Docs — AgentFactory

Getting started

Pick a template. The wizard opens with four hand-tuned patterns — Monitor URLs, Weekly research digest, Daily news brief, Writing coach — plus templates shared by the community, reviewed before publication. Click one and you skip straight to a pre-filled review screen with the full instructions on display.
Describe your own. Type a paragraph about what the agent should do. A meta-agent reads it and drafts the instructions, suggested tools, and a starter avatar.

Either path lands you on the same review screen. Tweak anything, save, and you'll be dropped into the Playground to run your first turn. Two minutes, give or take.

Agents

An agent is a saved configuration. It bundles:

a name, avatar, and short description of itself
a system prompt (the instructions it operates from)
a reasoning model, plus optional image and video models
a temperature and a max-tokens cap
a set of tools it's allowed to call
optional memory and a file workspace

You can have as many agents as you like — they're isolated and cheap to keep around. Hit Settings from the agents list to edit any field later, including swapping models or rewriting the prompt.

Wizard vs. manual

The wizard at Agents → New agent is the recommended path. If you'd rather fill everything by hand, the bottom-right link reads Create manually instead.

Playground

The playground is the agent's run-time surface. Three top tabs: Text, Image, and Video. Inside Text there are two sub-modes:

Chat. Multi-turn conversation. Bubbles for each turn, persistent history. Toggle Private to skip memory persistence for the duration of a turn.
Run. One-shot autonomous execution against the agent's own instructions — no user input required. The output renders as markdown, gets saved as a downloadable .md, and lands in the top-level Runs section so you can revisit it later.

Image and Video tabs work the same way: write a prompt, watch the result land. Costs are previewed up-front before you confirm.

Past conversations live under Conversations in the sidebar. Open any one and click Resume in playground to continue where you left off — the model sees the full transcript.

Memory

Each agent has two layers of memory, both auto-maintained. The agent reads whichever one is relevant on each call, never both at once.

User memory: A short brief about you — preferences, projects, recurring context. Built from chat history with this agent. Refreshed automatically every five chat turns. Injected on chat-mode calls only.
Run memory: A rolling list of what the agent has already produced. Append-only after each successful run; compacts older entries once it grows past ~3 000 characters. Injected on run-mode and scheduled calls only — so a weekly researcher stops repeating itself.

Both layers are visible in the editor's Memory card. You can edit by hand, force a refresh, clear either, or turn the feature off entirely with the master toggle. When disabled, nothing is injected and the auto-summariser jobs no-op.

Personas

Memory is what an agent learns about you over time; a persona is what you tell it up front. Create personas under Settings → Personas — short "about me" profiles, one per field of life. A work persona can carry your role and clients; a hobby persona doesn't need to know your employer.

Each agent uses exactly one persona: the one you pick for it in the wizard or editor, otherwise your default persona — or none at all if you set it to "None" (handy for client-facing agents). Personas never stack, so nothing leaks between your worlds. Agents cloned from community templates start on your default, never the creator's.

Personas are capped at 2 000 characters, ride along with every message (a few tokens of cost), and are sent to whichever AI model the agent runs on.

Files

Each agent has a file workspace — upload reference docs the agent can read at run time. Supports PDF, plain text, markdown, CSV, JSON, and common image formats. Up to 10 MB per file, 100 MB and 100 files per agent.

Text is extracted once at upload and cached on the row, so reads are cheap. To let the agent actually use them, attach one or more of these tools on the editor's Tools card:

list_files — returns the workspace contents with sizes and previews. The agent calls this first to see what's there.
read_file — returns the full text of a file by name (capped around 50 000 characters per call).
search_files — case-insensitive substring search across all extracted text.

A common pattern: drop a targets.txt with URLs, one per line, and pair with the Monitor URLs template — the agent reads the list each run and fetches each URL via web_fetch.

Tools

Tools let the model take action between turns. The free-tier catalog today:

web_search — live web search via Tavily. Use for current events, prices, anything time-sensitive.
web_fetch — read the full text of a URL via Tavily. Handles most pages; struggles with Cloudflare-protected ones.
list_files, read_file, search_files — see Files above.

Each tool is opt-in per agent — they don't fire unless attached. Attaching a tool doesn't force the agent to use it; the model decides each turn whether a call would help.

Schedules

Each agent can have one schedule. Use plain-language hints ("every weekday at 7am", "every Sunday at 9pm") — we translate to a real cron expression and preview the next ten runs before you save. Pause / resume any time without losing history.

Scheduled runs use the agent's Run mode and respect its run memory, so a weekly digest naturally avoids repeating last week's coverage. Failed runs can be retried automatically; budget guards skip a run rather than overspend.

Images

To generate an image from an agent, open the playground and switch to the Image tab. Write a prompt, pick an aspect ratio, hit run. The result lands inline and is saved to the agent's run history.

Two things to know:

Models differ. "Flux.2 Pro" and "Imagen 4" go for photorealism; "Flux.2 Klein" leans illustrative; "Nano Banana" (Gemini Flash Image) is fast and cheap. The catalog at /models/catalog shows the full lineup with badges, pricing, and modality info.
Cost varies wildly. A flat-rate model bills per image (e.g. ~$0.04); a token-billed model bills per output token (~$0.003 for a small image). Both land as exact debit lines in your Usage ledger.

Video

Video is async — the run queues, polls for completion, then downloads the result. Switch to the Video tab in the playground, write a prompt, pick duration and resolution. The cost estimator preview shows what you'd be charged before you confirm; you can toggle with audio on supporting models for an extra cost.

Most generations take 30 seconds to two minutes depending on the model and length. The page stays open and the result appears live when ready — you can navigate away and come back later via Runs.

