Try it live

Two first-party host applications built on Affiant run in public: Meridian, an aircraft-maintenance desk, and HR Portal, a small HR department. Each opens its AI agent in a resizable panel on the right, from a floating button in the corner; the rail down the side is the app’s own navigation. Everything either agent proposes to write is filed and arrives as an Evidence Card, and nothing reaches either app’s own records until that proposal has been approved — by a reviewer on the card, or by a standing order a person wrote in advance. That moment is what the demos exist to show — and because the two apps sit on different backends and different docket stores, together they show that the moment depends on neither.

Both apps say so in their header bar: they are demo apps built to show Affiant at work, not products.

Before you start

• Sign-in. Each demo opens on a sign-in page, and the one-time code it asks for arrives by email. Enter any email address, paste the code it sends you, and you are in for 24 hours. There is no password and no account to create.
• Five messages a day, per person, per demo. The sixth is refused until 00:00 UTC. Each demo also has daily ceilings of its own across everyone using it: once one is reached, a further message — or a first sign-in — is refused until the same reset. Spend yours on the prompts below.
• Shared, synthetic data, and none of it keeps. Every visitor works on the same seeded data. Each app runs in a container that sleeps after fifteen idle minutes and is destroyed outright every night at 00:00 UTC; the disk is ephemeral either way, so the next request after one of those starts a fresh, reseeded app — chat history and any cards you left pending go with each reset. A pending card can also lapse before either of those on the app’s own docket deadline: thirty minutes in Meridian, fifteen in HR Portal. Deciding on a lapsed card is refused rather than applied. Treat it all as a shared whiteboard: don’t type anything real into it.
• It may need a moment. After a quiet spell the first visit takes a few seconds to wake — that spell is the same fifteen minutes.
• Seats are finite. At most fifty people can hold a seat at a time, across both demos together. A seat is claimed by each person who signs in and released only after 31 idle days, never by signing out. A first sign-in can be refused for either of two reasons: the fifty seats are taken, or that demo’s own daily ceiling on new people has been reached for the day.
• Say what you think. Every reply has a thumbs-up and thumbs-down; the comments reach the maintainers.

Meridian

Meridian is an aircraft-maintenance application — a fleet of ten aircraft, their work orders, and a parts inventory — at meridian.affiant.dev. Ask its agent to raise a work order and the proposal comes back as a card.

Captured on 2026-08-26, against the fleet Meridian seeded then. The ten aircraft were given fictional registrations later that day, so the registration in the shot is not one the demo has now; the table below names one it does seed, WZ-BRN.

What to try

Say	What happens	What it shows
Show me all AOG aircraft	The agent calls a read tool and answers from the fleet.	A read tool returning a dual-audience result.
What work orders are overdue?	Another read. Neither read tool filters on a due date: the search marks an overdue row in its result, and a statistics tool reports the count.	Same contract, a second tool.
Create an AOG work order for the left engine on WZ-BRN	The agent drafts the work order and an Evidence Card appears. Nothing is written yet.	The Docket and Evidence Cards, the Review Gate, and Rule 3 — write tools never write.
(on the card) type a new value into a field, then Approve	The write executes with your amendment on the record.	A reviewer’s correction is appended to the affidavit, and the tag it displaced stays behind it on the record.
Guide me through creating a work order	The app navigates to the form and walks you through it field by field.	Rule 6: the agent guides the UI through a registry, never by inspecting the DOM.
Is part PN-CFM56-OIL-001 available?	A stock check.	A read tool against inventory.

HR Portal

HR Portal is a small HR department — an employee directory, leave requests, expense reports, and an approvals inbox — at hrportal.affiant.dev. Its agent works on one employee at a time: find them first, then ask for a change on their behalf.

What to try

Say	What happens	What it shows
Find employee Priya Nair	The agent searches the directory and remembers who “her” means for the rest of the conversation.	A read tool. For leave and profile changes the host goes further: a write with no search before it is sent back to search first, so the agent does not guess an identifier.
Submit an expense report of 120 dollars for a client lunch for Priya Nair	The agent drafts the report and an Evidence Card appears. Nothing is written yet.	The same Docket and Review Gate on a different backend.
(on the card) type a new value into a field, then Approve	The write executes with your amendment on the record, tagged UserStated at confidence 1.	The reviewer’s tag goes in force unconditionally — a correction is not a confidence contest it can lose to the model.
Request 3 days of annual leave for Priya Nair starting next Monday	Another Evidence Card: this leave request wants a reviewer’s confirmation.	A policy the host built on Affiant.Policies, deciding what the write needs before the entry is even filed.
Request 1 day of sick leave for Priya Nair tomorrow	Nobody is asked to confirm: a standing order approves it on the spot and the leave is written. A card still arrives — the framework marks it as needing no confirmation, and the demo’s card does not show that, so it looks like any other.	A policy that pre-approves one class of write — a day of sick leave — so the routine case skips the queue. The entry is still filed, attested to the policy that fired.
Update Priya Nair's email to [email protected]	An Evidence Card for a one-field change.	The same gate on the smallest write.

