LLM Pricing

LLM Pricing MCP Server

[CI/CD Pipeline] [License: MIT]

LLM pricing you can verify, route on, and act on: 32 MCP tools, a REST API and browser UIs over pricing data from 30 LLM providers.

Live at: https://llm-pricing-api.fly.dev · API reference: /api-docs · Swagger: /docs

What it does

• Verified prices: every curated price is checked against a live external registry; anything drifted more than 5% is withheld, never silently served.
• Routes, not just quotes: picks a model by cost and quality for your task.
• Executes: the dispatch gateway runs the best eligible model on your own provider keys.
• Tracks real spend: records usage, analyzes sessions, fires budget alerts.
• Price alerts: webhooks on price changes, self-managed with an alert_token.
• Local currency: optional currency / country on pricing and estimates (13 currencies).
• Many surfaces: MCP (STDIO + HTTP), REST, chat agent, calculator / compare / history UIs, billing dashboard, FAQ (/faq), published legal pages and security statement (/terms, /privacy, /refunds, /dpa, /subprocessors, /security).
• Customer support: a grounded support assistant (Support mode on /chat, support_chat MCP tool) that knows the FAQ, the plan catalogue and your own account, and escalates anything it cannot do to the owner with a reference.
• Retired providers: Anyscale (2026-10)

Quick start

Claude Desktop (remote, no install): open Settings → Connectors → Add custom connector and paste:

https://llm-pricing-api.fly.dev/mcp

• Claude Desktop ignores a url entry in claude_desktop_config.json; that file only launches local commands.
• A connector sends no custom headers. For tools that need your service key, use the npx bridge below.

npx bridge (STDIO, no clone, no Python): add to claude_desktop_config.json:

{  "mcpServers": {    "llm-pricing": {      "command": "npx",      "args": ["-y", "llm-pricing-mcp"],      "env": { "LLM_PRICING_API_KEY": "<your-service-key>" }    }  }}

• LLM_PRICING_API_KEY is optional; it is forwarded as x-api-key. Options are in npm/llm-pricing-mcp/README.md.
• Other clients (Cursor, VS Code, Claude Code): docs/MCP_QUICK_START.md · Claude details and troubleshooting: docs/CLAUDE_INTEGRATION.md.

Run locally (Python 3.11):

pip install -r requirements.txtuvicorn src.main:app --reload --port 8000   # REST + POST /mcp; STDIO server: python mcp/server.py

MCP tools

All 32 tools are listed over POST /mcp and STDIO. MCP protocol version 2024-11-05.

Pass client (e.g. Claude Code, cursor, github copilot, cline) to recommend_model, optimize_workload, compare_costs or POST /router/recommend and the answer is restricted to models that tool can run before ranking; on the hosted MCP server it is inferred from the handshake when omitted, and an explicit value always wins. Recognised clients and the override order are in docs/MCP_QUICK_START.md.

The chat agent binds 27 of the 32 MCP tools (all except ask_agent, support_chat, get_telemetry, get_cache_stats and dispatch).

Key features

Dispatch gateway

• Add a dispatch policy to POST /v1/chat/completions. The best eligible model runs for real and you get a dispatch_receipt with the chosen model, reason, actual cost and savings.
• Objectives: balanced, cheapest, best_quality, lowest_latency, pinned. Failover never drops below your min_quality_score.
• Supports stream: true (OpenAI-format SSE; the receipt arrives in the final chunk).
• Text-only, single-choice. tools, n > 1 and image parts are rejected up front, with the feature named.
• Also available as the dispatch MCP tool. Executable providers are listed at GET /dispatch/providers.

Stored provider keys

• Store each provider key once per org with POST /dispatch/keys. After that, callers need only their service key.
• Keys are AES-256-GCM-encrypted at rest and never returned; listings show only a last-4 fingerprint.
• A per-request X-Provider-Key-<Provider> header still overrides the stored key.

Price alerts with alert_token

• Registering a price alert returns a one-time alert_token. It is stored only as a hash and nobody can recover it.
• View or delete your alert with no account: POST /pricing/alerts/self/view or /self/delete with the token in the JSON body, or the manage_price_alert tool.
• Listing or deleting all alerts is operator-only.

Webhook SSRF guard

• Webhook URLs must be public https. Private, loopback, metadata and unresolvable hosts are refused with 422 invalid_webhook_url.
• The URL is checked at registration and again at every delivery, connecting only to the checked IP; redirects are never followed.
• Older alerts that break the rules are kept but skipped at delivery.
• WEBHOOK_ALLOW_PRIVATE_TARGETS is for local development only; never set it in production.

Cost estimation

• POST /cost-estimate (one model) and POST /cost-estimate/batch (1–50 models). No key needed; requests are rate-limited.
• predict_cost and scripts/audit_repo.py estimate the cost of a workload or a codebase's LLM call sites.

Data quality

• GET /data-quality gives one trust score, confirmed_pct, plus counts of drift-withheld, never-priced and stale models.
• Every price carries price_source_kind (provider or registry). check_price_drift gives per-model detail.

Usage and budget

• POST /usage records actual spend. GET /usage/summary and GET /usage/session/{session_id} analyze it.
• Budget alerts (POST /usage/alerts) fire a webhook at a threshold; hard_stop: true also blocks dispatch for your org.

