LLM Pricing

Analyze one specific session's actual recorded LLM usage (tagged via record_usage's session_id) on demand: total input/output tokens, total actual cost, and a per-model breakdown if more than one model was used. Also runs the session's own observed token pattern through the same routing engine behind recommend_model / POST /router/recommend to produce a grounded recommendation — or confirm the current choice is already optimal — with the estimated dollar savings or increase versus this session's actual cost. Returns has_data: false with no recommendation if the session_id has no recorded usage. Advisory only — never switches or reroutes anything. Requires your service API key on the MCP connection (x-api-key or Authorization: Bearer); org_id defaults to, and must match, that key's org.

Ask the LLM Pricing AI agent a question in natural language. Use for complex queries about pricing, model recommendations, cost optimisation advice, or anything that benefits from multi-step reasoning.

Audit this server's own pricing data against an external reference registry. Prices here are hand-maintained and go stale silently; a price that disagrees with the registry by more than the threshold is automatically withheld from serving the moment it's detected, not just when this tool runs. This reports what's currently withheld and why. The price value itself is never overwritten by the registry. Use when the user asks whether the pricing data is accurate or up to date, or wants to audit a specific provider's rates.

Compare costs for multiple LLM models. Every named model is priced; pass client (e.g. 'claude_code') to also flag each model runnable_in_client (it never filters here). Recognised clients: claude_code, claude_desktop, cursor, github_copilot, windsurf, cline, continue, zed, jetbrains_ai, gemini_cli, codex_cli, aider, amazon_q, chatgpt_desktop.

Delete one of your org's registered budget-spend alerts by its ID. Requires your service API key on the MCP connection (x-api-key or Authorization: Bearer); another org's alert ID reads as not found.

Delete a specific chat conversation by its ID. Use when the user asks to delete, clear, or remove a past conversation. Use list_conversations first to find the conversation ID. Operator only: needs the server operator's key (MCP_API_KEY) on the MCP connection (x-api-key or Authorization: Bearer); any other caller, including the chat agent and the local STDIO server, gets error_code operator_key_required.

Delete a registered price-change alert by its ID. Operator only: needs the server operator's key (MCP_API_KEY) on the MCP connection (x-api-key or Authorization: Bearer); any other caller, including the chat agent and the local STDIO server, gets error_code operator_key_required. To view or delete an alert you registered yourself, use manage_price_alert with the alert_token from registration.

Send a text chat request through the dispatch gateway: it applies your policy, runs the best eligible model, and returns the completion, usage and the full dispatch_receipt (chosen model, reason, cost, savings, attempts) — the same as POST /v1/chat/completions with a `dispatch` object. This makes a REAL model call billed to your own provider keys stored for your org (POST /dispatch/keys), and the spend is recorded against your org. Requires your service API key on the MCP connection (x-api-key or Authorization: Bearer); the operator key and anonymous callers get dispatch_requires_api_key, and the local STDIO server refuses with dispatch_requires_hosted_mcp. Pass prompt (with optional system) or messages, not both. Text-only, single-choice, non-streaming; no tools/functions. Refusals carry the same error codes as HTTP. Calls can take a while with failover: bound them with dispatch.failover.timeout_ms and dispatch.failover.max_attempts, and set LLM_PRICING_MCP_TIMEOUT_MS=120000 when using the npx bridge.

Estimate the cost of using a specific LLM model

Get current pricing for all LLM models from all providers

Report process-lifetime hit/miss counts, total lookups, and hit rate for the ~5-minute recommendation cache, separately for the recommend_model tool and POST /router/recommend (never combined). Server-operations introspection for confirming the cache is reducing recomputation in production, not a pricing question. Counts reset on process restart; reading them never changes them.

Report an aggregate accuracy summary for the pricing catalogue: what percentage of models have a confirmed, fresh price, how many are currently withheld due to price drift, how many are newly discovered and not yet priced, and how many confirmed prices are over 90 days old. A single trust signal, distinct from check_price_drift which lists individual disputes. Use when the user asks how accurate or trustworthy the pricing data is, overall.

