RPCS-1 Agent Tuner & Translation Bridge
A stateless, read-only RPCS-1 toolkit for tuning AI agent runtime settings and cleaning up/rendering human communication.
recommend_agent_configuration — Diagnose a deployed agent from task, environment (entropy, predictability, stakes, context horizon, commitment style) and target platform/model; returns receiver profile values (TI, SG, FT, UE, AR), temperature/top_p/strategy, predicted regime, confidence, warnings, and a next test.
interpret — Detect ambiguity in a vague or underspecified message; returns literal vs. implied meaning, AR level, ambiguities, clarifying questions, and scored candidate readings.
normalize — Clean fragmented, elliptical, or run-on text into coherent prose and report fragment count.
rewrite — Produce style-specific rewrite instructions (technical, plain, socially_gentle, concise, detailed, direct) to pass to an LLM as a system prompt.
All tools are deterministic, idempotent, and non-destructive — nothing is stored or written.
This server has a remotely hosted endpoint. Open its connector listing for connection details, authentication requirements, and setup instructions:
RPCS-1 SDK — AI Agent Tuner
Start with the free tuner: find your AI agent’s likely failure mode, get runtime settings to try, and validate them with a harder case.
RPCS-1 helps teams make agent settings deliberate rather than guessed. Describe the task, change rate, predictability, stakes, relevant context horizon, and commitment style; it returns a five-primitive profile, a runtime recommendation, and a next test. The suite also includes SendRight for catching ambiguous prompts before handoff and the Translation Bridge for profile-aware communication rendering.
Repository Structure
rpcs1-sdk/
├── packages/core/ # TypeScript engine (@rpcs1/core): tuner + translation layer + receiver-profile intake
├── packages/web/ # Next.js app serving rpcs1.dev (tuner, translator, docs, Stripe, /mcp endpoint)
├── packages/mcp-server/ # Standalone STDIO MCP server (what Glama and MCP clients build)
├── sdk/python/ # Python SDK (pip install rpcs1)
├── skills/ # Canonical agent skill package (HF-HATP v2.0 SKILL.md)
├── docs/ # Architecture, deployment, launch playbook
└── .github/workflows/ # CI/CDRelated MCP server: jev-compactor
Quick Start — Python SDK
pip install rpcs1from rpcs1 import recommend_params
config = recommend_params(
task_description="Customer support agent",
environment_entropy="dynamic",
environment_predictability="somewhat_predictable",
stakes="high",
target_platform="anthropic",
)
print(config.platform_parameters.temperature) # e.g. 0.52
print(config.predicted_regime) # 'stable'
print(config.reasoning) # cites Matching PrincipleQuick Start — TypeScript Core
import { recommend } from '@rpcs1/core';
const rec = recommend({
task: { task_summary: 'Customer support agent' },
environment: {
entropy: 'dynamic',
predictability: 'somewhat_predictable',
stakes: 'high',
context_relevance: 'medium',
commitment_style: 'cautious',
},
target_platform: 'anthropic',
});
console.log(rec.platform_parameters.temperature);
console.log(rec.predicted_regime);Development
# Install dependencies
npm ci --include=optional
# Build and test TypeScript core
npm run build --workspace=@rpcs1/core
npm run test --workspace=@rpcs1/core
# Test Python SDK
cd sdk/python
pip install -e ".[dev]"
pytest -vWeb environment variables are documented in packages/web/.env.example
(Stripe, Resend, license signing, rate limits). MCP production controls are listed under
Production controls below.
The web app deploys to Vercel on Node 24 (region iad1); pushes to main trigger the production deployment.
The Matching Principle
The SDK implements Pred-09-5 from IMM Paper 9:
Stable receivers in an environment with entropy H satisfy TI ~ 1/H.
High-entropy environments → short attention windows (TI ~ 10). Low-entropy environments → long attention windows (TI ~ 90).
Every parameter recommendation traces back to this principle or the basin stability geometry (oscillation/overload/freeze boundary conditions).
