Help & FAQ
How Sia Agent Forge works, and answers to the questions that come up most.
When does it make sense to build an app?+
Build an app when a repeatable task needs a shared interface, structured inputs, saved results or a workflow that several people can use. Examples include a document analysis tool, a reporting dashboard or a guided assessment. For a one-off question or draft, a conversation with an existing assistant may be enough. Start with one clear task, its users and what a useful result looks like.
Can I still use Langflow? Does Agent Forge replace it?+
You can continue building agents in Langflow. Agent Forge can build an application around an existing agent available through Agent Marketplace, or implement an agent in the app itself. These are different approaches; do not assume that every Langflow feature has an identical visual editor here.
How do I make my existing Langflow or SiaGPT agent available?+
Agents in the public SiaGPT space are added automatically. Check whether yours appears in Agent Marketplace. If it is missing or private, contact the Agent Forge team with its name and location to confirm access and the integration steps.
Should I use Agent Forge or Claude skills to create an agent?+
Both can contribute to the same project. Agent Forge provides a place to build, preview, host and share an application. You can also import application code created with Claude as a ZIP and adapt it with the agent. A skill on its own is not a deployable app. Choose based on the workflow and sharing you need; there is no blanket guarantee that one option costs less.
How should I compare Agent Forge with Microsoft Copilot Studio?+
Agent Forge lets you build a custom application with editable code, preview it, publish it and prepare it for deployment elsewhere. For a client comparison, assess the actual use case: required Microsoft 365 integrations, identity and permissions, hosting, customization, operating costs and maintenance. Ask the Agent Forge team for validated positioning material before making feature-parity or pricing claims.
Is the Sia design built in? Can I use a client design or provide an example?+
The agent has Sia design guidance, but you can request different colors, typography and layouts. Describe the desired style and provide a reference or mockup, including any client brand requirements. Review the proposed design before approving the build; you are not limited to the default appearance.
Does Agent Forge use the same models as SiaGPT?+
Distinguish the assistant that builds your app from the AI your app calls. The builder uses Claude models. For the app itself, you can choose the platform LLM key, an existing Agent Marketplace agent, your own provider key or Amazon Bedrock. Model availability depends on that choice and its configuration; do not assume the list is identical to SiaGPT.
Can it split large tasks between multiple agents, like Claude Cowork?+
The building assistant can delegate specialized development tasks to subagents. This does not automatically give the finished app a multi-agent workflow. If your app needs parallel analysis or large document processing, specify that requirement so the workflow, batching, retries and result checks can be designed and tested. Multiple agents alone do not guarantee exhaustive or accurate results.
Are Microsoft 365 connectors available?+
Do not assume a ready-to-use Microsoft 365 connection is enabled for your app. Describe the service and action you need, such as reading SharePoint documents or sending a Teams notification, and confirm the supported integration with the Agent Forge team. Access must be configured with the relevant organization; publishing an app does not automatically grant it access to a client's Microsoft 365 data or permissions.
Can I talk to the agent instead of typing?+
Yes. Press the microphone in the Agent chat's message box and speak in French or English: your words appear in the box as you talk. Press it again to stop, then correct the text if needed and send it. Esc throws the recording away. One recording lasts up to 3 minutes, and each person can dictate up to 60 minutes a day; past that, you can still type. The audio is transcribed by AWS Transcribe in Sia's own AWS account, in the EU (Frankfurt), and is not kept, and neither is the text until you send it. The agent still answers in text. The first time, your browser asks you to allow the microphone for the site.
Do I need to know how to code?+
No. Describing what you want to the agent is enough to get a working app, and plenty of apps here were built entirely that way. Studio is there if you do want to write code, not because you have to.
Will my app break if I keep asking for changes?+
Your published app stays exactly as it was when you published it, so your colleagues are never affected by work in progress. Every change is saved as a step along the way, so a bad one can be undone — just ask the agent to put it back.
My app was working and now the page will not load.+
If it is published, it has most likely gone to sleep after eight quiet hours; opening it wakes it, which takes a moment. If it is a preview, it stops after an idle hour and the Preview button starts it again.
Still failing: open the Preview tab to see what your app reported, or simply ask the agent to look for you.
Why does the first request after a while take so long?+
Because the app was asleep and is starting up. This affects only the first request; the rest are normal speed.
