One base URL in front of Anthropic, OpenAI, Bedrock, or OpenRouter. Crushers and caches shrink context before the model sees it. Every response ships a signed receipt with tokens saved, cache hits, and module chain. Outcomes feed patterns that ride into the next call.
Compress → prove → learn · no provider-token markup · BYOK · hash-only by default · Ed25519 receipts
One base URL. Provider keys stay yours. Provider bills inference; prxy bills the control layer.
The compounding loop
Other gateways route and log. prxy compresses work on the way in, proves what happened on a signed receipt, and learns only from verified outcomes — so the next call is cheaper, auditable, and smarter.
01 · compress
MCP tool pruning, JSON/code crushers, semantic + exact cache, and IPC keep long sessions inside the window. Optional steel-browser folds live web pages into compact tool results; CCR archives blobs you can retrieve on demand.
mcp-optimizer · structured-crusher · code-crusher · steel-browser · semantic-cache · ipc
02 · prove
Ed25519-signed JSON: identity, cost with pricing provenance, cache_status, tokens saved per module, policy decision, four canonical hashes. Verifiable offline via public JWKS.
Payment-Receipt · module_chain · cache_status · cost_estimated_usd
03 · learn
Attach an outcome to a receipt. Positive signals queue memory_candidates. A reviewer promotes patterns. Patterns inject into future requests — anchored on proof, not vibes.
POST /v1/outcomes → memory_candidates → patterns
Next call re-enters at compress — with patterns already loaded and savings on the receipt
Built for what broke this month
| What broke | Where | The module |
|---|---|---|
| Auto-compaction regression dropping user intent mid-session | Issue #36068 · Mar 19, 2026 | Compaction Bridge |
| MCP tool definitions burning 67K–143K tokens before you type | Apideck post · Mar 16, 2026 | MCP Optimizer |
| Massive JSON + file-read tool outputs refilling context every turn | Agent workloads · 2026 | Structured + Code Crusher |
| Public reports of AI coding-tool budgets outrunning forecasts | Benzinga report · Apr 2026 | Cost Guard |
| Claude Code users reporting rapid rate-limit drain | MacRumors report · Mar 26, 2026 | Semantic + Exact Cache |
| Claude Code subscription availability/pricing tests | Public timeline · Apr 2026 | MIT self-host |
| Context rot after ~2 hours of session | Widely reported · Apr 2026 | IPC + Rehydrator |
New high-signal agent incidents become module candidates. Read the dated log at /monster-log/ or rerun the proof at /benchmarks/ →
What you actually get
You send a request to api.prxy.monster with your existing provider key. The default pipeline compresses context, caches repeats, injects learned patterns, then hits your provider. The response ships model output plus signed receipt headers. Same wire format you already use.
STEP 01
curl -X POST https://api.prxy.monster/v1/messages \ -H "Authorization: Bearer $PRXY_KEY" \ -H "X-Provider-Key: $ANTHROPIC_KEY" \ -H "Content-Type: application/json" \ -d '{ "model": "claude-sonnet-4-6", "max_tokens": 200, "messages": [{"role":"user","content":"Hello, prxy."}] }'
SDK drop-in: just swap ANTHROPIC_BASE_URL
STEP 02
→ mcp-optimizer # prune tool defs → exact-cache # byte-identical hits → semantic-cache # near-match hits → structured-crusher # shrink JSON tools → code-crusher # fold file bodies → ipc + patterns # context + memory → your provider # your key, your bill
Toggle modules per key via PRXY_PIPE
STEP 03
{ "id": "msg_01HZ...", "type": "message", "role": "assistant", "content": [{ "type": "text", "text": "Hi. How can I help?" }], "usage": { "input_tokens": 12, "output_tokens": 8 } }
Receipt proves compression + cache on every call
prxy.monster does not bill you for tokens. Your provider bills you for tokens. We bill you for the gateway and the module pipeline. We never mark up inference.
Not an inference provider. Not a web proxy. Not a VPN. Not prxy.com.
How it works
For supported Anthropic Messages and OpenAI Chat Completions clients, change the base URL. Check the compatibility matrix for partial and planned APIs.
Every conversation forges patterns. Outcomes are tracked. Failures retire. Good solutions reinforce.
Patterns inject before each request. Context never resets. Your AI bill goes down over time.
Every request flows through the monster first — your key, your provider, your bill.
Featured modules
Fifteen+ composable modules · default pipeline ships compress + learn · toggle via PRXY_PIPE
Survives the auto-compaction regression in #36068. Re-injects user intent on every compaction boundary so your agent doesn't drop the thread mid-session.
The 67K-tokens-of-MCP problem. Scores each tool against the request and ships only the relevant ones. The local benchmark suite reports 33.4% average tool-token reduction across synthetic 120-tool fixtures.
The 10k-row JSON tool_result problem. Head/tail-samples arrays, truncates strings, drops empty fields — only on tool_result blocks by default. Optional CCR stores originals for prxy_retrieve.
The re-read-the-whole-file problem. Folds function bodies in file reads and fenced code while keeping signatures. Every elision leaves a visible … N lines elided marker.
Sessions don't have to start from zero. Injects relevant past solutions into the system prompt. Forges new patterns from successful resolutions. Compounds over time.
Repeat questions don't repeat costs. Embeds the request, replays the cached response above similarity threshold, and skips the upstream call on a hit. Hit rate depends on repeated workload shape.
Live-web research without a second SDK. Injects steel_scrape via Steel.dev; PRXY fulfills scrapes and auto-loops. Pair with crushers for bounded context.