Get subscription pricing for AI coding IDE tools: GitHub Copilot, Cursor, Windsurf, Claude.ai, JetBrains AI, Amazon Q Developer, and Tabnine. Returns monthly subscription cost, inline completion support, context window, and per-tool strengths. Use this when the user asks to compare IDE AI tools, subscription-based coding assistants, or wants to know which IDE AI fits their budget.

Get performance metrics (throughput, latency, context window) for models

Generate a download URL for the pricing history export. Returns a URL the user can open in their browser to download snapshot history as CSV (for spreadsheets) or JSON (for code). Use this when the user asks to export, download, or save pricing data.

Query historical pricing snapshots recorded over time. Use this to look up past prices for a model or provider, or to see how prices have changed over a given number of days.

Find models whose prices changed the most over a given period. Returns models sorted by absolute percentage price change with a direction label ('increased', 'decreased', or 'unchanged'). Use this to answer questions like 'which models got cheaper this month?' or 'which providers raised prices recently?'

Get MCP server telemetry and usage statistics including tool usage, response times, and error rates

Get a summary of actual recorded LLM spend: total cost, request count, and breakdowns by model and provider, from usage events previously logged via record_usage. Use when the user asks what they've actually spent, as opposed to a hypothetical cost estimate. Requires your service API key on the MCP connection (x-api-key or Authorization: Bearer); org_id defaults to, and must match, that key's org.

Get recommended use cases and strengths for LLM models

Bulk-import a completed session's usage in a single call — the retroactive counterpart to record_usage, for when a session wasn't instrumented live. Use when a calling agent (e.g. VS Code Copilot Chat) already has a session's turn-by-turn history in its own local store and wants this server to report on it exactly as if it had been tracked live. Accepts a list of entries under one shared session_id; each entry resolves its input and output tokens independently, either from exact counts when known or by estimating from raw turn text using this server's existing token counter. Cost is always computed server-side from current pricing, never trusted from the caller. A malformed entry (no usable tokens/text on one side, or an unresolvable model_name) is reported as its own per-entry failure and does not fail the rest of the batch. Idempotent like record_usage — resubmitting the same entries (same request_id, or the same session_id/model/occurred_at/token-or-text when request_id is omitted) is a no-op. Recorded usage is then readable via analyze_session_usage and GET /usage/session/{session_id}. Requires your service API key on the MCP connection (x-api-key or Authorization: Bearer); org_id defaults to, and must match, that key's org. POST /mcp request bodies are capped at 1,000,000 bytes (413 above that). 500 exact-count entries fit comfortably (~150 KB); text-mode imports of long sessions can exceed it, so split the entries across several calls under the same session_id — imports are idempotent, so overlapping or retried batches are safe.

List your org's registered budget-spend webhook alerts with their IDs and settings. Requires your service API key on the MCP connection (x-api-key or Authorization: Bearer); only alerts for that key's org are returned.

List stored chat conversation sessions with their IDs, last-updated timestamp, turn count, and a preview of the most recent user message. Use this when the user asks to see their past conversations or chat history. Operator only: needs the server operator's key (MCP_API_KEY) on the MCP connection (x-api-key or Authorization: Bearer); any other caller, including the chat agent and the local STDIO server, gets error_code operator_key_required.

List all registered price-change webhook alerts with their IDs and settings. Operator only: needs the server operator's key (MCP_API_KEY) on the MCP connection (x-api-key or Authorization: Bearer); any other caller, including the chat agent and the local STDIO server, gets error_code operator_key_required.

View or delete the one price-change alert you registered, using the alert_token that register_price_alert returned once. No API key needed: the alert_token is the credential and is scoped to exactly that alert. action 'view' returns its id, url, threshold_pct, provider, model_name and created_at; action 'delete' stops it. A token that matches nothing gets the same invalid_alert_token refusal whether or not any alert exists. Alerts registered before v1.81.0 have no token and are removed via the operator.