Web App
Free Tuner: https://rpcs1.dev/tuner
SendRight: https://rpcs1.dev/send
Translation Bridge: https://rpcs1.dev/translator
Calibrate a communication-preference profile: https://rpcs1.dev/calibrate
The site can also explain the same product facts in technical, executive, plain-language, or literal-and-precise registers. The explanation changes; pricing, deliverables, and limitations do not.
Brand — Explicit Formula (product) / RPCS-1 (mechanism)
The site fronts one consumer product: Explicit Formula — the box on the
landing page. Explicit: says exactly what it means (the product's one job);
formula: a repeatable method. The wordmark is an advisory-sticker homage
(components/StickerLogo.tsx), deliberately distinct from the trademarked
RIAA label.
The mechanism brand — RPCS-1, the receiver engine, its laws, and its scorecard — is unchanged and renders as "Powered by RPCS-1" in the footer. House rule: outcome on the wrapper, mechanism one click deep.
The brand is a token:
packages/web/lib/brand.ts. Renaming the product is one env var (NEXT_PUBLIC_BRAND_NAME) or one line — no other code changes.Every station that used to compete for the nav (SendRight, Bridge, Translator, Calibrate, Tuner, R&D, …) stays live at its original route and is indexed at /labs (
packages/web/lib/labs.ts).The consumer domain follows the deployment: set
NEXT_PUBLIC_APP_URLwhen it goes live. rpcs1.dev remains the mechanism home either way.
SendRight (Interpretation Mirror + Hand-off)
SendRight is the type-and-send front door: type a prompt the way you'd say it out loud, see the readings it actually supports, lock in the one you meant, and hand it to your own model app with one click.
Modules (packages/core):
mirror(text)— deterministic fork detectors (no ML, no API calls). Returns{ clean, readings[], ambiguousSpans[] }. Detectors: compare-or-choose ("X or Y?" questions without an explicit verb), grouping forks ("A and B or C"), scope forks ("only ... and ..."), dangling pronouns, bare objects ("fix it"), external references ("the above"). Contract: silent on clean prompts — zero-fork controls intests/mirror.test.tsenforce it. Pure function, callable from any front end (web box, NL2Build, CLI).applyReading(text, clarifier)— appends the chosen reading's clarifier so the locked interpretation travels with the prompt.buildHandoff(vendor, prompt)/listVendors()— per-vendor capability table for opening the user's own model app with the prompt pre-filled. Prefill URL parameters are undocumented vendor behavior and churn without notice; each entry carries averifieddate and must be re-checked at release. Verified 2026-07-25: ChatGPT, Claude, Perplexity, Grok support URL prefill; Gemini and Copilot are clipboard-fallback only. Logged-out users may lose the prefill at login. All vendors degrade gracefully to clipboard.
Web: /send (packages/web/app/send) renders the box via
components/SendBox.tsx — mirror runs client-side (debounced, zero network);
the hand-off happens in the user's own app. SendRight never makes the model
call and never sees the answer.
Feasibility boundary (honest scope): reasoning-stream digests and mid-generation stop/realign are only possible where rpcs1 itself owns the API call (the fan-out / power-user mode, not yet shipped). They are structurally impossible in vendor chat UIs and via the MCP surface — SendRight's hand-off path intentionally trades those away for zero keys, zero cost, and zero data custody.
MCP Server
RPCS-1 is also available as a public, anonymous, read-only MCP server:
https://rpcs1.dev/mcpIt exposes eight read-only tools across four families:
recommend_agent_configuration— diagnose an AI agent against environmental entropy, predictability, stakes, context horizon, and commitment style; receive runtime settings to try and a next test.interpret,normalize, andrewrite— detect ambiguity, turn fragmented text into coherent prose, and return style-specific rewrite instructions.route_intent— entropy routing over competing interpretations of a message: the calling model proposes candidate readings (paraphrases and priors); the deterministic router computes the posterior and decides commit, present options, or clarify. The commit-vs-clarify authority in the pipeline.calibrate_profile,prepare_prompt, andrender_reply— create a continuous communication-preference profile, recover intended meaning before an action, and render a reply for that profile.
