MCP · Live For enterprise AI, platform and FinOps teams

Cost and governance engines for every AI workflow.

Know the cost of each workflow step before you ship
Cut AI spend with routing, caching and caps that keep quality
Turn rules into policy your gateway and cloud enforce
Prove it daily: modeled versus actual, with an audit trail

Your workflow→AICost engines→Your gateway and cloud→Daily proof

124cost and governance engines

8workflow tools

0LLM calls: same answer every time

Get a free MCP key See enterprise workflows

Works with Claude, Codex, Cursor, CI, LiteLLM, Amazon Bedrock, Azure AI Foundry, Vertex AI and vendor APIs · https://mcp.aicost.ai/mcp

Six enterprise jobs you can run today

The problem, the engines that solve it, the outcome.

Legend: registry engine workflow tool platform capability quality gate your system

🔎

See where every AI dollar goes

One invoice line per vendor; no idea which team, product, or agent spent it.

A platform lead asks Claude, in the middle of standup:

"Break down last month's AI spend across our three products and flag anything growing faster than usage."

Your bills / usage export → Ledger ingest (file inbox / LiteLLM pull) → business-bill-diagnose → analyze_workload_portfolio → observability-cost

Outcome

Spend attributed to workload and team, leakage and concentration flagged, and the cost of keeping that attribution running priced honestly.

In Claude Code

claude mcp add --transport http aicost https://mcp.aicost.ai/mcp # then just ask: "diagnose this AI bill" and paste the numbers

📉

Optimize and reduce - without double counting

Routing 45% plus caching 30% is not 75%. Vendor decks sum levers that overlap.

An AI engineer, before the quarterly cost review:

"Stack routing, caching, batching and token reduction on our support workload. Give me the REAL combined number and the gateway config to deploy it."

multi-model-router → prompt-cache-roi → batch-vs-realtime → eval-benchmark-cost → stack-savings → policy_patch -> LiteLLM / Cloudflare export

Outcome

A stacked-savings number that survives finance review, the eval budget that proves quality held, and an executable gateway config - not a slide.

Every engine returns

"policy_patch": { "compiled": { "target": "litellm", "config": { ... } } }

📈

Plan and forecast in ranges, not points

Teams that forecast cloud within 1-3% miss AI by 2-3x. Agentic spend compounds.

FP&A, building next year's plan:

"Model this workload at 1x / 10x / 100x, re-forecast the plan year with the agentic share compounding, and size the reserve."

estimate_llm_workload_cost → scale-projection → agentic-variance-reserve → budget-planner

Outcome

A range forecast with a breach month, a 20-40% reserve sized to your agentic mix, and alert thresholds instead of hard stops.

REST, from any script

POST /v1/run-batch Authorization: Bearer aicost_sk_... {"runs":[{"engine_id":"aicost.scale-projection","inputs":{...}}]}

⚡

Enforce it dynamically - caps that actually fire

AICost CostWall

A budget in a spreadsheet stops nothing. One runaway agent loop can burn a five-figure day.

The same policy, now live in the request path and in CI:

"Soft-cap downgrades to the economy tier, hard cap blocks at the daily limit, and every PR that raises cost per task fails the build."

cost-firewall → Firewall proxy: 429 rate / routed_to_economy / 402 hard cap / 403 kill switch → CI/CD cost gate (/v1/run in the pipeline) → Hourly anomaly + budget alerts

Outcome

Four enforcement tiers proven on live traffic, cost regressions caught pre-merge, and alerts before the bill becomes a surprise.

What a blocked call returns

HTTP 402 { "error": "daily_hard_cap", "resets_at": "..." }

📒

Audit it in the ledger - showback that holds up

AICost CostProof

Forecasts nobody reconciles are theater. Finance wants actuals, attribution, and an export their tooling reads.

Month-end, without a data pull request to engineering:

"Show forecast vs actual per workload, attribution coverage, and give me the FOCUS export for the FinOps platform."

aicost_ledger_actual (MySQL, per-tenant metering) → Forecast-vs-actual reconcile (accuracy % reported on /v1/status) → FOCUS 1.x CSV export → Hash-chained audit trail (aicost_mcp_audit)

Outcome

Showback by tenant and workload, forecast accuracy you can quote, an open-standard export, and an audit trail where a silently altered row is a detected security event - not a shrug.

The buyer's uptime question

GET /v1/status -> { ledger: { staleness_h, attribution_coverage_pct, forecast_accuracy } }

🛡️

Continuous governance - beyond the assessment PDF

Every consultancy sells a point-in-time assessment. It is stale the week a model, a price, or a rule changes.

