LiveModel routing is now available in Claude Code. Same agent. The right model for each subagent.Model routing is live.Get startedReportBenchmarking Sonnet 5.5: less thinking for the same answers.Benchmarking Sonnet 5.5.Read
Caveman

Caveman Cloud Free to start

Run for less.
Understand every run.
Improve automatically.

“We found some issues.
Here’s how much they cost.
Here’s the PR.”

One connected platform for your agents and applications. Route requests, trace the work, and turn wasted spend into tested fixes in your repository.

Start free Try the platform

From a run to a pull request · 17s

Your agents, requests, and costs in one workspace.

Chapter 1: Connect the work

Sample project.
Illustrative improvement.
Caveman Cloud · Connect the workdemo data

01 / Run Make every request count

Same work.
Less wasted spend.

Route to the right model. Reuse cached context. Set budgets before requests run away. Use the Caveman gateway with your own provider keys.

Model routing and fallbacksCaching and spend controlsOr keep your gateway and connect its traces
Explore the gateway
Request pathIllustrative workflow
Coding agentsApplications
Caveman gatewayYour keys. Your controls.
OpenAIAnthropicGoogle
Route the modelReuse the cacheEnforce the budget
RequestDecisionResult
support / classifySmaller eligible modelRouted
research / follow-upRepeated contextCache hit
code / reviewProject spend limitWithin budget
Already on LiteLLM? Keep it. Connect your traces.
Your organizationExample workspace

Every agent. One view.

Cost per successful task
Worth investigatingRepeated tool calls

Why are support tasks getting more expensive?

90% task success in this sampleFictional data

02 / Understand Know where to act

See what your
AI spend delivers.

Connect cost to completed work across teams. Find the agents, retries, and workflows worth improving.

Cost per successful taskFailures and retries includedEvidence behind every finding
Explore the workspace
One shared record

From “what happened?”
to “here’s why.”

Agents and employees work from the same traces, tool calls, costs, and evaluation results. Follow a finding back to the run that caused it.

Inspect a trace
support-agent / pull requestDraft
Repeated lookups$4,200 / monthEstimated waste · illustrative

Reuse completed lookups
across retries.

tools/refund.ts+1 −1
− lookupRefund(orderId)+ lookupRefundOnce(task.id, orderId)
Regression test included Evaluation attached
Your review comes next.

Illustrative proposal · No repository is changed

03 / Improve Close the loop

Issues found.
Cost explained.
Pull request ready.

Caveman investigates repeated work, runs experiments, and tests fixes on your tasks. You get a pull request with the change, regression tests, and before-and-after evidence.

Investigate→Experiment→Evaluate

Your quality gates. Your team decides what ships.

Explore agent improvement

Every fix becomes a regression test.
Every new run is another chance to improve.

Explore evaluations

04 / Connect

Works with the stack
you already run.

Keep the pieces you chose.
Connect the missing loop.

The connection mapChoose your starting point ↓
Provider-compatible requests

See the work behind the spend.

Bring supported coding-agent traffic through the gateway. Attribute usage to developers, inspect runs, and find repeated work across your team.

Explore developer usage
Bring your own provider keys
OpenAIAnthropicGoogle
Built for agents, too. Use Cloud headlessly or alongside your team in the console.Explore headless use

Start with one workload

Bring the work.
Find the next improvement.

Connect your agents, gateway, and repository.
See the issues, their cost, and the path to a fix.

Start free Free to start · Workload-specific evaluation