Can I use a database, or store files?+
Yes to both, and you do not have to set either up. Say what your app needs to store and the agent declares the database or file storage, which the platform then provisions.
When the data model changes, ask the agent to include a migration that preserves existing records and test it in Preview. Use Update to ship a new version while keeping production data. Unpublish deletes the published app's database and stored files; it is not a way to apply an update. Preview and the published app have separate runtime environments, so do not assume preview data will become production data.
Where do I put my API keys?+
Settings → Secrets & env, marked as secret. Never inside the app itself — anything written into it can be read by anyone who reviews it or makes a copy.
Do I have to build a login page?+
No, and you should not. Sign-in happens before a request reaches your app, so your app already knows who the visitor is. A login page of your own would be redundant at best, and a weaker second gate at worst.
Can people outside Sia use my app?+
Yes, once an admin approves it. Request guest access in Settings → Public access. Guests sign in by confirming their email with a one-time code, since they have no Sia account.
Someone outside Sia was asked for their email again on a second app.+
Confirming once now covers the other guest-visible apps they open for the rest of that session, rather than a fresh code per app. Signing out ends that and lets them come back as somebody else.
Can I rename my app?+
The display name, any time. The slug — the address — is frozen once published, because it is the app's hostname and the link people have already shared. Unpublish first if you truly need to change it.
Who can see the code I write?+
You and the builders you add in People — on the platform and on the app's GitLab project. Viewers and visitors use the running app without access to Studio or GitLab; browser-delivered code remains inspectable. Allowing forks lets others copy the published code into their own app.
Should I add a colleague as a builder, or let them fork?+
A builder works on your app with you: same code, same conversation, same address. A fork is a separate app they own, copied from your published version, that goes its own way and never comes back to yours. Building it together: builder. Reusing it for something else: fork.
A colleague pressed Backup to Git and I see nothing on GitLab.+
Every backup goes to the owner's branch, marketplace/<owner>, whoever presses the button, so look there. Backups made before this was fixed, in October 2026, went to a branch named after the builder; that work is also in the app's shared copy, so the next backup brings it to the owner's branch.
If a builder cannot open the GitLab project, their GitLab account was not found from their Sia address. They are added at their next backup once it can be.
Why does my new app show me a mockup instead of building?+
So you can check the agent understood you before it builds the whole app. Ask for changes under the mockup, then press Approve and build.
I get “page not found” on an app's Agent or Settings tab.+
Those tabs exist only for the app's owner and builders; for anyone else the page answers as if it did not exist. Ask the owner to add you as a builder in Settings → People.
Where is my app on GitLab, and can I open it?+
After the first backup, the app's page tells its owner whether they were given access to its GitLab project, and under which GitLab account. Builders are given the same access. A GitLab account is found from the Sia address, or failing that from the full name; if neither matches, nobody can be added until it does.
My app's AI calls suddenly fail with a 429 error.+
The app has reached its AI budget, or its per-minute limit (300 requests or 1,000,000 tokens a minute). Settings → AI access shows the spend against the budget: if the budget is used up, ask for more there and an admin reviews the request. If it is not, it was the per-minute limit, which lifts by itself — make your app wait and retry rather than fail.
What happens to an app I stop using?+
It sleeps after eight quiet hours, which changes nothing and is undone by opening it. After 30 days with no visitors it is archived, and that one does have a consequence: the app is kept and you can publish it again, but it starts fresh, without the data it had collected.
So if an app holds something you would miss, open it now and then, or keep your own copy of what matters.
Can I get my code out of the platform?+
Yes. Backup to Git saves a full copy to the app's GitLab project, which you and your builders can clone. The Deploy export is available to platform admins, who can prepare the files for a technical team to run the app on its own infrastructure.
Something is wrong and none of this covers it.+
Ask the agent on your app's Agent tab first — it can see your app, its settings and what it reported when it ran, which is more than a colleague could tell from a screenshot. If the platform itself is misbehaving rather than your app, tell the Agent Forge team.
Are there token or spending limits while building or previewing an app?+
Two separate limits apply. Building with the agent: each person has an allowance of 5 million AI tokens over a rolling 24 hours, shared by all their apps. It counts what the agent produces for you — code, replies, tool calls; your messages and attached documents do not weigh on it. A bar under the Agent chat shows how much you have used, and when the allowance is used up the message says from what time you can send again: it frees up as older messages turn a day old, not at midnight. See Settings → Build cost for measured development usage.