The AICostAdvisory difference, running as engines:

"Keep our AI inventory, per-route eligibility, and governance cost current - and re-issue the evidence whenever anything moves."

msp-client-ai-exposure-audit → Model eligibility: 6 gates, per ROUTE not per model → compliance-cost → jurisdiction-risk-scorer → Policy wizard -> gateway export + hash-chained evidence

Outcome

A governance posture that re-computes itself: same engines the assessment used, callable nightly from your own automation, with tamper-evident history. The PDF becomes a snapshot of a living system.

Deterministic by design

No LLM at runtime. Same inputs, same outputs, every run - auditable.

Governance that stays current

Re-run on any change: price, model, route or rule
From your own automation, on your schedule
Hash-chained audit trail shows nothing was edited in between

See the 20 governance modules →

Mapped to how your team already works

Every audience runs the same engines in a different order. The full chain catalog powers this page, the web calculators, and the MCP server from one source of truth.

Developers & AI Engineers 3 ready-made chains Price the feature before you build it

Enterprise FinOps / CFO 3 ready-made chains A forecast with teeth (range, not a point)

MSPs (MSPCost.com) 3 ready-made chains Stop the leak before invoices ship

Healthcare (HealthITCost.com) 4 ready-made chains Ambient scribe: cost per encounter vs time returned

Governance & Advisory (AICostAdvisory.com) 3 ready-made chains AI Cost & Governance Assessment

VC / PE Diligence 1 ready-made chain Cost-claim diligence in a deal window

SMB & Solo Builders 1 ready-made chain Pick the builder stack without overbuying

Consumers & Prosumers 1 ready-made chain Stop paying for three assistants

The productized layers on these rails

The engines model and decide for free. When the workload is live, two of the rails above become products: the wall that enforces, and the ledger that proves. Names and claims mirror the pricing page - one source for every price.

AICost CostWall

Blocks the waste, lets the work through. Budgets, daily caps, spike anomaly, kill switch - enforced in your gateway.

Runs on: Enforce it dynamically - caps that actually fire See plans →

AICost CostProof

Prove it. Persistent ledger, modeled vs actual, realized savings.

Runs on: Audit it in the ledger - showback that holds up See plans →

AICost Scale

The cloud bill under the model bill, too. AWS CUR, Azure and BigQuery ingest, attribution by feature, customer and team, and a CI/CD cost gate that fails the build over budget.

Runs on: See where every AI dollar goes See plans →

Deploy into the gateway you already run

We are the vendor-neutral decision layer that runs BEFORE the routing layer: decide which models are permitted, what they cost, and what the caps are - then export that as configuration into whichever gateway your team already operates. Policy engines emit an executable policy_patch; anything a target cannot express natively is preserved under x_aicost_policy so nothing is silently dropped. Actuals flow back through the LiteLLM connector, the billing file-inbox, or your FOCUS 1.x export. Comparing them first? The router-compare engine holds the verified capability and pricing capture.

LiteLLM Self-hosted gateway native policy export actuals pull

policy_patch compiles to model_list, fallbacks, budgets, cache and retry config. The Ledger pulls spend back from its database, so forecast-vs-actual closes automatically.

Cloudflare AI Gateway Edge gateway native policy export

policy_patch compiles to gateway config: caching, rate limits, fallback order. Their edge guardrails and DLP run in front; our caps and routing policy ride in the same config.

Portkey Gateway (open source) native policy export

Policy export to gateway config with guardrails at every tier on their side; our eligibility gates decide WHICH models the config may contain.

Kong Enterprise API gateway native policy export

Policy export for teams already standardized on Kong: budgets and caps expressed as gateway plugins alongside existing API governance.

OpenRouter Gateway + router

We run BEFORE it: eligibility and budget policy decide the permitted model set; OpenRouter moves the traffic. Billing exports drop into the file-inbox for Ledger reconcile.

Vercel AI Gateway Gateway

Routing rules and budgets configured from our policy output; pass-through billing exports feed the Ledger the same way.

Helicone Gateway + observability

Their traces answer WHAT happened per request; our envelope answers WHAT IT SHOULD COST and why. Cost exports reconcile into the Ledger.

Not Diamond Router only

A routing brain that sits on top of an existing gateway. We qualify the model set it is allowed to route across before it ever picks.

Use it today - four doors, one fleet

Dual auth on one header: OAuth for humans in assistants, bearer keys for apps and pipelines. Nothing is degraded by using a key.

