Your organization's AI use is opaque.

ixway clears the fog.

ixway in 40 seconds

0:00 / 0:37

You know what you spent on AI last month. Do you know where it went? Requests from harnesses and SDKs flow toward the models and disappear.

ixway sits between them and every request passes through it. Intelligent routing: routing you can see and set, the right model for each job. Insight extraction: know how your workforce uses AI and what it actually produced, with spend by type of work (feature work, debugging, customer support, research) and the pull requests and deploys it is linked to. Local and batch inference: your own hardware first, batch work in cheaper windows. Efficiency findings: know where the waste is. One finding, a volatile header breaking the cache prefix every turn, is sent to the harness that caused it, and that harness's tokens per session fall by 38% once it is applied. ixway teaches your harnesses to be more efficient. The figures are illustrative.

Reduce your spend. Stretch your tokens. Transparency and efficiency for your workforce AI usage. ixway.ai.

01Observability

Know what your agents did, what it cost, and why.

ixway traces every session end to end. Each agent, each subagent it spawned, each model call, each tool call with what it actually did.

Goal · inferred from session sess_4e1b

Add idempotency keys to the payments API and backfill 1.2M existing orders

Person
maya.chen
Team
platform-eng
Harness
Claude Code
Active
1h 18m
Waited on maya
20h 19m
Turns
47 31 fable · 16 local
Tool calls
376
Cost
$11.34 baseline $58.39
Outputs
a1f9c2e · PR #482 · staging
Heartbeatcost so far $8.91 at 21:26 · $11.34 at end · writing to you · idle compressed
21:26:35 · $8.91
Efficiency findings · 3 in this sessionestimates · recommendations never change routing on their own
  1. $3.52 est. avoidableReplay billed above cache rate361,000 tokens re-sent uncached at $10/M after the x-request-id header changed the prefix 61 times; as cache reads they cost $0.25/MF-0398 · high confidence
  2. $1.38 est. avoidableModel and effort mismatch9 of 31 Fable turns were one-line edits; Sonnet 5.5 passed the same turns on replayF-0351 · verifying
  3. $0.36 est. avoidableUnused tool definitions13 of 31 definitions never called (ask_user, skill and 11 smaller) · 263,000 of 641,390 definition tokens re-sent with every requestF-0412 · high confidence
total est. avoidable $5.25 of $11.34 measured
Where the 1h 18m wentidle time excluded · it waited on maya for 20h 19m
  1. writing to you 34m 37s · 44%
  2. composing tool calls 16m 11s · 21%
  3. waiting on the model 13m 5s · 17%
  4. running tools 10m 9s · 13%
  5. delegating 2m 53s · 4%
  6. reasoning 59.0s · 1%
You, the agent, its toolswhole session · calls · time · tokens back · share of cost
you read ×124 · 24.0s 1.67M · $2.22 · 19.6% grep ×21 · 12.1s 167k · $0.22 · 2.0% glob ×12 · 6.8s 55.8k · $0.07 · 0.7% edit ×58 · 10.4s 41.8k · $0.06 · 0.5% write ×6 · 710ms 5.6k · $0.01 · 0.1% bash ×132 · 9m 12s 809k · $1.07 · 9.5% todo ×20 · 2.6s 16.7k · $0.02 · 0.2% delegate ×3 · 2m 53s 19.5k · $0.03 · 0.2% main fable-5-1 writing to you context 38k tokens