Customer support and plans

• Support mode on /chat and the support_chat MCP tool answer plan, billing, key, rate-limit and how-to questions only from the published FAQ (docs/FAQ.md), the plan catalogue (GET /billing/config) and, with your service key, your own account. No invented prices, policies or response times.
• Everything outside its authority (refunds, credits, disputes, failed payments, a lost key without the current key, bug/feature/security reports, "talk to a human") becomes an escalation in the owner's operator-only queue with a reference; POST /support/escalations/status returns the owner's reply.
• Plans describe only what is implemented: rate limit, support assistant allowance and escalation priority. Upgrade and Manage Billing work when Stripe is configured; POST /billing/key/rotate rotates a key with the current one; GET /billing/account/export downloads everything held about your organisation and POST /billing/account/delete (body {"confirm": "DELETE", "org_id": ...}) irreversibly erases it, ends any Stripe subscription and opens a refund review when you delete inside the 14-day window (v1.87.0); service keys are stored only as a SHA-256 hash; the Stripe webhook handles failed renewals (14-day grace, then the free limit), refunds and disputes, and ignores duplicate deliveries.
• Legal pages (v1.86.0): /terms, /privacy, /refunds, /dpa, /subprocessors are generated from docs/legal/*.md (scripts/sync_legal.py), linked from the site footer, /billing and /faq, and show a draft notice until the owner turns on LEGAL_PAGES_LIVE and every placeholder is replaced. Paid checkout stays disabled until then; the Stripe customer portal is unaffected. Refunds: full within 14 days of the first payment, pro-rated after, every request reviewed by the owner via the support assistant. /security states what the service does and does not have (no SOC 2, ISO 27001 or third-party penetration test). The pre-launch list is docs/LAUNCH_CHECKLIST.md.
• Spend: per-caller daily LLM caps by tier (anonymous strictest) under the orchestrator's global monthly ceiling; a cap hit returns a non-LLM fallback with FAQ links. Every support action leaves an operator-only, append-only record.

Orchestrator core (operator-only)

• One place where the operator's inbound work (GitHub issues/PRs, the weekly price audit, owner requests) is accepted exactly once, routed by trigger class and owner-applied labels only, and held at human checkpoints.
• /orchestrator/* plus an Orchestrator block on /admin: approval queue, parked runs, run log (CSV), global pause, per-workflow mode, budgets. Nothing merges, moves money or sends anything without a recorded approval.
• The SDLC pipeline is wrapped as an owner checkpoint: the queue shows the exact /release-feature command; the PR it opens completes the run and posts one status comment on the issue.

Compliance audit (operator-only, v1.88.0)

• A weekly (Mondays 07:00 UTC, compliance-audit.yml) and on-demand mechanical audit of the whole system: legal placeholders and the legal-live gate, Litestream coverage of every database, production secret presence (names only), version consistency, data-quality regression against the previous report, the weekly price audit, orchestrator budget, service-key hashing, Stripe reconciliation when STRIPE_AUDIT_KEY is set, and the launch checklist's "done" items. It is report-only (changes nothing audited), makes no legal-compliance claim, and opens an owner approval only when a check fails.
• POST /audit/run, GET /audit/reports, GET /audit/reports/{id} (latest alias) and a Compliance audit panel on /admin. Reports hold counts, file and setting names and opaque prefixes, never emails, keys or Stripe ids.

Document agent (operator-only, v1.89.0)

• Owner-requested decks, one-pagers, release briefs and status reports from a closed four-type catalogue (release_notes_brief, pricing_comparison_deck, pricing_one_pager, internal_status_report). Every figure comes from a product data call made during the generation and is recorded in a provenance manifest; every number in the text is cross-checked against it (untraceable numbers are removed and counted); each price carries its price_as_of and a status label, and models without a usable price show "price not available" and stay out of rankings and charts.
• GET /documents/types, POST /documents/generate, GET /documents/artifacts, GET /documents/artifacts/{id} and .../download?format=html|md|pptx|docx (self-contained printable HTML, Markdown, a PowerPoint deck with native tables, native charts and speaker notes, or a Word document with native tables and numbered notes; a deck also saves as a Word handout, a one-pager has no .pptx and answers 415), plus a Documents panel on /admin. One bounded drafting call per document (claude-haiku-5-5 by default, claude-sonnet-5-5 on request) with a labelled data-only fallback; each document waits for the owner's publish approval. Nothing is sent anywhere, no MCP tool exposes it, and customer-visible behaviour is unchanged.

Access and keys

• Send keys as x-api-key: <your-key> (MCP also accepts Authorization: Bearer <your-key>).
• Tiers: Free 30 / Pro 120 / Enterprise 600 req/min. Full endpoint and auth map: /api-docs.

Configuration

Set via environment variables or .env. An unknown key in .env stops startup with an error naming it.

Full list: src/config/settings.py. Production settings: docs/DEPLOYMENT.md.

Development and tests

pytest tests/ -q

• 4117 passing tests, CI/CD on every PR (test → lint → bandit → OSV → gitleaks → deploy); CodeQL and Dependabot run as additional scanners
• Git Flow: feature/* → develop → master. See docs/CONTRIBUTING.md.
• Internals, services, data stores and CI: docs/ARCHITECTURE.md.
• Deploys to Fly.io from master when all CI gates pass: docs/DEPLOYMENT.md.

Links

• docs/INDEX.md: every doc
• docs/archive/README.md: historical docs (blue-green deployment, MCP validation report)
• /whats-new: release notes
• /faq: plans, billing, keys, rate limits, how to get help
• /terms · /privacy · /refunds · /dpa · /subprocessors · /security: legal pages and security statement
• /mcp-setup: MCP integration hub
• License: MIT, see LICENSE