SwitchyardAI coding gateway

Every AI coding call, metered and on the right model.

Your engineers run several AI coding tools, and every one asks for the biggest model, autocomplete included. Switchyard sits between every tool and every model.

Ask for a demo Book a free review

Switchyard's By tool panel on the projected month: Claude Code, Codex, Cursor, Devin and Antigravity, each routed cost against unrouted

Engineering teams run five AI coding tools, and every one sends every call to the biggest model, autocomplete included. Nobody could say which tool, engineer or team the bill came from.

One gateway every tool points at. It meters every call's tokens and cost by tool, engineer and team, routes each call to the cheapest model that can do it, and holds teams to their budgets.

How it works

Point every tool at one gateway

Claude Code, Codex, Cursor, Devin and Antigravity each take a custom base URL and key, so they route through Switchyard without changing how anyone works. Its policy file names all five.

README.md

1 Claude Code ─┐                        ┌─▶ frontier   (planning, long context)
2 Codex ───────┤                        │
3 Cursor ──────┼──▶  SWITCHYARD  ──route─┼─▶ standard   (agentic edits with tools)
4 Devin ───────┤    meter · policy      │
5 Antigravity ─┘    budget · audit      └─▶ small      (autocomplete, commit messages, lint fixes)

Each call onto the right track

Plans go to a frontier model, agentic edits to a standard one, autocomplete and commit messages to a small one. The rules are a file you can read.

config/policy.json

1"tiers": {
2  "frontier": { "anthropic": "claude-opus-5-5",   "openai": "gpt-5" },
3  "standard": { "anthropic": "claude-sonnet-5-5", "openai": "gpt-5-mini" },
4  "small":    { "anthropic": "claude-haiku-4-5",  "openai": "gpt-5-nano" }
5},
6"routing": {
7  "smallMaxInputTokens": 2000,
8  "frontierMinInputTokens": 60000,
9  "smallTasks": ["autocomplete", "commit_message", "summarize", "rename",
10                 "docstring", "lint_fix", "explain"],
11  "frontierTasks": ["plan", "architecture", "debug_hard", "migration"],
12  "honourPinnedModels": true
13}

Every token, by tool, engineer and team
The dashboard streams each call as it is routed: the model asked for, the model used, and what it cost.
Teams held to their budgets
Each team has a monthly budget on the dashboard, and spend per ticket and PR can be posted back to GitHub, Jira, Linear or ClickUp.

README.md

1 Claude Code ─┐                        ┌─▶ frontier   (planning, long context)
2 Codex ───────┤                        │
3 Cursor ──────┼──▶  SWITCHYARD  ──route─┼─▶ standard   (agentic edits with tools)
4 Devin ───────┤    meter · policy      │
5 Antigravity ─┘    budget · audit      └─▶ small      (autocomplete, commit messages, lint fixes)

config/policy.json

1"tiers": {
2  "frontier": { "anthropic": "claude-opus-5-5",   "openai": "gpt-5" },
3  "standard": { "anthropic": "claude-sonnet-5-5", "openai": "gpt-5-mini" },
4  "small":    { "anthropic": "claude-haiku-4-5",  "openai": "gpt-5-nano" }
5},
6"routing": {
7  "smallMaxInputTokens": 2000,
8  "frontierMinInputTokens": 60000,
9  "smallTasks": ["autocomplete", "commit_message", "summarize", "rename",
10                 "docstring", "lint_fix", "explain"],
11  "frontierTasks": ["plan", "architecture", "debug_hard", "migration"],
12  "honourPinnedModels": true
13}

Switchyard on the projected month: $6,004.83 saved, 39% cheaper than unrouted, $9,486.78 spent after routing

Switchyard's team budgets on the projected month: Platform $2,366.83 of $40,000, Product $2,382.73 of $32,000, Data $2,383.76 of $24,000, Mobile $2,353.46 of $18,000

The last calls of the month, as they were routed

Thirteen calls from the projected month the dashboard opens on: which tool sent each one, the model it asked for, the model it got, and the rule that decided.