Hover to preview · click a tool to filter the tape

The tapeup to the playhead, newest first · scroll for earlier · select a row for its turn
  1. 21:26:00agent → bashbash · 34.2s · 295 chars backSend the same webhook retry to staging twice and confirm the second reuses the stored response.$0.12so far
  2. 21:25:59agentreasoned · 131 tokens$0.12so far
  3. 21:25:56you → agentsent a messageAsked whether webhook retries reuse the original idempotency key.—
  4. 21:14:32agent → editedit · 1.2s · 312 chars backScope the key lookup in idempotency.ex to merchant_id so two merchants can reuse a key.$0.89
  5. 21:14:22agent → readread · 143ms · 6.9k chars backRead payments/handler.ex to find where the key must be checked before capture.$0.89
  6. 21:09:19agent → grepgrep · 0.9s · 2.1k chars backFind every call site that builds an idempotency key before touching the handler.$0.89
  7. 20:01:50you → agentsent a messageRun the suite again and fix whatever the backfill broke.—
  8. 19:59:24agent → writewrite · 0.6s · okWrite the migration that adds idempotency_keys and backfills existing orders in batches of 10k.$0.05
  9. 19:59:18agentreasoned · 904 tokens$0.05
  10. 19:57:31you → agentsent a messageBackfill has to run in batches; 1.2M rows in one transaction will lock the table.—
  11. 19:56:57agent → todotodo · 40ms · 6 itemsTrack the remaining work: migration, backfill, handler change, tests, PR.$0.67
  12. 19:56:48agent → readread · 98ms · 4.1k chars backRead lib/payments/charge.ex to see how captures are committed today.$0.67
  13. 19:56:46agent → globglob · 31ms · 9 paths backList the payments modules before reading any of them.$0.67
  14. 17:55:09test-runner → bashbash · 41.0s · 38 lines backRun the payments suite after the fix; all 12 tests pass.$0.00local
  15. 17:54:57test-runner → editedit · 0.8s · 212 chars backReturn the stored response when a key is seen twice instead of raising.$0.00local
  16. 17:54:46test-runner → readread · 121ms · 3.4k chars backRead the failing test to see what it expects on a duplicate key.$0.00local
  17. 17:54:39test-runner → bashbash · 39.4s · 38 lines backRun the payments suite; one idempotency test fails.$0.00local
  18. 17:53:08agent → delegatedelegate · spawned test-runnerRun the payments suite and fix the failing idempotency test.$0.75
sess_4e1b · 5,308,152 tokens analysed · $11.34 measuredestimated from size measured exactly
Bars scaledcost = each bucket’s own billed mix · session input averages $1.56/M
all buckets5,308,152100%$6.40
system prompt1,364,81325.7%$0.61
tool definitions641,39012.1%$0.87
tool results2,788,97252.5%$3.70
user27,3630.5%$0.21
assistant text149,7322.8%80.7% measured exactly$0.36
tool calls335,8826.3%9.5% measured exactly$0.64
replay224,183billed beyond analysed, all uncached (224,183×$10/1M)$2.24
output61,40054,000 Fable ×$50/1M · 7,400 local $0.00$2.70
measured5,593,735$6.40 buckets + $2.24 replay + $2.70 output$11.34
Conversation tree · frame width = tokens47 turns · 31 main on fable-5-1 · 16 test-runner on local qwen3-coder
turn 1

turn 1 · 1,362,598 · 25.7% · $1.75 · Read the payments handlers and planned the idempotency-key migrationnot in thread 793,730 · test-runner ×16 500,000 · $0.00 local (cloud equiv $0.12)

billed input 5,532,335 · analysed 5,308,152 · $11.34 measured · baseline $58.39 all-cloud, no cache

FIG. 1 — One session, replayed

This session is PR #482, one tile in platform-eng's month.

See what leaders see ↗

02Routing

Routing you understand and control.

Not an auto button. You pick the strategy, set a policy per team or agent cluster, and every decision is written down: the candidates, the choice, the reason.

Policy

profile eng/coding · strategy: capability floor, then lowest cost · shadow: off · session pinned for cache continuity

A
Request

“Plan the idempotency-key refactor across the payments handlers”

ixway reads
intent
refactor
complexity
high
files
6
context
38k tokens
priority
normal
max cost
$0.60
Candidates · route
  1. claude-fable-5-1chosen anthropic · cloud · est $0.1934 · p50 9.1 s meets capability floor The only candidate at or above the floor for a high-complexity refactor. $0.1934 fits the $0.60 cap; cache reads keep it under half the $0.4750 uncached cost.
  2. claude-opus-5-5 anthropic · cloud · est $0.0803 below floor for this task $0.1131 cheaper, but it scores under the floor for multi-file refactors. Cost only breaks ties above the floor.
  3. claude-sonnet-5-5 anthropic · cloud · est $0.0401 below floor for this task Cheaper again, and further under the floor for a change that spans six handlers.
  4. qwen3-coder-480b vllm · rack-a · local · $0.00 · cloud equiv. $0.0103 below floor for this task Free on the local rack, but under the floor here. It clears the floor for simpler work: see B.

