FAQ
Straight answers.
What PlatformSmith does, what it doesn’t yet, and where your data goes. The rest is easier to show than to tell.
Last updated
The idea
What is PlatformSmith?
An agentic software development platform. It is the control plane for a fleet of AI coding agents: many agents, many services, one record, and approval gates wherever you want them.
We already use Claude Code and Codex. What does PlatformSmith add?
Claude Code and Codex are the agents. PlatformSmith is what runs them together. It splits one piece of work across services, routes and records every message between agents, and keeps a run waiting for your decision for as long as it takes. A coding assistant helps one developer in one repository. This runs the whole fleet.
Which coding agents does PlatformSmith support?
Claude Code and Codex. You can mix them in the same run, with Claude Code on one service and Codex on another.
Can one task span multiple repositories and services?
Yes, that is what it was built for. A run can fan out across every repository in a workspace, one agent per service, and barriers hold the group until the contract between services is agreed.
How a run works →Where does work come from?
Wherever your team already asks: Slack, GitHub, Jira, Linear, a schedule, the API, or an MCP client such as Claude, Cursor or ChatGPT. Every request is checked against your roles first, and your tracker keeps priority, assignee and sprint.
Control
What can an agent do without approval?
What you let it. Approval gates are explicit steps in a run; a run parks at one for up to seven days and resumes where it stopped. A run the system starts by itself, from a schedule or an alert, can’t change anything. Only a request from a real person carries that authority.
Can we pause, steer or cancel an agent mid-run?
Yes. Send input to a live session, stop a session, or end a whole run. The work sits on a git branch, so a developer can pick it up from where it stopped.
Are the guardrails enforced, or are they just instructions to the agent?
The important ones are enforced by the platform, not by the prompt. Which work item a run belongs to is decided by the control plane, never claimed by the agent; the tools an agent may call are granted per session, and a reviewing agent gets read-only ones; access is deny-by-default. No guardrail can yet stop an agent from being misled by something it reads. See prompt injection below.
Can we see how the agents arrived at a change?
Yes. Agents never talk to each other directly. Every message goes through the platform and can be replayed, and each output is an artifact attributed to the agent that made it and the definition it worked to. You approve artifacts, not a transcript.
Can each team work its own way?
Yes. Each team writes down how it works (prompts, subagents, skills, hooks and connectors), layered from company down to project. Access is role-based down to the workspace.
Does it work in our real environment or a clean-room sandbox?
The real one. The runtime goes inside your own container image and your service runs beneath it, so agents test against the real thing. A fix is rebuilt and swapped in underneath the running service. The pod never restarts, and the build runs before anything stops, so a broken build changes nothing.
Security and data
Where does PlatformSmith run?
The runtimes, where your code executes, run on your infrastructure. The control plane is ours, in AWS Ireland (eu-west-1). Your side dials out over one outbound connection; we never dial in.
Shared responsibility model →Does our source code leave our environment?
Agent calls go straight from your runtime to your model provider. That traffic never passes through us. We do keep a full session transcript in the EU so you can watch, resume and review, and it contains code. It is not redacted yet.
Privacy policy, section 2 →How are credentials handled?
You bring them. They are encrypted at rest, bound to the person who connected them, kept out of our logs, and delivered to your runtime when a session starts. Scope them narrowly, because a credential is the limit of what an agent can reach.
Privacy policy, section 2A →Is our code used to train models?
No. Not ours, and not anyone else’s. Your model provider’s retention and training terms are the ones on your own account with them, so you can read them yourself.
Who at PlatformSmith can see our sessions?
Our staff can see metadata (what happened and when) as part of normal support. Reading content needs a stated reason and your consent for that incident. And your session content never goes into our own AI tools.
Staff access policy →What about prompt injection?
Unsolved, here and everywhere. The best control today is narrow credential scope and reviewing work before it matters. We say so on a public page.
Known limitations →Can we delete our data?
Yes. Ask and we delete it, then confirm. A disconnected credential leaves live systems immediately; encrypted backups age out as they rotate.
Buying
Can we use our existing model subscriptions?
Yes. Bring an API key, or a setup token you generate with your provider’s own CLI; your provider bills you directly and we never resell access. For team-wide or unattended runs, use an API key.
What does PlatformSmith cost?
Nothing during the beta. You pay your model provider and run your own compute.
Can we see it on our own stack, not a staged demo?
That is the demo: thirty minutes, a real multi-service change, nothing pre-recorded.
What it doesn’t do yet
Does PlatformSmith have SOC 2, ISO 27001 or SSO?
Not yet. No SOC 2 audit, no ISO 27001 certification, no external penetration test, and no SSO/SAML or multi-factor authentication. Sign-in is email and password, or Google or GitHub.
Known limitations →Is there an SLA or high availability?
Not yet. It is a free beta in a single region with no redundancy and no service level agreement.
Can we set hard spending limits per task?
Not in the platform yet. Model spend sits on your own provider account, where that provider’s limits apply, and we don’t cap runtime CPU, memory or disk, so size the host accordingly.
Do you offer IP indemnity for AI-generated code?
No. Whatever cover you have comes from your agreement with your model provider.
Terms, section 4A →Can PlatformSmith restrict what a runtime reaches on our network?
Not today. That control is yours. What a runtime can reach is decided by your network, so segment it on your side.
The rest, we’d rather show you.
Thirty minutes, a real multi-service change, nothing pre-recorded.