ScoutCode ships Scout, a Rust-built desktop coding agent that pools free model tiers and automatically fails over when a provider rate-limits, according to the ScoutCode product site on November 11, 2026 and the open-source ScoutCode GitHub repository. Version 0.1.24 runs on Windows and Linux, works without an account for basic limits, and centers on a harness layer: permission modes, memory files, context pruning, and optional ScoutMote remote control.
This is not another chat wrapper. The site argues models are only half the product; tool policy and session hygiene determine whether free tiers are usable for real repos.
TL;DR
| Feature | ScoutCode claim |
|---|---|
| Free models | 4 models, $0 per request with rate caps |
| Fallback | Rotates when a provider throttles |
| Memory | .scout/memory.md plus ~/.scoutcode/memory.md |
| Permissions | 5 modes, diff-first edits |
| Remote | ScoutMote browser control of desktop session |
| Size | ~18 MB executable, WebView2 UI |
Why failover matters for free-tier agents
Free-tier model pools with automatic failover when a provider rate-limits
Free models are attractive until 429 responses mid-refactor. ScoutCode pools ScoutCode providers, Kilo Code, OpenCode, and OVH Cloud free tiers and moves to the next model automatically.
That pattern mirrors model routing in production gateways — see Amazon Bedrock gateway fallback chains and Cloudflare token wars routing — but targets solo developers who refuse to pre-pay for every experiment.
Tradeoff: you accept behavior drift when the fallback model uses different edit tools. Scout addresses that by matching patch formats per model family on the site.
Permission modes in plain language
Tiered permission modes from ask-first review to unrestricted yolo execution
Scout exposes five modes one keystroke apart:
| Mode | Edits | Commands | Typical use |
|---|---|---|---|
| Ask first | Waits for yes | Asks | Default safe coding |
| Accept edits | Auto write | Asks | Fast typing, cautious shell |
| Auto | Auto | Mixed holds | Daily driver |
| Plan | Read-only until plan approved | Refused | Exploratory refactors |
| Yolo | Auto | Auto | Disposable VMs only |
The site shows git push --force held for review with a reason — similar spirit to Herdr terminal runtime guards.
Yolo is not for laptops with production keys. Treat it like Nightly OpenBot full-access warnings.
Memory that edits itself
Agent memory files pruned and rewritten instead of only appended
Most agent memory features append bullets until context rots. ScoutCode claims Scout revises, tightens, and deletes stale decisions while you work.
Files land in:
- Repo:
.scout/memory.md— committable for teams - User:
~/.scoutcode/memory.md— personal style
That split resembles AGENTS.md conventions but automates hygiene. Pair with git review like any generated config.
Context pruning and compaction
Long sessions accumulate tool output. Scout prunes old file views and command logs, then summarizes older turns while keeping recent messages verbatim — the site cites thresholds near 60k prune and 90k compact.
Compare with Claude Code idle compaction forks. The goal is the same: stop paying tokens for dead stderr.
ScoutMote and MCP bridge
ScoutMote lets you start chats, approve permissions, and switch models from a phone browser while Scout runs on a PC.
The Web chatbot MCP bridge exposes local file and shell tools to claude.ai, ChatGPT, Gemini, and Grok with permission gating — a different threat model. Only enable if you understand remote prompt injection into your desktop session.
For MCP hardening basics, read MCP rug-pull and tool poisoning.
Sub-agents and background shells
Scout supports read-only sub-agents for broad searches so the main chat keeps focus — the same pattern as Claude managed subagent fan-out at smaller scale.
Background shells run dev servers and builds, then attach output when needed — similar to Cursor computer-use local agents ergonomics without claiming parity.
Pricing beyond free tiers
Browser sign-in includes $5 weekly credit with a $1 per five hours burst cap, then account balance billing, per ScoutCode FAQ. API key mode bills per model on ScoutCode's catalog.
Free limits without account: 10/min, 200/day. With free account: 15/min, 400/day.
Document spend in the in-app daily chart if you mix free and paid models in one session.
What developers are saying on Hacker News
The Show HN thread (item 50458487) was thin at crawl time. Treat early feedback as anecdotal until more hands-on benchmarks appear.
When comments grow, compare reported effective limits against ScoutCode's published caps — free tiers change silently.
How Scout compares to neighboring harnesses
| Harness | Free model angle | Isolation story |
|---|---|---|
| Scout | Built-in pools + fallback | Local desktop, permission modes |
| Cindy harness switching | BYO keys | User-chosen sandboxes |
| Cowagent memory channels | BYO models | Channel-based workflows |
| Claude Code | Subscription | Vendor sandbox options |
Scout optimizes $0 experimentation. Teams with compliance needs still route production work through governed CI agents — Claude Code remote AgentCore sample.