🤖 Claude Code / Desktop

claude mcp add --transport http \ aicost https://mcp.aicost.ai/mcp # sign in with OAuth, or use your key

⌨️ Codex CLI / IDE (~/.codex/config.toml)

[mcp_servers.aicost] url = "https://mcp.aicost.ai/mcp" bearer_token_env_var = "AICOST_API_KEY" # then /mcp in a session to verify

🐍 Your app (Python, Node, anything)

POST https://mcp.aicost.ai/mcp Authorization: Bearer aicost_sk_... Accept: application/json, text/event-stream

🔁 CI / nightly jobs (REST)

POST /v1/run-batch {"runs":[{"engine_id":"aicost.cost-firewall", "inputs":{...}}]}

🗂️ Browse the catalog first

GET https://mcp.aicost.ai/v1/engines GET https://mcp.aicost.ai/v1/health

Integrate into your workflows: AWS, Azure, GCP, vendor APIs and gateways

Same answer in every scenario: your model calls never route through us - the MCP layer is out-of-band advisory and audit. What you get in each case is the same pair of outcomes: DAILY cloud + AI cost visibility and daily optimization recommendations from the engines in your pipeline.

You run on AWS Bedrock

Your app→Bedrock runtime→AWS bill (CUR)

AICost MCP advises your CI, agents and reviews - never in the request path

DAILY costs: Turn on Bedrock model invocation logging: token-level JSONL lands in S3 near-real-time. Drop it (gunzipped) into the actuals inbox as bedrock-*.jsonl - daily, per-model, per-agent rows, no CUR wait.

On the showback dashboard, daily: $ / dayby modelby workloadby agenttokens in/out + cachemodeled vs billed drift

Full integration guide →

You run on Azure AI Foundry / Azure OpenAI

Your app→Foundry deployment→Azure Cost Management

Same out-of-band pattern: engines advise, Azure serves

DAILY costs: Schedule the Azure Cost Management daily export to storage and forward it to the inbox; Azure Monitor token metrics per deployment fill the intraday view. PTU showback is allocated by utilization share, which the engines already model.

On the showback dashboard, daily: $ / dayby deploymentby workloadby agentPTU utilization sharePAYG vs PTU split

Full integration guide →

You call vendor APIs directly (Anthropic, OpenAI, Google, Mistral...)

Your app→Vendor API→Vendor invoice

Richest optimization surface: caching, batch and routing engines apply at list prices

DAILY costs: Two real-time options: emit the usage block your app already receives on every response into the modeled lane, or pull the vendor usage API daily into the inbox. Either way you never wait for the invoice.

On the showback dashboard, daily: $ / dayby modelby workloadby agentcache + batch savings realizedper-key attribution

Full integration guide →

You run through a gateway (LiteLLM, Portkey, Helicone...)

Your app→Gateway→Vendor APIs→Vendor invoices

The fully automated path: policy in, actuals back - still zero inference through us

DAILY costs: Real-time out of the box: the connector pulls continuously - no exports, no cron, no waiting.

On the showback dashboard, daily: $ / dayby model + providerby team + keyby workloadby agentrouting savings realized

You run on GCP Vertex AI

Your app→Vertex AI endpoint→Cloud Billing

Gemini and Model Garden priced before you ship - and daily actuals from the BigQuery billing export

DAILY costs: Daily SKU-level rows via the BigQuery export; real-time via the modeled lane when your app emits usage metadata. You never wait for the invoice.

On the showback dashboard, daily: $ / dayby modelby workloadby agentlabels honoredmodeled vs billed drift

Full integration guide →

Mixed estate (most enterprises)? One tag schema, one policy layer, one modeled book, one FOCUS 1.x export - the dashboard shows the whole estate per tenant, workload and agent regardless of where each piece runs.

Run it on your own workflow

Free key in minutes. Or we price and govern one workload with you.

