The control plane for AI-assisted engineering

Every AI coding request, routed, governed, and accounted for.

srooter is the independent layer between your developers' AI tools and every model provider. It sends each request to the right model, enforces your standards, and records every decision — without changing how anyone works.

Drop-in · OpenAI- & Anthropic-compatible · one environment variable
srooter — live requests · acme
RequestedServedRouteCostLatencyStatus
claude-opus-4-8qwen3 (local)trivial → qwen$0.0000293mssuccess
claude-opus-4-8claude-opus-4-8council : security$0.00001.8ssuccess
grok-4.3deepseek-v4-profallback chain$0.00012.9sfallback
claude-opus-4-8qwen3 (local)learned-route$0.0000339mssuccess
Opus asked. Qwen delivered. You paid nothing.
Works with the tools your team already uses
Claude CodeCodexaider Anthropic OpenAI DeepSeek xAI Gemini Ollama · local
The problem

AI coding is now your fastest-growing — and least visible — cost.

Agents burn tokens in autonomous loops. A single weekend refactor can run into thousands. Nobody can see which model did what, no two teams follow the same practices, and a confident bad diff ships to production before anyone reviews it.

01

Runaway spend

Token bills scale nonlinearly with agent loops, with no cap and no warning until the invoice lands.

02

Zero visibility

No record of which model handled which request, at what cost, with what outcome.

03

Inconsistent practice

Every developer prompts differently. Standards live in people's heads, not in the pipeline.

04

Unreviewed risk

Agents touch auth, payments, and migrations with the same casual confidence as a typo fix.

One layer, three jobs

Route smart. Remember everything. Govern it all.

Useful today

Productivity routing

The right model for every task — picked automatically, across every provider you connect.

  • Trivial work to fast, cheap, or local models
  • Hard work stays on your best model
  • Cross-vendor by design
  • Automatic fallback — nothing blocks the dev
Gets smarter

Harness intelligence

srooter learns your codebase and your standards, and applies them to every request.

  • A live code graph informs routing
  • Best practices defined once, enforced for all
  • Routes that learn from outcomes
  • Cross-agent context a single model can't hold
Durable

Governance

An independent control plane for spend, policy, and compliance — one that outlives any single model.

  • Every decision recorded and exportable
  • Model allowlists & reasoning caps
  • Hard budget caps with graceful fallback
  • Role-based access for the whole org

Models keep getting better at routing and reasoning. The layer that remembers your org and enforces your rules doesn't get absorbed into them — it lives around the model, not inside it.

Capabilities

A complete control surface for how your org ships with AI.

Routing & council

The cheapest model that can actually do the job — and a council when it can't.

An intent classifier reads every request before it reaches a model. Trivial work drops to local or fast models; high-stakes changes convene a council of diverse models, out-of-band, before anything ships.

Intent-aware — routes on what the task is, not keywords.
Council review on security & architecture-critical work.
Your rules or ours — start on srooter defaults, override anytime.
Routingsrooter default
TRIVIAL-TASK MODEL
qwen3 · local
INTENT CLASSIFIER
gemini-classify
COUNCIL MEMBERS
deepseek, grok, gemini
Cortex graph

It knows your codebase — and uses it to route and to gauge blast radius.

srooter builds a living graph of your org's symbols and dependencies. That knowledge tells it when a change is trivial and when it touches something that can hurt you — and routes, reviews, and escalates accordingly.

Blast-radius aware — a README edit and a payments migration aren't the same risk.
Compounds with use — the longer it runs, the sharper the routing.
Cortex graphlive
505 symbols1,394 edges63 files
Skills

Your engineering standards, defined once — enforced for every developer.

Encode practices like spec-first or TDD as skills. srooter injects them into matching requests across the whole org. Soft skills steer the model; hard skills are reserved for security and compliance.

Central, not per-repo — best practice stops living in people's heads.
Scoped by intent — apply only where it matters.
tdd-enforcersoftpriority 100
Write a failing test before implementing (TDD). Keep functions ≤ 20 lines.
APPLIES TO INTENTS
trivialsubstantivearchitecturesecurity
Audit log

Every request, every decision — recorded. Prompt content never stored.

A complete, exportable trail of what was asked, what was served, why, and what it cost. Prompts are SHA-256 hashed, so you get full accountability without ever holding sensitive content in the clear.

Hashed by default — compliance without exposure.
Shows the route — trivial, council, learned, or fallback.
Audit log121 requests
ServedRouteIn/OutCostStatus
qwentrivial → qwen9 / 12$0.00ok
opus-4-8council : security24 / 16$0.00ok
qwenlearned-route18 / 12$0.00ok
deepseekfallback12 / 258$0.0001fb
Prompts SHA-256 hashed — content is never stored.
Providers

Bring your own keys — or your Claude Max and OpenAI subscriptions.