Optimise a multi-task LLM workload by assigning the cheapest qualifying model to each task instead of running everything through one model. Takes a list of workloads (task type + monthly volume + optional quality floor and capability requirements) and returns a per-task model allocation, the total monthly cost, and the saving versus the best single-model deployment. If a monthly budget is given, leftover headroom is spent on the most cost-efficient quality upgrades. Use when the user asks how to cut LLM spend across several task types, or whether tiered model routing is worth it. Pass client (e.g. 'claude_code', 'cursor', 'github_copilot', 'cline') so the answer is restricted to models that tool can run; omit it and the hosted server tries to infer it from the MCP handshake, otherwise every priceable model is considered. An explicit client always wins. Recognised clients: claude_code, claude_desktop, cursor, github_copilot, windsurf, cline, continue, zed, jetbrains_ai, gemini_cli, codex_cli, aider, amazon_q, chatgpt_desktop.

Predict the real cost of an LLM call from your actual prompt text — no need to know token counts in advance. Counts tokens locally, infers the task type (or accepts one explicitly), estimates output length from empirical I/O profiles, and returns all models ranked cheapest-first with effective cost after prompt-cache savings. Also surfaces the best quality-per-dollar pick. Use this before making any LLM API call to avoid surprise bills.

Recommend the optimal LLM model from a free-text description of your use case — no need to already know the routing engine's task_type vocabulary. Classifies the description into a task type, then runs it through the same routing engine behind /router/recommend (live pricing + benchmark quality, hard filters for cost/quality/context window, scored by quality-per-dollar). Returns the best match plus up to 3 alternatives. Read-only — use predict_cost first if you also want per-model cost estimates for an actual prompt, and optimize_workload if you have several distinct task types to route at once. Restricts (not just boosts) the candidate set to models your client can run before ranking. Pass client (e.g. 'claude_code', 'cursor', 'github_copilot', 'cline') so the answer is restricted to models that tool can run; omit it and the hosted server tries to infer it from the MCP handshake, otherwise every priceable model is considered. An explicit client always wins. Recognised clients: claude_code, claude_desktop, cursor, github_copilot, windsurf, cline, continue, zed, jetbrains_ai, gemini_cli, codex_cli, aider, amazon_q, chatgpt_desktop.

Record actual token usage for a completed LLM call. Cost is computed server-side from current pricing, never trusted from the caller — this is the 'actual spend' counterpart to estimate_cost's hypothetical numbers. Usage is attributed to your org. Resubmitting the same request_id for the same org is a no-op. Use when the user wants to log or track real LLM spend, not just estimate it. Requires your service API key on the MCP connection (x-api-key or Authorization: Bearer); org_id defaults to, and must match, that key's org.

Register a webhook URL to receive notifications when actual recorded spend (from record_usage) crosses a USD threshold within a rolling window. The alert monitors your org's spend. Fires at most once per period_days while spend stays over threshold. Returns the alert ID. Requires your service API key on the MCP connection (x-api-key or Authorization: Bearer); org_id defaults to, and must match, that key's org.

Register a webhook URL to receive notifications when a model's price changes by more than a specified percentage. Optionally scope the alert to a specific provider or model. Returns the alert ID and a one-time alert_token needed to view or delete it via manage_price_alert.

Ask the customer support assistant about THIS service: plans and prices, signing up, upgrading or cancelling, invoices, service API keys (rotation, lost key), rate limits, connecting over MCP or REST, data accuracy, and how to reach the owner. Answers come only from the published FAQ, the plan catalogue and the caller's own account; anything the assistant cannot do (refunds, credits, disputes, lost key without the current key, 'talk to a human', security reports) is escalated to the owner with a reference. Identity-aware: with your service key on the MCP connection (x-api-key or Authorization: Bearer) it can describe your own plan, payment status, rate limit and key preview; without a key it answers in general. It never reveals keys or other customers' data, never changes billing, and is capped per caller per day (a cap hit returns a non-LLM fallback with FAQ links). The local STDIO server refuses with support_requires_hosted_mcp.