x-ixway-resolution-reason: capability_floor_cheapest

baseline, all on fable, no cache: $0.4750

B
Request

“Run the payments suite and fix the failing idempotency test”

ixway reads
intent
test-fix
complexity
low
files
2
context
9k tokens
priority
normal
max cost
$0.60
Candidates · route
  1. qwen3-coder-480bchosen vllm · rack-a · local · $0.00 · p50 2.8 s meets floor · local first · cloud equiv. $0.0026 Clears the floor for a low-complexity test fix, so the local rack takes it: $0.00 billed, $0.0026 at the hosted list price.
  2. claude-haiku-5-5 anthropic · cloud · est $0.0012 · p50 2.1 s eligible · local scored higher The cheapest and fastest cloud option, but it scored below qwen3-coder-480b on test fixes, and local bills nothing.
  3. claude-sonnet-5-5 anthropic · cloud · est $0.0244 eligible · unnecessary for this task Clears the floor, but nothing here needs it. $0.0244 for the same outcome.
  4. claude-fable-5-1 anthropic · cloud · est $0.1220 eligible · unnecessary for this task The most expensive way to fix one test in two files: $0.1220 for what the local rack does at $0.00.

x-ixway-resolution-reason: local_first_meets_floor

list prices · fable 5.1 $10/$50 · opus 5.5 $4/$20 · sonnet 5.5 $2/$10 · haiku 5.5 $0.10/$0.50 · qwen3-coder-480b $0.22/$1.00 per M tokens in/out

FIG. 2 — Two requests, two routes
  1. 2.1Policies per team. Profiles bind to an organization, a unit, a project, or a single agent.
  2. 2.2Strategies you choose. Lowest cost, lowest latency, capability-gated, peak and off-peak economics, classifier, escalate on a weak answer, or a judge trained on your own traffic.
  3. 2.3Shadow first. A new policy records what it would have done until you switch it on. Sessions stay pinned so caches keep hitting.
  4. 2.4Explained in the response. Resolved model, backend, reason, and decision id come back as headers. Replay any policy offline before you ship it.

02.1Local inference

Get everything out of the hardware you own.

When local can handle the request, local gets it. ixway balances your fleet, keeps priority work inside its SLA, and diverts to cloud only under pressure.

Fleet · last 60 s 3 engine groups · 7 replicas · latency SLO 6,000 ms
homelab · llama.cppqwen3-coder-30b-a3b1 replica
load 18%
p50 1.1 s
0 queued
12 req · $0.00cloud equiv. $0.0043
rack-a · vLLMqwen3-coder-480b-a35b4 replicas
load 62%
p50 2.8 s
3 queued
206 req · $0.00cloud equiv. $0.6427
rack-a · SGLangglm-4.5-air2 replicas
load 91%
p50 7.9 s · over SLO
14 queued
p=high ×3
dispatched first
39 req · $0.00cloud equiv. $0.0881+22 sent to cloud
22 of 61 divertedglm-4.5-air over SLO
cloudanthropicfallback & frontier
+22 from rack-a · SGLang
→ haiku 5.5 · $0.0242
97 req · $3.804822 diverted · 75 by policyfable 15 · sonnet 40haiku 42

local 257 req · $0.00 (cloud equiv. $0.7351) · cloud 97 req · $3.8048

FIG. 3 — Local first, cloud when it must
  1. 2.5Real engine signals. vLLM, SGLang, llama.cpp, TensorRT-LLM, Ollama. Queue depth and capacity read straight from the engine.
  2. 2.6Priority dispatch. Tag work with a priority and a latency SLO. The important traffic keeps its place in the queue.
  3. 2.7Honest economics. Every local request shows its cloud-equivalent cost. You finally know what the rack is worth.