UTC	Tool	Team	Task	Asked for → routed to	Rule	Tokens	Cost	Unrouted
17:59:55	Cursor	Product	autocomplete	~~gpt-5~~routed togpt-5-nano	task "autocomplete" is routine	1.6k	$0.0003	$0.0072
17:59:51	Cursor	Platform	·	~~gpt-5~~routed togpt-5-mini	agentic edit with tools	47.7k	$0.0130	$0.0652
17:59:46	Claude Code	Platform	·	~~claude-opus-5-5~~routed toclaude-haiku-4-5	small edit (32,530 tokens), aggressive policy	33.4k	$0.0371	$0.1853
17:59:42	Claude Code	Mobile	·	~~claude-opus-5-5~~routed toclaude-haiku-4-5	small edit (32,006 tokens), aggressive policy	32.5k	$0.0345	$0.1726
17:59:37	Cursor	Mobile	autocomplete	~~gpt-5~~routed togpt-5-nano	task "autocomplete" is routine	1.2k	$0.0003	$0.0075
17:59:33	Devin	Mobile	·	claude-opus-5-5kept	long context (87,520 tokens)	88.5k	$0.4611	·
17:59:28	Claude Code	Product	plan	claude-opus-5-5kept	task "plan" needs frontier reasoning	23.6k	$0.1200	·
17:59:24	Claude Code	Product	·	~~claude-opus-5-5~~routed toclaude-sonnet-5-5	agentic edit with tools	48.8k	$0.1490	$0.2483
17:59:19	Cursor	Platform	·	~~gpt-5~~routed togpt-5-mini	general request	3.7k	$0.0020	$0.0102
17:59:15	Claude Code	Product	summarize	~~claude-opus-5-5~~routed toclaude-haiku-4-5	task "summarize" is routine	8.5k	$0.0105	$0.0527
17:59:10	Antigravity	Mobile	lint_fix	~~claude-opus-5-5~~routed toclaude-haiku-4-5	task "lint_fix" is routine	2.6k	$0.0028	$0.0141
17:59:06	Cursor	Platform	autocomplete	~~gpt-5~~routed togpt-5-nano	task "autocomplete" is routine	1.6k	$0.0003	$0.0072
17:59:01	Codex	Platform	·	~~gpt-5~~routed togpt-5-mini	agentic edit with tools	43.1k	$0.0117	$0.0587

The 17:59:46 Claude Code call, through route() in src/router.ts

1Team over budget?Platform: $2,366.83 of $40,000 this monthno
2Model pinned by the caller?no x-switchyard-pin headerno
3A frontier task?plan, architecture, debug_hard, migration: no task sentno
4Long context?32,530 input tokens, under 60,000no
5A routine task?autocomplete, commit_message, summarize…: no task sentno
6A small edit with tools?32,530 tokens, within 40,000 (aggressive policy)yes

Routed to: claude-haiku-4-5 (small)
Cost: 32,530 × $1/M + 905 × $5/M = $0.0371
On the model it asked for: 32,530 × $5/M + 905 × $25/M = $0.1853
Ticket: jira:PLAT-1440, stored on the call as its ref

npm run report -- --ref jira:PLAT-1440 --post

AI spend for jira:PLAT-1440 (Switchyard)
19,702 calls · 456.57M tokens · $1182.46 after routing ($1918.54 unrouted, $736.07 saved)
- claude-code: $690.00
- devin: $305.36
- antigravity: $143.35
- codex: $24.32
- cursor: $19.44

The comment it posts to the ticket, for every call on it this month.

In numbers

$72kyearly saving
39%cheaper than unrouted
158kcalls routed monthly

For a team of 24 engineers making about 158,000 calls a month, at standard model prices. Routing small edits to small models too saves $6,005 (38.8%) a month; the default rules save $5,051 (32.6%).

Built with