FreeGent

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. A model without a key gives way to the next one that has one.

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. A model without a key gives way to the next one that has one.

Sampling

Override LLM sampling parameters. Empty = not sent: the provider's or server's own default applies.

Temperature
Top P
Top K
Min P
Presence penalty
Repetition penalty

A value you set is sent to custom endpoints (vLLM / Ollama / LM Studio); Temperature is sent to every provider. Fields are empty by default, so each model runs with its own provider's defaults.

Media Model Routing

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

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

Save Current

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

Network privacy

Where FreeGent may send requests. The level applies at once and is saved in this browser only (profiles do not carry it).

Routed — Cloudflare Worker + endpoint
Browsers block direct calls to some providers, so on freegent.ai model requests, web fetches and searches go through FreeGent's Cloudflare Worker, which forwards them and stores nothing. On a local server, providers that need one go through its own proxy. Everything else goes straight to its endpoint. Provider-only — no FreeGent.ai or Cloudflare
Your browser calls the model providers directly. Providers that do not allow browser calls (NVIDIA, Kilo) are dropped from the rotation on a hosted page, and so are web fetches and Brave search for sites without CORS headers. A local server's own proxy still works. Local — local network access only
Only localhost and private-network addresses are reachable: local models (Ollama, vLLM, LM Studio) and MCP servers on your network. Cloud providers, web search and fetch_url are refused, and code the agent runs (JavaScript, Python) and HTML previews cannot call other hosts. Libraries the page loads from jsdelivr / esm.sh (and runtimes such as Pyodide) still load once and are cached.

CORS Proxy / CF Worker URL

Optional, Routed level only. freegent.ai and the local dev server already proxy provider calls, so leave this empty unless you host FreeGent on your own domain or fork (the built-in Worker refuses other origins). Then deploy cf-worker/ to your own Cloudflare account (free tier) and paste the URL — see Hosting your own site and Worker.

Other requests

Besides the model provider you chose. The location lookup is off at the Provider-only and Local levels, and the page-view counter at Local.

Look up approximate location (sends your IP to ipinfo.io; city, region and ISP are then added to prompts that ask about weather, places or maps) Anonymous page-view counter (GoatCounter; no cookies; skipped when Do Not Track is on; off at the Local level). Applies on next load.

Context Management Tool result truncation

Cap tool results stored in history to avoid context bloat.

Max result size — workers (chars) Max result size — main agent (chars, 0 = off)

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: 100000.

Proactive compaction

Compact history before hitting the hard context limit.

Compact at (first threshold wins)

% of window	50% 60% 75% 85% 100% (not used)
Token count

Prompt Construction Script generation steering

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

Concise prompts

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

Prompt template

Agent Loop Max steps per turn Rate-limit cooldown (minutes)

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

Play a sound when Autopilot finishes a task Lean worker context

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

Director workers fork the full context (shared prefix)

Workers with role "director" fork the main agent: same system prompt, tools and conversation so far, plus their subtask (prefix-cached). Coder and researcher workers always get only their task. When off, director workers also get only their task. On by default.

Worker output reduce

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

Role model routing

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

One request per endpoint at a time

Never sends two requests to the same model endpoint at once; a second one waits for the first to finish. Parallel workers each start on a different model, taking the highest one in the priority list that is free. On by default.

Rotate endpoints every N steps

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.

Max delegation depth 0 — disabled (workers cannot spawn sub-workers) 1 — one level (coder workers can spawn sub-workers) 2 — two levels

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.

Worker thinking budget Inherit from active model Off Low (1K tokens) Medium (8K tokens) High (24K tokens)

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

Preserve thinking across turns (vLLM)

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.

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 packages, no DOM or canvas. fetch() reaches public URLs (directly when the site allows cross-origin requests, otherwise as a plain GET through the same proxy as fetch_url). 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.

Bash Sandbox

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

Provider WASM — bash shell in the browser (default) None (bash disabled) Local — via the dev server (not on freegent.ai)

Runs bash entirely in your browser, in a bash-compatible shell with 248 common commands, builtins and keywords (grep, sed, awk, find, sort, tr, …). Workspace files are available at /workspace; writes sync back automatically. A few commands (jq, lua and others) are WebAssembly packages, fetched on first use and cached in the browser (IndexedDB). curl / wget reach sites that allow browser requests directly; other sites are fetched through FreeGent's proxy (plain GET only).

Runs bash on your machine through the dev server (npm run dev), so it works only when FreeGent is run locally, not on freegent.ai. Optionally isolated. Full coreutils, npm, and pip available. Files written by scripts sync back to the workspace automatically. Python runs on your machine too unless Pyodide is loaded; JavaScript always runs in the browser sandbox.

Git Integration Enable git access

Exposes the run_git tool and injects git instructions when messages mention git, commit, push, or pull. Requires the Local code execution sandbox (dev server). Disabled by default.

Code Structure Queries (AST) Enable ast_query tool

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.

Web Search Search provider Auto (Tavily → Brave → Wikipedia) Tavily (direct) Brave (via proxy) Wikipedia only

Used by web_search and the research tools. Their keys are in the Credentials tab.

Tools

Enable tools per mode. Chat: standard conversations. Cowork: task-runner sessions (also applies to the runner).

MCP Servers

Connect external Model Context Protocol servers (Streamable HTTP). Each server's tools start disabled — enable the ones the agent should see. Tool calls send their arguments to the server's host. Context7 is built in: Tools → Context7 Docs.

Custom server

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	Speak each model response automatically
Browser voice	(browser default) Applies to Web Speech Synthesis only
Speed	1
Pitch	1