Here is everything the framework does, in seven areas. Each feature names the setting, type or call behind it, so you can find it in the code. Each area ends with the docs that tell it in full.
The engine runs the loop for every agent. These are the things it lets an agent do, from loading only the tools it needs to asking other agents for help.
ToolsetCatalog
Only the tools it needs. A session names the toolsets it needs and nothing else loads, so the model sees only the tools it should.SkillLoader
Skills you can read. The system prompt is put together from markdown files, the principles first, so you change an agent by editing text.AskAgentTools
Agents that ask agents. An agent asks another by its address, a whole identity such as acme or one of its roles such as acme.cook, up to three asks a turn, and the answer comes back under a line that names who gave it. A team of specialists is a few addresses, not a workflow to draw.SubSessionRegistry
Follow-ups that remember. In a chat on the API, a specialist's own chat stays open after it answers, so the next question to it carries on where it left off, with the files you attached passed to it by name.MINION_LLMS
Small jobs on a cheap model. A minion role runs on the model you set for it, cheapest first, a local one if you like.CreateTool
An agent that grows its tools. The agent can write a new skill or a script-backed tool, and it waits for your /approve before any later session loads it.SessionAttachments
Files passed by name. A file you attach is handed to another agent by name, so a large file isn't copied through the model.STREAM_PACE
Paced words. A turn's text leaves the engine at a steady pace in whole words, not in the model's bursts.Pick the model with a setting, not a rewrite. It can be a cloud model or one on your own machine, and the engine works the same way with each.
LlmClientFactory
Any model, one setting. Anthropic, any OpenAI-compatible endpoint, or a local model through Ollama, picked by LLM_PROVIDER.IPromptCachingStrategy
Prompt caching. Each provider's prompt cache is used its own way, so a long chat pays less for what it has sent before.ModelReplayChatClient
Replays for nothing. A run's model answers are recorded, so the same requests replay with the same answers and cost nothing.One agent can sit in the terminal, in Rider, behind an HTTP API or behind an MCP server. Each place is a host, and the agent doesn't change between them.
CliHost
The terminal. A chat with your agent in the terminal, from two lines of your own.AcpHost
Rider. Your agent inside Rider's AI chat.SessionEndpoints
An HTTP API. Sessions over HTTP, each turn streamed back as server-sent events while it is written.Stream-Paced
Paced once. The answer's header says the engine already paced its words, so a page leaves its own pacer off.McpHost
MCP. Another client brings its own model and calls your tools, and reads the packs' feeds as resources.AccountsStartup
Accounts. Tenants, people and sign-in as building blocks, and nothing switched on until you name a database.A site names the parts it wants, and the framework builds the rest. That covers the chat, the guide in the corner, the admin and the pictures.
AddSiteAreas
A site in a few lines. A site names the areas it wants, and the framework wires them, so a site holds only its own words and data.ChatWidget
A chat in one tag. A chat any page drops in, folded in a corner, opened as the agent the page names.AgentCaption
The guide's caption. The agent's last answer as a card with two sides, its words on the front and how it was made on the back.CaptionHandOver
The answer's motion. An answer the guide's bar stops showing doesn't vanish. It fades, and goes once its caption has left.DrawnCard
Cards the agent draws. The agent writes a card, such as a ticket, and the page draws it in the site's own look.PlacesArea
Every place an address. Every section, card and page of a site has an address of its own, so the guide can take a visitor there.SiteRoutes
Found by search engines. The host draws each page as HTML for a search engine, before the app takes over in the browser.SiteFocusesBuilder
The guide knows where it is. Each part of a site tells the guide how to lead there, selling on one page and guiding on another.OverviewCountsBuilder
An admin with an overview. What the site's agents did, counted by the tools the site cares about, such as its leads.PackViews
A pack's data as screens. A pack's data in the admin as a plain table, unless the site draws a view of its own.SessionReset
Reset my session. Puts the browser back to a first visit, every saved chat ended, as a new visitor sees the site.LookSwitch
Two themes, light and dark. Glass and flat, each light and dark. The visitor picks, and the browser remembers.DiagramRenderers
Pictures from data. Flow, sequence, class, cards, nested and composite pictures from JSON, in the site's look and fonts.These keep an agent within bounds. They are written in code and not left to the prompt, so they hold whatever the model says.
ToolCallGuards
A guarded context. A tool's result over 50,000 characters is summarised, and a call the output limit cut short is never run.RobotsWebBrowser
It takes a site's no. Every fetch asks robots.txt first, and a page that says no to AI readers is refused.JSON-LD
A page's own data first. A page's own structured data is read in code before its text, so a fact isn't guessed from prose.WebSearchProviderKind
Four search engines. Tavily, Google, SearXNG or DuckDuckGo, picked by WEB_SEARCH_PROVIDER.You can see what each turn cost and what it did. The numbers come from records written as it ran, so nothing is a guess.
UsageLedger
Every request priced. Every request's tokens and cost, priced from a dated table, a file a month, and unknown is never nought.TurnLog
A line per turn. One line for each finished turn beside the ledger, which the admin's overview reads back.OtlpTelemetry
Traced end to end. Every model request, tool call and web request traced to the OpenTelemetry collector you name.These aren't features. They are rules the framework's own code is written by, so any file you open reads like the rest.
SettingChoice
A choice is an enum. A setting that names one of a few choices is an enum, read once where it comes in.turn.Paced(clock)
A name says what it is. A method is a verb, a stream operator is named for what it gives back, a type is a noun with no empty word.