Chats
Drop files to attach
New Chat
How can I help?
Drop files here
✓ Tasks
Idle
To Do 0
In Progress 0
Review 0
Done 0
Model Priority

Models tried in order. First non-rate-limited model is used. If all are cooling down, waits for the shortest cooldown.

Worker Model

Model used for worker subagent calls. "Default" follows the priority list rotation. "Priority model" always forces the #1 model in the list.

Utility Model

Model used for background tasks: chat title generation, prompt suggestions, etc. "Default" uses the first available model from the priority list. "None" disables all utility LLM calls.

Sampling

Override LLM sampling parameters. Empty = provider default.

Temperature
Top P
Top K
Min P
Presence penalty
Repetition penalty

Applied as defaults to custom endpoints (vLLM / Ollama / LM Studio). Other providers use their own defaults. Set a value to override; leave blank for provider default.

API Credentials
Gemini Key
Mistral Key
Groq Key
Cerebras Key
OpenRouter Key
NVIDIA Key
OpenCode Key
TokenHarbor Key
Kilo Key
Vercel AI Gateway Key
Nous Portal Key
HuggingFace Key

HuggingFace key enables FLUX / SDXL via HF Inference as a fallback for generate_image (Pollinations works without a key). Env var: HF_API_KEY.

Media Model Routing

When a message includes multimedia, requests are routed to these models. "Auto" uses the first capable model from the priority list.

Model Catalog

All available models. Use ✕ to remove a model from the catalog.

Save Current

Snapshot all settings (except API keys) as a named profile.

Web Search

Direct browser call — no proxy needed.

Required on GitHub Pages for providers that block direct browser calls (OpenCode, Kilo, NVIDIA NIM, TokenHarbor, Vercel). Also used for Brave search. Deploy cf-worker/ to your own Cloudflare account (free tier) and paste the URL — see cf-worker/README.md for instructions.

Sent as X-Brave-Key header.

Upgrades GitHub search from 60/hr to 5,000/hr.

Upgrades Stack Overflow search from 300/day to 10,000/day.

JavaScript in Browser

Always active — no configuration needed. execute_code(language="javascript") runs in a browser sandbox with a virtual filesystem shim (fs, path modules). Files written via fs.writeFileSync sync back to the workspace automatically. Limitations: no npm, no DOM or canvas, no network access. For canvas/DOM work the AI can generate a self-contained HTML file instead.

Python in Browser (Pyodide)

Runs Python in the browser via Pyodide. Workspace files are pre-loaded at their normal paths; writes sync back automatically. Use micropip.install for pure-Python packages.

Not loaded
Bash Sandbox

Controls bash execution only. JavaScript uses the browser sandbox above; Python uses Pyodide.

Git Integration

Exposes the run_git tool and injects git instructions when messages mention git, commit, push, or pull. Requires Local Bash Sandbox (server.py). Disabled by default.

Code Structure Queries (AST)

Exposes ast_query(path, query) for precise symbol navigation with exact line numbers — find definitions, call sites, imports, exports, and references without grep. Supports: "functions", "classes", "imports", "exports", "symbols", "calls:NAME", "references:NAME", "symbol_at:LINE". Works in all environments; available to all agent roles when enabled. Disabled by default.

Context Management

Cap tool results stored in history to avoid context bloat.

When a read_file result in the main agent context exceeds this length, it is replaced with a structural index of definitions (with line numbers) plus as many full function bodies as fit. Other tools are passed through as-is. 0 disables. Default: 20000.

Compact history before hitting the hard context limit.

% of window
Token count
Prompt Construction

Nudge the agent to prefer execute_code scripts over chained tool calls.

One sentence of intent between tool calls; telegraphic task strings for workers; workers respond in compressed style. Reduces history growth without affecting output quality.

Agent Loop

Emit a step-by-step plan file before executing multi-step tasks.

Base cooldown duration per endpoint after a 429 rate-limit response. Doubles on each consecutive hit (exponential backoff), capped at 10 minutes.

Give workers scoped context and minimal tool subsets instead of the full system prompt.

Prefix workers with the full main conversation history (prefix-cached, so cheapest tokens). Workers get complete context; main agent needs less detailed task strings. Disabled by default.

When total worker text output exceeds 2K chars, synthesise all outputs into one compact summary before returning to the main loop.

Automatically uses the first model for planning workers; later models in the priority list for execution workers.

Each model in your list is used for N consecutive steps before rotating to the next. A new turn always starts from the first model. Spreads load to stay within per-model TPM limits.

Hierarchical delegation depth. At depth 1, coder workers can spawn their own parallel sub-workers for independent sub-tasks. Sub-workers cannot spawn further. Depth 0 disables the feature entirely.

Thinking budget applied to worker agents. High by default — workers often tackle harder sub-problems than the main loop.

Pass preserve_thinking: true to the vLLM chat template so the model's reasoning from each step is re-inserted as a <think> block in subsequent turns. Improves multi-turn continuity for Qwen3 and similar models. Only affects provider:vllm with a thinking budget > 0.

QA Gates

Run gate checks at each Kanban column transition. When disabled, tasks move directly to done without checks.

Check each criterion in the task's ## Acceptance section against the changed files before marking done.

Find and run test files related to changed files. Blocks completion if tests fail. Requires a code sandbox.

Run the project build/lint/typecheck command after each task. Configured via WORKFLOW.md or detected from package.json. Disabled by default.

Maximum QA failure cycles before a task is escalated to blocked instead of retried.

When a write_file call changes >50% of an existing file's lines, a second LLM call reviews the edit before it is applied. Rejects it if content was accidentally dropped or broken. Adds latency — off by default.

Before executing fetch_url, write_file, execute_code, or delete_file, validate the call against the original user intent to detect prompt injection. Heuristic: fast regex patterns. LLM: slower but catches subtler attacks.

Pause and ask for approval before executing high-risk or write tools. Disabled by default. Does not apply inside sub-workers.

Autonomous Agent Loop

When enabled, the agent loop runs lifecycle gates (task 036) after completing each task. Failures route the task to in-review for rework.

Maintain explicit ledger files for long-horizon coherence (Magentic-One pattern). Only active in Plan mode.

After each turn, save the session log to logs/ and create a task to review it for AI inefficiencies. Review tasks will create follow-up tasks for novel issues found.

Pause the loop if this many tasks in a row fail without progress. Gives you a chance to inspect and redirect before continuing.

Skills are procedural workflows invoked with /name. Rules are standing context injected automatically by trigger. Checked items are always injected into every prompt.

Tools

Checked tools are available to the LLM in every conversation.

Worker roles define the system prompt, tool set, and auto-injected skills for each specialist agent. Uncheck a role to prevent workers from using it (they fall back to the generic worker prompt). Edit a system prompt and Save to override the built-in; Reset restores the default.

Speech Input (STT)

Models tried in order. First with an available key is used.

Language
Speech Output (TTS)

Models tried in order. First with an available key is used.

Auto-speak replies
Browser voice Applies to Web Speech Synthesis only
Speed 1
Pitch 1