Paste a credential and srooter discovers every model behind it automatically. API keys, OAuth subscription tokens, or a local Ollama endpoint. No registering models one by one.

Auto-discovery — models appear the moment you connect.
Subscription & local — use a Claude Max seat or run fully offline.
Anthropic Claude Max · OAuth10 models
DeepSeek2 models
xAIgrok-4.3
Ollama localqwen3-coder:30b
Gemini2 models
Policies & budget

Hard limits, enforced invisibly at the gateway.

Allowlist models, cap reasoning effort, restrict service tiers, and set a monthly spend ceiling. Go over budget and requests fall back gracefully down each model's chain — work never stops.

Graceful, not blocking — caps degrade, they don't break.
Per-org policy applied to every request automatically.
Spent this period$612 / $2,500
Above the cap, requests fall back silently to the next model in each chain — never a hard stop.

Dashboard

Spend, latency, and recent routing at a glance.

API keys

Mint & revoke per-developer gateway keys.

Audit log

Hashed, exportable record of every request.

Cortex graph

Your org's code knowledge, visualized.

Skills

Central practices, enforced everywhere.

Routing

Intent classifier, council, thresholds.

Policies

Allowlists, reasoning caps, tiers.

Budget

Monthly caps with graceful fallback.

Providers

BYOK, OAuth, local — auto-discovered.

Users

Role-based access for the whole team.

How it works

Transparent by design. It never takes the wheel.

1

Point your tools at srooter

One environment variable. Claude Code, Codex, and aider keep working exactly as before — srooter stays invisible to your devs.

2

It governs every turn

Each request is checked against policy, routed intelligently, and recorded. Permissions and execution stay in the developer's tool.

3

You see and steer everything

One dashboard for spend, routing decisions, and the practices your org enforces — across every developer and every model.

# point your agent at srooter — that's the whole setup
export ANTHROPIC_BASE_URL=https://gateway.srooter.dev
export ANTHROPIC_API_KEY=srt_live_19d895d2…
# now just work. srooter routes, governs, and records every turn.
Security & trust

Built for the team that has to answer for it.

Content never stored

Prompts are SHA-256 hashed. You get a full audit trail without ever holding sensitive content in the clear.

Self-host in your VPC

Run the entire gateway inside your own infrastructure. Your keys, your data, your perimeter.

Your keys & subscriptions

Bring your own provider keys, OAuth subscription tokens, or local models. srooter never resells inference.

Exportable record

Every routing and policy decision is logged and exportable for review, billing, or compliance.

Role-based access

Admins set standards and see everything; developers keep their flow. Clear roles across the org.

Hard budget ceilings

Set a monthly cap. Overspend degrades gracefully down each model's fallback chain.

How we compare

Free tools stop at routing. Governance tools aren't built for code.

Most of the market is one of two things: a free proxy that picks a cheaper model, or a horizontal gateway that bolts governance onto any LLM app. srooter is the only one doing the deep, code-aware work — and the governance — for autonomous coding agents specifically.

srooter> LiteLLM OpenRouter Portkey Claude Code Router
The basics · table stakes
OpenAI- & Anthropic-compatible
Self-hostable
Your own keys · no inference resale
Cost routing & automatic fallback
Code intelligence
Intent-aware routing for code tasks
Code-graph & blast-radius awareness
Council review on critical changes
Central engineering skills, org-wide
Routes that learn from outcomes
Governance & fit
Hashed audit · policy · budget · RBAC
Built for autonomous coding agents
full partial none

The honest read: LiteLLM and Claude Code Router are free and good at exactly one thing — routing. Portkey is genuinely strong on governance. None of them understand your code, weigh blast radius, convene a council, or enforce your team's practice. srooter is the only layer doing the code-aware work and the governance, for the agents your developers actually run.

Compared in good faith, June 2026. Capabilities move fast — if anything here is out of date, tell us and we'll correct it.

Who it's for

One engine. Three people who need it.

For developers

Keep your tools and your flow. Point your agent at srooter with one env var and get the right model and your team's standards, automatically. It never takes the wheel.

For heads of AI

Define your org's engineering practice once. See which models do what, keep quality consistent across teams, and let routing learn from every shipped change.

For CTOs

An independent, auditable control plane for AI spend, policy, and compliance — one that outlives any single model or vendor.

Get started

Put one layer between your team and the chaos.

Drop srooter in front of your AI coding tools today. Self-host it, bring your own keys, and change nothing about how your developers work.

Self-hosted & cloud · OpenAI- & Anthropic-compatible · born from Maggy