03Efficiency insights

Find the waste. Hand the fix to the harness.

ixway produces findings the way a scanner produces vulnerabilities: specific, evidence-backed, priced. Your harnesses read them and get leaner.

Replay billed above cache rate a volatile header broke the cache prefix in 61 sessions; 22M tokens billed at $10/M, not $0.25/M $215 high
Model and effort mismatch a frontier model on 720 one-line edits; Sonnet 5.5 passed on replay $110 verifying
Bloated tool output a 9 MB log read into context 14 times $96 high
Unused tool definitions 13 of 31 definitions never called across 200 sessions, re-sent every request $72 high
Fan-out duplicate work 4 sibling agents read the same 3 files, 32 times $41 medium
5 open · platform-eng · last 30 days · $2,418 mtdest. avoidable $534 / mo · verified this month $140
recordevery request, in the ledger findpriced, with evidence changethe harness applies the fix verifyon comparable work quotes · receipts compaction advice · findings
FIG. 4 — Findings, ranked by what they cost you
  1. 3.1Findings, not advice. Each one links to the requests, tool calls, and sessions where it happened.
  2. 3.2Three honest numbers. Measured spend, estimated avoidable spend, and verified savings after the fix. Never one pretending to be another.
  3. 3.3A loop, not a report. Quotes before a request, receipts after, compaction advice when context bloats, findings back to the harness.
  4. 3.4Verified after. Accept a fix, re-run comparable work, and see whether the saving was real.

Redefine tokenmaxxing. Maximize value per token.

04System of record

A system of record your harnesses learn from.

The ledger is the product. Every view above reads from it. So do your harnesses.

  • 4.1

    Tamper-evident ledger

    Every request is a hash-chained row. Verify the chain any time. Reconcile it to provider invoices at period close.

  • 4.2

    Ask it anything

    A query builder with CSV export, and an assistant that answers with cited queries, routed through your own gateway.

  • 4.3

    Outputs, not requests

    Sessions are classified by activity and outcome and linked to the commits, pull requests, and deployments they produced.

  • 4.4

    Application traces, joined

    Send OpenTelemetry traces from your own agents. ixway joins them to the ledger, so a workflow run shows its cost per outcome.

  • 4.5

    Budgets that bite

    Daily, weekly, and monthly budgets per key. Rate limits, priority queues, and batch APIs for work that can wait.

  • 4.6

    Privacy by default

    Capture is optional. Individual reporting hides behind minimum group sizes. Content never leaves your deployment unless you say so.

05How it works

Point your inference at ixway. Keep everything else.

ixway is a gateway. Harnesses and SDKs talk to it exactly as they talk to a provider. It speaks the Anthropic and OpenAI dialects natively and relays them byte for byte.

harnesses & sdks

  • Lemieux · lmx
  • Claude Code
  • Codex
  • Anthropic SDK
  • OpenAI SDK
  • LangChain · LangGraph
  • OpenAI Agents
  • OpenInference

gateway

recordsroutesenforcesfinds waste
passthrough
Your Claude or ChatGPT login goes straight through. ixway holds no credential.
gateway keys
Provider keys live in ixway. People and agents get virtual keys with budgets and limits.
telemetry
Your own agents send traces on a separate credential. Nothing can spend with it.

models

  • Anthropic
  • OpenAI
  • 30+ cloud providers
  • vLLM
  • SGLang
  • llama.cpp
  • TensorRT-LLM
  • Ollama

local engines

FIG. 5 — One base URL in, every model out

Your AI.
A clearer picture.

We’ll let you know when Ixway is ready.

Have a question? hello@ixway.ai

Join the waitlist

We’ll let you know when Ixway is ready.

How would you use ixway? Optional

By joining, you agree to receive emails from Ixway Labs, Inc about the ixway waitlist and early access. Unsubscribe at any time. Privacy notice.