AI coding-tool budgets outrunning forecasts. Per-key, per-day, per-month USD ceilings. 429 before the bill blows. Stops runaway agents in their tracks.
pricing transparency
| prxy.monster | OpenRouter | Portkey | Helicone | LiteLLM | |
|---|---|---|---|---|---|
| BYOK provider invoice | Provider bills you directly | Direct BYOK with post-free-tier fee | Customer provider key / gateway billing | Customer provider key / gateway billing | Self-hosted provider key |
| Published hosted billing unit | YES | Model token prices / credits | Recorded logs / requests | Requests + storage usage | Open source self-host |
| Provider inference markup on API-key BYOK | No markup; no provider settlement | No token markup; BYOK fee after 1M requests | No published token markup | No published token markup | No hosted bill in OSS mode |
| Public itemized payment ledger | receipts.prxy.monster | Not a public ledger | Not a public ledger | Not a public ledger | N/A |
| MCP token optimization | YES | NO | NO | NO | NO |
| Infinite context (compressed) | YES | NO | NO | NO | NO |
| Pattern learning across sessions | YES | NO | NO | NO | NO |
| Semantic cache | YES | NO | YES | NO | NO |
| Self-host (MIT/Apache) | YES — MIT | NO | YES — Apache 2.0 | YES | YES |
| Composable modules | YES | NO | NO | NO | NO |
prxy.monster charges for the gateway pipeline, not provider inference on BYOK routes. Sources: OpenRouter FAQ, Portkey pricing, Helicone pricing, LiteLLM GitHub.
the proof layer
Compression saves tokens on the way in. Receipts prove what happened — including per-module tokens.saved and cache_status. Outcomes prove what mattered. Patterns reuse what worked. Public JWKS at /.well-known/prxy-receipt-keys.json closes the loop without trusting our database.
Cost by model, project, and agent. Cache hit / miss. Policy decision. Module chain. Provider, status, latency. Public_demo / public_minimal / public_redacted / public_full / private — your call.
receipts.prxy.monster/r/<id>Per-project monthly budgets with off / warn / hard_fail enforcement. Hash-only payload capture by default; encrypted-at-rest is opt-in. BYOK + provider routing.
blocked receipt for the audit trailnone mode disables capture beyond receipt hashesEd25519 over RFC 8785 JCS-canonicalized receipt body. Public JWKS, key id prxy-receipt-2026-q2. Verify in browser, in CLI, or in your own code.
prxy-cli receipt verify <id> runs the full canonicalize-and-verify path<VerifyBadge> on every public receipt pageSubmit an outcome anchored on a receipt. Positive outcomes feed memory_candidates. A reviewer in lair promotes useful candidates into patterns. Patterns ride into future calls.
POST /v1/outcomes — succeeded / failed / partially_solved / +8 morePlays nice
Same wire format for supported Anthropic Messages and OpenAI Chat Completions routes. Most integrations are a single base URL change.
AI Coding Tools
SDKs
Frameworks
Deploy
Self-host the entire pipeline · MIT · Same modules, your hardware
Hosted gateway. Zero ops. Account-scoped memory and cache.
Single local gateway. Private data volume. MIT licensed.
Dedicated deployment for teams that need their own account boundary.
Requests, not tokens. Your provider already charges you per token — we don't double-dip.
prxy_FREE
$0 forever
5,000 requests / month · no card
prxy_PRO
$20 / month
100,000 requests / month · then $0.20 per 1k
prxy_TEAM
$99 / month
1,000,000 requests / month · then $0.10 per 1k
One request = one HTTP call into our gateway. Streaming counts as one. Cached hits count as one. Failed-upstream calls don't count. API-key BYOK users pay providers directly at provider list rates; prxy.monster does not mark up that invoice. Managed MPP is separate: direct calls are $0.50 and funded sessions provide an effective $0.05 per call; both include upstream inference.
Create a free account, get a PRXY API key, register your provider key, then paste this in your terminal. Same Anthropic Messages shape.
Common questions
/.well-known/prxy-receipt-keys.json by anyone — no prxy code required.
memory_candidates queue. A reviewer in lair promotes useful candidates into patterns, which then ride into future calls via the patterns module.
hash_only is the default. We persist the four canonical sha-256 hashes plus receipt metadata. Plaintext request and response bodies are not stored unless you opt in to encrypted_at_rest, in which case they are sealed under your X25519 public key. Outcome notes are hashed server-side and never persist in raw form. Full matrix at /security/data-retention/.
module_chain tells you exactly which modules ran.
public_demo receipts you can click into. Do not paste secrets, private code, or customer data — sandbox receipts are public_demo by design.
prxy-monster-local is the MIT self-host edition. Same module API as cloud. No telemetry to prxy.monster, no hosted sync, no managed MPP. State (cache, patterns, archived context) lives in your local volume. Use it when policy or compliance forbids sending traffic through a hosted gateway.
/v1/agent/quote, /v1/agent/sessions, production settlement, and Payment-Receipt headers are live for the managed MPP route. Direct calls are $0.50; funded sessions provide an effective $0.05 per call. ATXP MCP, native x402, and Tempo remain separate blocked or planned rails until their readiness objects mark production_live.
prxy-monster-local, @prxy/module-sdk, and prxy-cli are MIT-licensed on npm. Self-host the entire pipeline on your own infrastructure for free. The hosted gateway, lair operator dashboard, and the receipts ledger surfaces are closed source.
Create your account, choose a plan, and continue through Stripe Checkout. When payment succeeds, your prxy_ API key is provisioned and emailed automatically.
