. “Review these project notes, identify decisions, open questions and risks, and turn them into a concise progress report.” Start with supplied documents, review the proposed analysis and reporting approach, then implement and reuse it as new notes arrive.
These examples describe the intended pattern. Actual access, outputs and verification depend on the task's sources, tools and declared checks.
AgentFoundry is designed to use each user's own local Codex installation and authentication. It does not distribute the creator's credentials or require a shared account.
With an AgentFoundry build installed, open Settings → Codex → Set up local Codex. Discover or select the executable, inspect the runtime and account, and use the supported browser sign-in flow when authentication is required. Choose an available model, run the genuine connection test, and select Use this setup after the completed response is verified. Existing local authentication is reused where available. Custom-provider and managed-account configuration stays with Codex's own supported tools.
The Codex process and orchestration run on your Mac. Model inference uses the configured service and is not necessarily offline; workflow material may be sent to that service and selected integrations. Your provider's account limits and charges apply.
Availability: this repository is a product showcase. It contains the landing page, explanatory material and six reviewed screenshots. No installer, DMG, application binary or application source code is included. The setup description explains the app's existing connection flow for a provided build; it is not a download offer.
The MVP focuses on turning intent into a reviewable agent setup and making execution observable. Background execution with the app closed, authenticated external event triggers, broader provider acceptance coverage, large-workflow accessibility/performance work and portable public distribution are future directions.
Connected services still require their own authentication and permissions. A subscription or logged-in browser needs an accessible authenticated route at execution time. A possible paywall is a runtime condition, not automatically a reason to reject a valid workflow configuration.
Outcome checks establish only what they inspect. A readable PDF does not certify the quality of its analysis, and provider acceptance does not prove message delivery or reading. OCR is bounded and benefits from review. The app records real failures and incomplete results rather than simulating success.
All six images come from the supplied app captures. Two publication copies have local directory paths replaced with “Path redacted for privacy” using image editing; the originals are not distributed. Nonessential PNG metadata was removed from publication copies. The captures show the interface at different moments, including an unsaved agent and pre-execution states; they are not a claim that every pictured task has run.
AgentFoundry is an independent curiosity project by swihpi. It is powered by the user's Codex setup and is not an OpenAI product.





