Skill Management is always included and read-only. Other available skills expose only their name and description initially. The agent may request their instructions when relevant. Skill loading can use extra inference credits. Skills do not grant tools or permissions.
Agent context
Instructions for every chat · saved on this device
Markdown text. Saved edits apply to future messages across all chats, not past replies. Instructions are included in encrypted TEE requests; they do not change app settings or security controls.
How can I help?
E2EE encrypted chat with the AI.
Cloudflare relays ciphertext · local history · browser-verified Intel attestation
TESTNET CREDITFUEL SEPOLIA
Keep the conversation going.
Your conversation stays in this browser. Testnet FUEL has no monetary value.
GPU acceleration (experimental)Enabled by default on compatible browsers; you can switch it off. Requests GPU offloading where supported. Can use more RAM and battery; start with 0.8B. Initialization failures retry on CPU locally. Generation failures require manual retry. Applies on the next local initialization; the runtime still unloads after each reply.
Download from Fuel Computer’s Cloudflare R2 mirror, then run on your device with wllama. Qwen downloads have a pinned Hugging Face backup. Gemma’s prepared shards are served only from Cloudflare. Files are pinned to a revision and checked with SHA-256. No account or credits are needed. These are community Q4_K_M conversions of Qwen and Gemma models, not hosted inference. Runtime licenses.
Local chats do not use search, web fetching, telemetry, NEAR, or remote inference fallback. Download providers see your IP and requested model, not your prompts. Enabled local skills can be loaded as tools. First-time model and app downloads need internet; keep this site’s browser data for offline reopening.
Modern desktop Chrome or Edge is recommended. CPU inference uses a single worker thread. Optional GPU acceleration depends on browser and model support; phones may run slowly or run out of memory. Larger models need considerably more RAM than their download size. Gemma 4 E2B is an experimental 3.11 GB, five-file desktop download; 16 GB+ RAM is recommended, not a guarantee. Its speed and GPU/memory compatibility still need device testing. Gemma requires Memory64 and JSPI and supports text only. Browsers can evict stored files. Downloads pause during refresh or closing the page, then resume automatically when the app reopens online. Up to 4 MiB since the last checkpoint may download again. Keep this site’s browser data. Network failures retain progress; use Resume download to retry. Cancel removes the pending download.
History, skills and model files are plaintext on-device. Local inference is not a TEE and cannot protect against compromised devices, browser extensions, or malicious site updates.
Private TEE chats encrypt messages to NEAR’s model environment. Local models run on this device without remote inference; chat history remains plaintext in this browser.
Local models
Skills library
Agent instructions
Appearance
Enable Web SearchFor remote chats only; always off in local chats. While enabled for remote chats, prompts and results are visible to Fuel Computer and Parallel. Model context and replies are then encrypted to NEAR.
Enable web page fetchingRemote chats only; local chats cannot fetch pages. Allows the remote agent to request public pages through the Worker. URLs and page content are visible to Cloudflare; websites see server-side requests. Permission is remembered until disabled. No cookies or authenticated pages.
Control fetching here. Enabling it exposes requested URLs and page content to Cloudflare. You can disable it anytime. Return to your chat and retry when ready; enabling this does not send a request.
Use local chat contextUses up to three relevant excerpts from other chats. Local inference keeps them on-device; remote inference shares them with NEAR after browser encryption. Local-inference chats are excluded from remote cross-chat retrieval.
Anonymous analyticsNot sent for local chats. For remote mode, shares broad app health and usage such as success/error category, duration range, device class, and country. The analytics dataset never includes chats, prompts, replies, searches, wallet addresses, keys, ciphertext, IP addresses, raw errors, or user identifiers.
Export local dataDownloads your local chat text, agent instructions, saved skills, selected settings, and wallet connector name as JSON. Model weights are not included. Keep the file private.