Sidewrite

Hit your Claude usage limit? Keep coding.

Claude Code burns through your weekly limit fastest during implementation — reading files, running tools, iterating. Sidewrite keeps Claude for planning and review, delegates that grind to another model (free-tier or your own key), and switches you over automatically the moment you run out.

No external dependencies. Just Node. Apache-2.0.

Install

npm i -g sidewrite
sidewrite install

Requires Node.js ≥ 22.5 and the claude CLI. Something off? sidewrite doctor tells you what to fix.

Quick start

On a Claude subscription — keep working normally, then delegate the implementation:

/sidewrite-delegate

No subscription, or want to go fully external:

sidewrite mode standalone
sidewrite code

Either way, sidewrite opens the dashboard so you can watch it happen.

Two modes

| | subscription (default) | standalone | |---|---|---| | Plans & reviews | You + Claude, as usual | The external model | | Writes the code | A cheap external model, headless | Same model — no subscription touched | | Delegate with | /sidewrite-delegate | sidewrite code |

Switch anytime: sidewrite mode subscription / sidewrite mode standalone.

Run for $0 — the Free-Tier Pool

One local gateway fronts 9 free-tier providers (197 models, no card required) and rotates across all of them automatically:

• Tier-aware fallback — exhausts every same-tier candidate before ever dropping a request to a weaker model.
• Auto-renewing budgets — a cooling-down provider isn't dead weight; its rate limit refills continuously and it rejoins rotation the moment it has room, no manual reset.
• Failover on the fly — a bad key, rate limit, or timeout falls through to the next provider mid-conversation, with a context-handoff note if it has to swap.
• Sticky sessions — you stay on the same provider for a conversation whenever possible.

Providers

17 bundled providers speaking Claude's own Messages API out of the box — OpenRouter (256 curated models), DeepSeek, GLM/Z.ai, Moonshot/Kimi, MiniMax, Qwen, Fireworks, Baseten, Cloudflare AI Gateway, and more — plus any custom Anthropic-compatible endpoint, or a local model (Ollama, LM Studio, vLLM) via a small gateway.

Token-saving features

Three opt-out features stretch a Free-Tier Pool budget further, controlled from the dashboard's Studio → Tools tab:

• Terse replies — a short instruction rides every outbound request, cutting output tokens on providers that tend to over-explain.
• History compaction — long conversations get deduplicated and truncated before dispatch, only once a request is already large (default ~100,000 tokens). Nothing is lost for good: omitted content is cached locally and the model can retrieve it on demand via a pool_retrieve tool, bounded to a few retrieval rounds so it can never loop forever.
• RTK command compression (optional) — compresses noisy Bash tool output (git, grep, cargo test, …) before it reaches the model, if you separately install the third-party rtk CLI. Sidewrite never installs it for you, and degrades to a no-op if it isn't present.

Each toggle is on by default and can be switched off per-feature from the dashboard; nothing here talks to a network beyond the provider call itself.

The dashboard

sidewrite opens a local-only site (127.0.0.1, token-protected): live runs, provider setup, skills, sub-agents, MCP, token-saving tools, analytics, budgets, health, a privacy panel showing exactly what's shared (off by default), and a feedback button for bug reports or screenshots.

Commands

| Command | What it does | |---|---| | sidewrite | open the dashboard | | sidewrite install / uninstall | register / remove the plugin + dashboard | | sidewrite doctor | diagnose your setup | | sidewrite mode [subscription\|standalone] | show or switch mode | | sidewrite code [provider] | interactive Claude Code session on an external provider | | sidewrite run [provider] "task" | delegate one task headlessly | | sidewrite status | quick health snapshot | | sidewrite up / stop | start / stop the background dashboard daemon |

Inside Claude Code: /sidewrite-delegate, /sidewrite-open, /sidewrite-status.

Auto-update

A background check against the npm registry notifies you in the terminal the moment a newer version ships, with an opt-in apply path that backs up, installs, and rolls back on failure.

Safety

• Your real Claude subscription is never run headlessly — only interactively, by you.
• Sidewrite never touches your Claude login, and never logs you out.
• External sessions run fully isolated (CLAUDE_CONFIG_DIR) and refuse to point at api.anthropic.com.
• API keys are stored locally, permission-locked (0600), never returned by any read API.

Credits

The token-saving features were shaped by prior work in the same space:

• Terse replies adapts the MIT-licensed instruction text from JuliusBrussee/caveman.
• History compaction's Compress-Cache-Retrieve design (cache what's omitted, let the model retrieve it on demand instead of losing it for good) follows the same approach pioneered by headroom — reimplemented natively in Node rather than depending on its Python service.
• RTK command compression is an optional integration with rtk-ai/rtk, a separate tool you install yourself.

Not part of sidewrite, but pairs well with it: Ponytail, a Claude Code plugin that nudges the agent toward writing less code. Install it independently from the Studio → Tools tab's link, or /plugin marketplace add DietrichGebert/ponytail.

License

Apache-2.0 © 2026 Hassan Ali [email protected] · Discord: hassanalitamam