Providers & keys

Sign in to your AI WorkForce account to use cloud models and credits — no API keys needed. Bringing your own provider keys stays available below. Changes save automatically.

Native setup check

Checking your setup...

Rust is reviewing provider readiness, safety, memory, and connectors.

AI WorkForce Recommended

Link this computer to your account with a short code you approve in the browser. Tasks you start on the web can then run here — each command still asks on this machine first.

Checking link…

Advanced — control-plane endpoint

The AI Workforce base URL. Leave blank to use the default.

Active provider

Claude (Anthropic)

Get a key at console.anthropic.com → API keys. Keys start with sk-ant-. Pay-per-use; a Claude Pro/Max subscription does not include API access.

OpenAI (ChatGPT)

Get a key at platform.openai.com/api-keys. Pay-per-use; a ChatGPT Plus subscription does not include API access.

Google Gemini

Get a free key at aistudio.google.com/apikey (Google AI Studio). The free tier covers Flash models.

Custom (OpenAI-compatible)

The one-click option: sign in once and use GPT, Claude and Gemini models through a single OpenRouter account — no key copying. Includes free models (:free); paid models use OpenRouter credits.

Any endpoint that speaks the OpenAI chat-completions API: https://openrouter.ai/api/v1, local Ollama at http://localhost:11434/v1 (leave the key blank), etc.

Models

Pick a model per provider, or leave the default. "Fetch models" lists everything your key can use; "Custom…" lets you type any model id.

Model catalog (models.dev)

Adds context/pricing metadata to the pickers below and powers provider/model suggestions.

Used for cheap tasks like chat titles (config: small_model). Blank = the active model.

Claude

OpenAI

Gemini

Custom endpoint

Required for most custom endpoints — there is no default model.

Behavior

How the agent runs and responds.

Tool actions the agent may take before pausing (1–120).

Response length cap per turn (256–32000).

Show the reply as it is generated.

Suggest planned rounds (Manager → Executor → independent Auditor) when a request outgrows one session. Verified progress survives restarts.

Rounds before the task pauses for review (1–30).

Executor tool actions per round (1–120).

Retries per subtask after a failed audit (0–5).

How much the model thinks before answering. Higher is more thorough but slower.

Stream the model's reasoning in a collapsible block above each reply.

The agent outlines its steps as a live checklist and ticks them off as it works.

Thinking is shown for Claude and Gemini reasoning models. OpenAI applies the effort level but does not expose reasoning text.

default

Lower is more focused, higher more creative. Some models ignore or reject custom temperatures — the default is safest.

Allow the screenshot tool and image attachments. Requires a vision-capable model.

Lets the agent run shell / PowerShell commands, read & write files anywhere on disk, launch and control other applications, move the real mouse & keyboard, and capture the whole screen. Auto is powerful but risky — the agent will act on your computer without confirmation. Only available in the desktop app.

Type with a realistic cadence and move the cursor along eased, curved paths with natural pauses between steps.

"Instant" behaves like the classic fast path even when human-like acting is on.

How many focused sub-agents a task may spawn — they run one at a time, sharing the browser tab (1–5).

Appended to AI WorkForce's built-in system prompt on every run.

Agents

Custom agents the main agent can delegate subtasks to by name (via spawn_subagents). Each has its own instructions, and optionally its own model and step cap. The cap on how many run per task is in Behavior → Max sub-agents.

Your agents

AgentModelModeOn

Enabled agents are listed (name + "when to use") in the model's system prompt so it knows when to pick each one. Disabled agents are kept but never offered.

Add an agent

One line shown to the main agent so it knows when to delegate to this agent.

A provider/model ref; bare model ids are treated as ids on your active provider. On a custom/OpenRouter endpoint, write openrouter/vendor/model (or custom/<id>) to be explicit — bare OpenRouter slugs like openai/gpt-5.5 stay on the custom endpoint.

Tool actions this agent may take per subtask (1–40).

Tools

Uncheck the tools this agent must not use. All are allowed by default.

Built-in agents build, plan, general, and explore exist automatically — override or disable them in the Config (JSON) page.

User profile

Tell AI WorkForce who you are. The profile is stored locally and added to the system prompt of every run (and sub-agent run) so replies can be personalized.

Turn off to keep the fields below saved but never sent to the model.

It is sent to your chosen model provider with every request, so keep passwords and payment details out of it. For response-style rules ("always answer in French"), Behavior → Custom instructions works too.

Memory

Save pages and notes so the agent can recall them later. Searches use your embedding key when available, with a keyword fallback otherwise.

Lets the agent save and search snippets from pages and notes.

Automatically pull relevant saved notes into context at the start of each task.

How many top matches to include per search (1–10).

Saved knowledge

ItemSaved

Add note

Permissions

Control when AI WorkForce may click, type, or navigate on your behalf.

Default mode

Site rules

Per-site overrides beat the default mode. "Always allow" in the panel adds an Auto rule here.

SiteMode

AI WorkForce can click, type, and navigate on your active tab. Keep "Ask first" on while you get comfortable, supervise what it does, and never ask it to enter passwords or payment details.

Appearance

Applies to the side panel and this page immediately.

Theme

Data

Everything is stored locally in this browser.

Export & import

Export downloads conversations, tasks, agents, site rules, the JSON config, and settings (including your user profile) as JSON — API keys are never included. Import merges conversations, tasks, agents, and site rules from an export file (the config is restored only if yours is empty).

Danger zone

Reset restores every setting to its default and removes your API keys and user profile. Conversations, tasks, and agents are kept.

Advanced

Integrations for power users.

Expose this browser's tools to an MCP host over a local server. Status: off

Must match the browser-mcp server's port (default 9234). Run the server and register it in Claude Code:
claude mcp add browser -- node /ABSOLUTE/PATH/browser-mcp/server.js

Let tasks you start on the AI WorkForce website run on THIS computer. Your desktop app must stay open. You can approve or stop tasks from the web or here. Status: off

Remote approvals are single-use and can never grant "always allow" — every action is still gated by your permission rules.

Connect the agent to external MCP servers (local stdio commands or remote http URLs). Their tools become available in chat and in workflow Run tool nodes.

Config (JSON)

AI WorkForce's JSON config, layered over these settings. Model refs are provider/model, e.g. anthropic/claude-sonnet-4-6.


        

Report an issue

Tell us about AI-generated content in this app that was harmful, inaccurate, or otherwise a problem. Reports go to our team — the model never sees them.

Report AI-generated content

You can also report a single response directly from the chat: hover any assistant reply or task result and choose Report. That way the exact text you are reporting is attached for you.

The report opens prefilled in your mail app, addressed to info@risessolutions.com. If this device has no mail app, the dialog lets you copy the full report instead.

What we ask for

A category (harmful / inaccurate / other), a short description in your own words, and — only if you tick the box — the conversation excerpt you are reporting. Nothing is sent anywhere until you send the email yourself.

Confirm action

Saved