Oct 2026
Multi-domain serving
One deployment answers on several hostnames, with the self-reference guard covering all of them.
An assistant that works inside your product, not next to it. It runs under your name, in your branding, acting with each user’s own permissions.
Three people ask the assistant for help. Lina asks where to add a teammate, and it rings the Invite button in her app. Sara asks for a refund of unused seats; it reads the billing policy, checks the March invoice, waits for her approval, and refunds £128.40 as her. Omar asks it to check seat usage every month, and it schedules that as a recurring job.
A twenty-second demo. A user asks where to add a teammate and the assistant rings the real Invite button in their app. Asked about a wrong invoice, it reads the invoice, answers from the billing policy and cites it. It then asks for approval before refunding £128.40, and runs the refund as that user once approved.
The tools page of the console: a searchable list of webhook and MCP tools with call counts, one of them switched off.
Every outbound call is SSRF-checked and carries a stable idempotency key, so an ambiguous failure never fires a write twice.
Connect any MCP server and its tools arrive disabled, so you choose what the agent is allowed to touch.
The knowledge page: ingested sources on the left, and on the right the passage retrieved for one answer with the matched phrase marked.
Each user can carry private documents only they can retrieve, scoped by signed identity rather than by a guessable id.
No knowledge base, no retrieval step. The agent stops paying for a search that was never going to return anything.
An embed snippet beside the page it produces: the widget docked in a customer app, with the Export button ringed.
Because it lives in your page, it can ring the actual control instead of describing where to find it.
Dictate instead of typing and have answers read back. Browser-native, so it costs nothing and works with every model.
A signed user_context payload with line numbers, and the three things the signature enforces.
Anything that moves money or data waits for an explicit confirmation before it fires.
Per-project rate limits and a monthly ceiling, enforced server-side rather than hoped for.
The analytics page: turns, tool calls, cache reads and spend, a seven-day chart with one column’s detail open, and the last three tool calls.
Input split across tool schemas, system prompt, history and the message itself. The parts add up to the total exactly.
Open a single conversation and see the exact context the model was given, not a reconstruction of it.
The platform underneath
It logs itself
Every turn, tool call and token, priced and searchable.
Your tools talk
Webhooks and MCP servers arrive as one toolset.
It knows the user
Signed identity, their permissions, their credentials.
It asks first
Writes pause for approval before anything moves.
It stays in budget
Rate limits, a monthly ceiling, and a balance per user.
AGENTIFYS MCP
Point your coding agent at one URL and it provisions the whole project: prompt, user schema, tools, policies, connected MCP servers, limits and scheduling. Twelve tools, a scoped token, and no secret is ever an argument.
$ claude mcp add --transport http agentifys https://agentifys.ai/mcpA terminal transcript: a coding agent creates the project, configures signed identity, sets the user schema and prompt, adds a tool, connects an MCP server whose nine tools import disabled, verifies the project, and finishes with one manual step: pasting the model key, which no tool accepts as an argument.
A documented HTTP API, an embeddable widget with a JavaScript handle, and the MCP server above. Three ways in, and underneath all three the same runtime, the same tools, and the same per-user credentials. Pick whichever fits how you build.
We checked all six
Almost nobody can
act as the customer.
Measured, not estimated
93%
of prompt tokens served from cache on live projects
80%
lower cost per model call once caching is on
1
script tag to embed the whole agent
0
shared keys: every project spends its own
Oct 2026
One deployment answers on several hostnames, with the self-reference guard covering all of them.
Sep 2026
Repeated tool schemas serve from cache: about 80% off a typical model call.
Sep 2026
Configure screen elements in a setup screen instead of hand-editing the system prompt.
Sep 2026
See which part of the context your spend came from: schemas, prompt, history or the message.