Studio

Studio is a one-shot, agent-less generation surface. Open it from the sidebar. Type a sentence describing what you want, and a meta-agent picks the right model and drafts the prompt for you. Click Suggest prompt + model and you'll see the recommendation with a generate button.

Studio is ideal when you don't need a saved persona — quick thumbnails, throwaway concepts, "what does X look like?" experiments. Everything you make there lives in the Studio gallery, where you can filter by image or video, preview, download, and delete.

Choosing a model

AgentFactory routes through OpenRouter: one wallet, hundreds of models. We surface them three ways:

Digest at /models. A weekly- refreshed pick of the best, balanced and cheapest model per category, with honest strengths and weaknesses. Start here if you don't know what to pick.
Catalog at /models/catalog. Searchable browser of every model OpenRouter exposes; filter by kind (reasoning / image / video), sort by price or context length.
Per-agent picker on the editor's Models section — type-ahead matching against the live catalog.

Runs & history

Every agent run — chat turn, autonomous run, image, video, even avatar generation — produces an audit-able row in your Usage ledger: inputs, model, tokens in and out, status, exact cost. Open Runs to revisit any saved autonomous output, download it as markdown, or delete.

Conversations shows the chat-mode threads — title, last message, message count. Resume any of them in the playground.

Credits & cost

AgentFactory is pay-as-you-go. Your credit balance is denominated in millicents (1 USD = 100 000 mc) internally for precision; the UI shows USD. Every run debits the actual cost OpenRouter charged plus a small safety buffer; failed jobs are refunded automatically.

Image and video runs prefer the provider's reported usage.cost when available, so you're billed exactly what was used — not an estimate.

During beta, top-ups are admin-granted. Once we're out of beta, a self-serve top-up flow lands. No subscriptions, no auto-charges.

EU-only routing

Under Settings → Privacy you can pin your agents' text generation to our provider's EU-region endpoint. Inference then stays within the EU, and it fails closed: if no in-region provider can serve a model, the run errors with a clear message instead of quietly routing elsewhere. Around 70 models are eligible (including the Claude and Gemini flagships) — look for the EU badge in the model picker; model suggestions respect the setting automatically.

Scope, honestly: the switch covers agent conversations and runs. Image and video generation remain on standard routing until our provider offers EU-region media inference, and platform plumbing (model suggestions, avatars) also uses standard routing — your agents' conversations are never part of that.

Claude & MCP

Your account speaks MCP (Model Context Protocol) — connect Claude Code or Claude Desktop and Claude can list your agents, start one-shot runs and collect the results, in plain conversation. Everything is billed to your credits exactly like the web UI.

Create an access token under Settings → Claude & MCP — it's shown once, together with the ready-made claude mcp add command. From then on: "list my agents and run the lead radar" just works. Tokens grant full agent access to your account; revoke any you no longer use from the same page.

Available tools today: agent.list, agent.run, job.wait, media.models, media.estimate, media.image, media.video and media.gallery — so "make me a 6-second clip of a foggy forest" works too. Media jobs always quote their price first, and videos won't dispatch without a confirmed quote. Schedules are next.

Glossary

The jargon, in plain English.

Agent: A saved configuration — name, prompt, model, tools, optional memory and files. Like a persona you can talk to or have run on a schedule.
System prompt (instructions): The text the model reads first, every turn. It sets the agent's role and rules ("you are a writing coach…"). The agent's persona lives here.
Reasoning model: The text-generating model the agent uses to chat or run. Examples: Claude Opus, GPT-4o, Llama 3.1.
Temperature: How wild the model gets. 0 = deterministic, predictable, repetitive. 2 = creative, varied, occasionally weird. 0.7 is a sensible default for most chat agents; 0.3 is good for research or anything where you want consistency.
Token: The unit a model reads and writes. Roughly four characters of English, or ¾ of a word. "Hello world" is about 2 tokens. Cost is per token (usually quoted per million tokens).
Max tokens: The hardest cap on output length per run. Higher = the model can produce longer responses. 2 048 is fine for chat; 4 096 or 8 192 makes sense for long-form runs.
Context length: The maximum input + output the model can handle in one call, measured in tokens. Bigger context = more material the agent can read in a single turn (long docs, long chat history). Listed on each model in the catalog.
Prompt: What you type into the chat or image / video tab. The message you want the model to act on.
Conversation: A multi-turn chat thread with an agent. Each thread is persisted unless you mark a turn private.
Tool: A function the agent can call between turns — web search, file read, etc. The model decides when to call; you decide which tools are available.
Tool call: One invocation of a tool, with the arguments the model chose. Visible in the run history if you want to debug what the agent did.
Modality: The kind of input or output a model handles: text, image, video, audio. Multimodal models handle more than one — e.g. vision models accept image input.
Run mode: One-shot execution. The agent runs against its own instructions, produces an output, done. Good for daily digests, scheduled reports, anything the user doesn't steer in real time.
Chat mode: Multi-turn conversation. The agent reads each turn and replies; history is kept so you can pick up where you left off.
Memory: The agent's running notes. See the Memory section above for the two layers.
Avatar prompt: One sentence describing the cartoon mascot for the agent. The platform adds the visual style automatically and generates the image in the background.
Aspect ratio: Width-to-height ratio for image gen. 1:1 = square, 16:9 = widescreen, 9:16 = vertical / story format.
Credit / millicent: Internal accounting unit. 1 USD = 100 000 millicents (mc). Lets us bill fractional cents accurately without floats. The UI shows USD; the ledger uses mc.
OpenRouter: The provider AgentFactory routes through — one API across hundreds of models from Anthropic, OpenAI, Google, Meta, BFL, and others. Pricing is theirs; we pass it through with a small safety buffer on upfront estimates.

Use thefactory.