Get a free MCP key → Test a workload with us Technical guide

Vendor / Model	Field	Why it’s inferred
Anthropic — Claude Sonnet 4.6	`cachedInput`	Derived at 10% of input rate — Anthropic publishes 90% cache-hit discount on this tier.
Anthropic — Claude Sonnet 4.5	`cachedInput`	Derived at 10% of input rate; same 90% cache-hit convention as Sonnet 4.6.
Anthropic — Claude Sonnet 4.5	`batchInput`	Derived at 50% of standard input — Anthropic documents uniform 50% Batch discount.
Anthropic — Claude Sonnet 4.5	`batchOutput`	Derived at 50% of standard output — Anthropic documents uniform 50% Batch discount.
Anthropic — Claude Haiku 4.5	`cachedInput`	Derived at 10% of input rate — Anthropic 90% cache-hit discount convention.
OpenAI — GPT-5.4 Mini	`cachedInput`	Derived at 10% of input — OpenAI documents automatic 90% discount on cache hits across GPT-5.x tier.
OpenAI — GPT-5.4 Nano	`cachedInput`	Derived at 10% of input — OpenAI 90% cache-hit convention.
OpenAI — GPT-5.4 Nano	`batchInput`	Derived at 50% of input — OpenAI Batch API uniform 50% discount.
OpenAI — GPT-5.4 Nano	`batchOutput`	Derived at 50% of output — OpenAI Batch API uniform 50% discount.
OpenAI — GPT-5.4 Pro	`cachedInput`	Derived at 10% of input — OpenAI 90% cache-hit convention.
OpenAI — GPT-5.4 Pro	`batchInput`	Derived at 50% of input — OpenAI Batch API uniform 50% discount.
OpenAI — GPT-5.4 Pro	`batchOutput`	Derived at 50% of output — OpenAI Batch API uniform 50% discount.
OpenAI — GPT-5.2	`cachedInput`	Derived at 10% of input; no residency uplift.
OpenAI — GPT-5.2	`batchInput`	Derived at 50% of input.
OpenAI — GPT-5.2	`batchOutput`	Derived at 50% of output.
OpenAI — GPT-5	`cachedInput`	Derived at 10% of input.
OpenAI — GPT-5	`batchInput`	Derived at 50% of input.
OpenAI — GPT-5	`batchOutput`	Derived at 50% of output.
OpenAI — GPT-5.5 Pro	`cachedInput`	Derived at 10% of input — OpenAI does not publish a cached rate for *-pro models; using the family convention.
OpenAI — GPT-5.5 Pro	`batchInput`	Derived at 50% of input.
OpenAI — GPT-5.5 Pro	`batchOutput`	Derived at 50% of output.
OpenAI — GPT-5.2 Pro	`cachedInput`	Derived at 10% of input — pro-tier convention.
OpenAI — GPT-5.2 Pro	`batchInput`	Derived at 50% of input.
OpenAI — GPT-5.2 Pro	`batchOutput`	Derived at 50% of output.
OpenAI — GPT-5.1	`batchInput`	Derived at 50% of input.
OpenAI — GPT-5.1	`batchOutput`	Derived at 50% of output.
OpenAI — GPT-5 Pro	`batchInput`	Derived at 50% of input.
OpenAI — GPT-5 Pro	`batchOutput`	Derived at 50% of output.
OpenAI — GPT-5 Nano	`cachedInput`	Derived at 10% of input.
OpenAI — GPT-5 Nano	`batchInput`	Derived at 50% of input.
OpenAI — GPT-5 Nano	`batchOutput`	Derived at 50% of output.
Google — Gemini 3 Flash	`cachedInput`	Derived at 10% of input — Google caching discount convention ~90%.
Google — Gemini 3.1 Flash-Lite	`cachedInput`	Derived at 10% of input — Google caching convention.
Google — Gemini 3.1 Flash-Lite	`batchInput`	Derived at 50% of input — Google Batch API uniform 50% discount.
Google — Gemini 3.1 Flash-Lite	`batchOutput`	Derived at 50% of output — Google Batch API uniform 50% discount.
Google — Gemini 2.5 Pro	`cachedInput`	Derived at 10% of input.
Google — Gemini 2.5 Flash	`cachedInput`	Derived at 10% of input.
Google — Gemini 2.5 Flash-Lite	`cachedInput`	Derived at 10% of input — Google caching convention.
Google — Gemini 2.5 Flash-Lite	`batchInput`	Derived at 50% of input — Google Batch API uniform 50% discount.
Google — Gemini 2.5 Flash-Lite	`batchOutput`	Derived at 50% of output — Google Batch API uniform 50% discount.
xAI — Grok 4 (legacy)	`cachedInput`	Extrapolated at 25% of base.

Cost and governance engines for every AI workflow.

Six enterprise jobs you can run today

See where every AI dollar goes

Optimize and reduce - without double counting

Plan and forecast in ranges, not points

Enforce it dynamically - caps that actually fire

Audit it in the ledger - showback that holds up

Continuous governance - beyond the assessment PDF

Governance that stays current

Mapped to how your team already works

The productized layers on these rails

Deploy into the gateway you already run

Use it today - four doors, one fleet

🤖 Claude Code / Desktop

⌨️ Codex CLI / IDE (~/.codex/config.toml)

🐍 Your app (Python, Node, anything)

🔁 CI / nightly jobs (REST)

🗂️ Browse the catalog first

Integrate into your workflows: AWS, Azure, GCP, vendor APIs and gateways

Run it on your own workflow

Methodology

Primary sources

Inferred values (marked with * in calculator tables)

Immediate steps you can take