What you are looking at

The card is an Evidence Card: the rendering of an Affidavit, one row per field the agent proposes to write. Each row carries a badge naming where its value came from. Both apps spell the provenance ladder out in the same plain words:

Badge	ProvenanceSource	Meaning
You said this	UserStated	The user said it, in so many words.
Looked up	External	Read from a system of record.
Calculated	Computed	Derived deterministically from trusted inputs.
From your conversation	Conversation	Drawn from earlier in the dialogue.
AI suggested	Inferred	The model’s best guess, sworn as exactly that.
System default	Default	A fallback value no one supplied.
Not derived — please verify	Empty	Provenance unknown, marked rather than left out. It counts as 0, so the write’s aggregate confidence is 0.

Approve and Reject are the Review Gate from the reviewer’s side. A required field still Not derived is shown as exactly that and drags the aggregate confidence to 0, and both demo cards go further than the framework does: Approve is disabled while a required field is blank — “Required fields are missing values” — until you type a value into it. The framework flags the hole and leaves that decision to the host; what it blocks on its own is a standing order over an empty mandatory field — that verdict degrades to a reviewer’s confirmation, so a person is asked instead. Set a new value before approving and it is recorded UserStated at confidence 1, bound to the decision you made it on, and on every field you change the reviewer’s tag goes in force while the model’s stays behind it in the chain — the reviewer’s word outranks the model’s. Clearing a field is something the framework has semantics for and these cards do not reach: an emptied input is sent as an empty string, which is a value, so a clear never leaves the demo.

Meridian’s guided walkthrough is Rule 6 in the open: the agent knows which form fields exist because the UI registered them, not because it read the page.

What they run on

Both apps are .NET 10 and React, talking over SignalR, with Gemini 2.5 Flash as the model, hosted on Cloudflare. Their tools are ordinary C# classes: on both of the backends they use, AffiantToolCatalog.FromType<T>() reflects over a type’s public instance methods — skipping generic method definitions, members declared on object, and property and event accessors, and refusing an overloaded name outright — and asks for no marker attribute to make one a read tool. A write tool is the exception: [AffiantWriteTool] is what declares it write-capable, and its three arguments are the operation the descriptor carries ("WriteCreate", "WriteUpdate"), the entity type, and the inference strategy. Without it the descriptor is built as a read query — the pre-tool inference is skipped as not_a_write_tool, and nothing holds the tool to returning a proposal. Every write tool behind the cards above carries it. The quickstart builds the same kind of tool in the Semantic Kernel shape instead, where a [KernelFunction] attribute marks each one. Where the two apps differ is the point:

	Meridian	HR Portal
Backend	Microsoft Agent Framework, through Affiant.AgentFramework	IChatClient directly, through Affiant.Extensions.AI
Docket store	Affiant.Docket, in memory	Affiant.Docket persisted by Affiant.EntityFramework (SQLite)
Packages in play	Affiant.Abstractions, Affiant.Core, Affiant.AgentFramework, Affiant.SemanticKernel, Affiant.EntityFramework, Affiant.Docket, Affiant.Policies, Affiant.Transport.SignalR	Affiant.Abstractions, Affiant.Core, Affiant.Extensions.AI, Affiant.EntityFramework, Affiant.Docket, Affiant.Policies, Affiant.Transport.SignalR

Meridian’s list is longer than its wiring, and neither extra package is there for the reason its name suggests: it takes Affiant.EntityFramework for the in-memory chat-session store that ships in it, not for a docket, and Affiant.SemanticKernel for the provider-configuration type its chat clients are built from, not for an interception seam. Both apps keep their own domain data in SQLite; the row above is about the docket.

Two backends, two stores, one Evidence Card — see Interception Backends.

What the demos do not show

• The third backend. Affiant also runs on Semantic Kernel — see Interception Backends.
• Nothing about hosted tools. Both apps’ tools are invoked locally, which is exactly the seam Affiant hooks. Tools that execute on a remote runtime are outside it, and the framework does not claim otherwise.
• Not a security claim. The demos show provenance and review; they do not show — and Affiant does not promise — that a model cannot be wrong.
• Not the source. Meridian and HR Portal are private first-party applications. Everything they run on is Apache-2.0 at github.com/Sakwala/affiant.

Where next

• Installation — the packages, --prerelease, and which ones you need.
• Quickstart — the same card in your own host, in one sitting.
• Using Affiant with Microsoft Agent Framework — the adapter Meridian runs on.
• Using Affiant with Microsoft.Extensions.AI — the bridge HR Portal runs on.
• Found something? Open an issue.