Install and first session
Linux install from ScoutCode:
curl -fsSL https://scoutcode.dev/install.sh | sh
Windows ships a user-level installer without admin; Arm builds exist.
Open a repo, pick Ask first mode, and run a small lint fix. Watch:
- Which model answers first.
- Whether fallback triggers on repeated prompts.
- How
.scout/memory.mdmutates after you correct style.
Risks to state plainly
- Free tiers can change limits or model lists without notice.
- Yolo mode on a machine with cloud credentials is negligent.
- MCP bridge multiplies attack surface.
- Scout sends file contents to whichever model you select, under that provider's terms — same class of risk as any local agent.
Edit tools matched to each model family
ScoutCode states models train on different patch formats. Scout therefore emits exact-replace edits or diff-style patches depending on the active model — the site lists Qwen, GPT, Nemotron, MiniMax, Agnes, and DeepSeek variants.
That detail matters when you benchmark fallback: a Haiku-style diff applied by a Qwen-trained formatter may fail apply and burn turns. If you rely on automatic failover, add CI that runs Scout against a tiny fixture repo nightly and records apply success rate per model.
Delta issue tracker and in-app docs
Scout ships Delta, an issue tracker with projects, cycles, roadmaps, and Notion-style docs in the sidebar. You can hand an issue straight to Scout from that UI.
For teams already on GitHub Issues, compare with agent autoloop cron hubs. Delta only wins if you want one window for planning and execution; otherwise duplicate state machines confuse agents.
Stack: Rust, WebView2, and why size claims matter
Scout advertises ~18 MB one executable, sub-half-second window open, and ~120 MB RAM by using the OS WebView2 instead of bundling Chromium.
That is meaningful on Windows laptops with tight disk quotas, less on Mac until a build ships. The terminal panel hosts a real shell beside chat — closer to Cursor local agents than browser-only chats.
Allow lists and command hygiene
Scout supports prefix allow lists — approve npm test once per project. Combined with Jev-rated command holds, you get a poor person's policy engine without standing up OPA.
Document your allow list in repo README so new contributors know which commands Scout may run unattended. Revoke prefixes when scripts gain network calls.
Voice, attachments, and @ file picks
Mic dictation and @ file attachments land in the prompt box like other agents. Treat microphone input as untrusted text on shared machines.
Cost dashboard and OpenRouter BYOK
Scout shows session cost and balance with a daily chart in Settings. You can add OpenRouter keys and bill models directly, bypassing ScoutCode pools.
That hybrid mode is how power users keep free routing for exploration and paid routing for release branches — similar mental model to DeepSeek Harness marketplace creator picks.
Team rollout suggestion
| Stage | Action |
|---|---|
| Pilot | Single repo, Ask-first mode, no MCP bridge |
| Team | Commit .scout/memory.md style rules via PR review |
| CI hook | Nightly Scout task on lint-only issues |
| Block | Ban Yolo on machines with cloud SSO sessions |
Steering, rewind, and undo semantics
Scout advertises mid-turn steering: queue a follow-up while a turn runs, or inject into the active turn. Rewind edits an earlier user message and restores files to that checkpoint; /undo mirrors undo for the last message.
Those features are powerful for experimentation and dangerous on shared branches. Policy: rewind only on feature branches, never on main with uncommitted teammate work. Compare semantics with Claude Code fork and subagent compaction so engineers know which harness supports which undo model.
Project instructions precedence
Scout reads SCOUT.md, AGENTS.md, or CLAUDE.md from git root down to the opened folder. Conflicting instructions across depths are a classic foot-gun.
Standardize one canonical file per repo and symlink others during migration. Document precedence in CONTRIBUTING so Scout does not inherit stale Claude-only rules meant for another product.
When free Scout is the wrong tool
Skip Scout as primary harness when:
- You need Mac support today.
- Compliance mandates vendor SOC2 on the harness binary itself, not only models.
- You run air-gapped builds — Scout assumes network model calls.
In those cases, pair Capsem air-gapped proxies with a single paid model instead of free rotation.
Support expectations
ScoutCode is a small vendor shipping fast iterations (0.1.24 at crawl). Budget time to re-read release notes weekly if free model IDs change. Pin installer scripts in internal docs instead of telling engineers to pipe latest blindly.
Related reading
- Taste vs judgment filters for AI drafts
- Agent autoloop background hubs
- Decision models for cheap routing
- Shadcn skill audit pattern
- Claude cage network isolation
Product details accurate to scoutcode.dev retrieved November 11, 2026 UTC.