AI calls made by the app, including in Preview: the platform LLM key has a default budget of USD 20 per 30 days, shared by everyone using the app, and per-minute limits of 300 requests and 1,000,000 tokens. Past them, calls return HTTP 429; request a budget increase in Settings → AI access.
Is there a limit on how many apps I can create or publish?+
There is currently no per-user app count limit. Internal publication does not require the client-catalog review, but shared hosting capacity is finite. Choose the appropriate audience and provide an app description when publishing. External sharing requires an admin review; a restricted audience does not replace approval to use client data.
Are security best practices or an audit skill included?+
The development agent includes best-practice instructions. A cyber-audit skill is also available; ask the Agent Forge team how to access it if you cannot find it. These checks do not guarantee that generated code is secure. Keep passwords and API keys in Settings → Secrets & env, review access and dependencies, and use approved data before sharing an app.
Can a client see or reverse-engineer our app?+
Using an app does not grant access to its Studio or GitLab repository. However, browser-delivered code and visible behavior can be inspected; an interface cannot guarantee protection against copying. Keep sensitive logic and secrets on the server, grant builder access deliberately and leave forking disabled unless copying is intended. Ask the relevant Sia team to confirm the sharing terms for the engagement.
Can we deploy to a client's environment and customize the interface?+
The app's interface can be customized for the client. Platform admins can prepare an export of the application code and deployment files through Deploy, for a technical team to run in the target environment. Hosting, authentication, integrations and secrets must be configured there. This exports the application; it does not install the full Agent Forge platform or its builder interface. Discuss a full platform deployment separately with the Agent Forge team.
What is an MCP server, and what can I do with one here?+
An MCP server gives an AI agent tools from outside the platform: the current documentation of a library, a repository's wiki, a search in a public catalogue, a service your app integrates with. Here, MCP servers serve two purposes: the agent uses them while it builds your app (for example to read up-to-date documentation before writing code), and your app itself can call them once it runs, in Preview and once published.
Where do MCP servers come from?+
From two places, both in Settings → MCP servers. The platform catalog: servers the Agent Forge admins have added and checked, offered to every app. This app's own servers: ones you add yourself, which only this app can use — never another app, not even another of yours.
Which MCP servers can be added, and which cannot?+
Accepted: a server reachable over the internet at an https:// address, with no credential or with a token passed in an HTTP header (for example Authorization: Bearer …). Most hosted MCP servers work this way.
Not accepted:
- servers you install and run locally (
npx …,uvx …, a command or a package): nothing third-party runs on the platform itself; - addresses inside a private or company network, or plain
http://; - servers that require signing in with an account in a browser window ("Connect your account"), such as SharePoint or Microsoft 365 for now;
- a token written inside the URL: put it in a header when the server allows it, otherwise the server cannot be added.
Does the agent get the catalog's MCP servers automatically?+
It depends on each server, and the admins decide. A server marked automatic is available to every app's agent with nothing to do. Others are added by the app's builders in Settings → MCP servers. A server can also be reserved for Sia staff, in which case external accounts do not see it. A server that needs your own token is always added by you, since only you can provide it. In Settings, each catalog server shows whether the agent already has it.
How does my app — not just the agent — use an MCP server?+
The agent using a server while it builds does not give it to your app. The app gets a server once it is marked Used by the app: press Use in the app in Settings → MCP servers, or let the agent do it when your app's code needs one — it tells you when it does. Your app then receives the server's address and credentials as MCP_<NAME>_URL and MCP_<NAME>_HEADERS from its next Preview or publish. A server you added to this app yourself is given to the app already.
How do I give my own token to an MCP server?+
In Settings → MCP servers, press Use in the app on the catalog server (or Add a server for one of your own), and paste the value of the header the server asks for, usually Bearer followed by your token. It is stored as a secret of this app: it is never shown again — not to you, not to the agent, not on this site — and no other app can use it. Never paste a token in the Agent chat.
I changed or renewed an MCP token. When does it apply?+
The Test button uses the new value at once. Your running app picks it up at its next start: start the Preview again, or publish again (Update) for the published app — the one already running keeps the previous value, as for any secret.
Who can add or change an app's MCP servers?+
The app's owner and its builders, as for its secrets and variables. Viewers cannot. External accounts never see a server the admins reserved for Sia staff. Only the admins manage the platform catalog.