Translation Layer
"Say what you mean. Hear what they meant."
The Translation Bridge treats the profile as a transportable parameter, not a category label. The five-question
Calibrate flow measures communication preferences for rendering only; it is not a psychological assessment or diagnosis.
prepare_prompt / render_reply use that profile on the inbound and outbound sides of an interaction. The canonical
agent-facing specification lives at skills/rpcs1-translation-layer/SKILL.md.
Tuner examples
The first useful call is a support copilot under live pressure:
Use recommend_agent_configuration to diagnose my support copilot.
Task: refund and billing dispute triage
Environment: dynamic, somewhat_predictable, high stakes
Context relevance: medium
Commitment style: cautious
Target platform: anthropicThe output should lead with the five-primitive profile, failure-risk score, predicted regime, runtime posture, and next test to run.
The second useful call is a coding agent in a changing repository:
Use recommend_agent_configuration to diagnose my coding agent.
Task: inspect a changing repository, edit files, run tests, and open a pull request
Environment: moderate, somewhat_predictable, medium stakes
Context relevance: long
Commitment style: balanced
Target platform: openaiThe output should still lead with the five-primitive profile, failure-risk score, predicted regime, runtime posture, and next test to run.
Connection details and client compatibility notes are available at https://rpcs1.dev/docs/mcp. Practical coding, support, and research examples are available at https://rpcs1.dev/docs/examples.
Hyperagent uses the fixed public OAuth client hyperagent-rpcs1 with PKCE and the registered
callback https://hyperagent.com/api/mcp-servers/callback. No client secret is required.
The MCP surface exposes the deterministic agent-tuning workflow alongside read-only translation and per-user rendering tools. New tools should be added only after their scoring or behavior contracts are implemented and tested in the core package.
Discovery metadata:
OpenAPI: https://rpcs1.dev/openapi.json
LLM overview: https://rpcs1.dev/llms.txt
MCP Registry manifest:
server.json
Production controls:
MCP_HOURLY_LIMITcontrols per-instance MCP throttling (default:120requests per IP/hour).MCP_MAX_BODY_BYTESlimits request bodies (default:65536bytes).MCP_ALLOWED_HOSTSis a comma-separated production host allowlist.MCP_ALLOWED_ORIGINSis an optional comma-separated browser-origin allowlist. Leave it blank to reject cross-origin browser requests.MCP_OAUTH_JWT_SECRETsigns short-lived OAuth authorization codes and access tokens./api/healthreports deployment and MCP readiness metadata.
For globally consistent abuse protection across Vercel instances, configure a Vercel Firewall
rate-limit rule for /mcp. The in-process limiter is defense in depth, not a distributed quota.
Glama Docker checks should build and launch the local STDIO server, not connect to the hosted
https://rpcs1.dev/mcp endpoint. Use this build spec:
{
"buildSteps": [
"npm ci --include=optional",
"npm run build --workspace=@rpcs1/core",
"npm run build --workspace=@rpcs1/mcp-server"
],
"cmdArguments": [
"mcp-proxy",
"--",
"node",
"packages/mcp-server/dist/index.js"
],
"environmentVariablesJsonSchema": {
"type": "object",
"properties": {},
"required": []
},
"placeholderArguments": {}
}License
MIT
Available Tools
4 toolsinterpretInterpret ambiguous human inputARead-onlyIdempotentInspect
Detect ambiguity in a user message and score candidate interpretations using the RPCS-1 Signature Ambiguity Framework. Returns literal summary, implied meaning, confidence, AR level (AR0-AR5), ambiguities, clarifying questions, and per-candidate scores (IC, UE, EC, NM, SG, TI). Use when a user says something vague, passive-aggressive, or underspecified.
Inputs: text (required), risk.
normalizeNormalize fragmented human inputARead-onlyIdempotentInspect
Clean up text with ellipses, fragments, and run-on thoughts into coherent prose. Returns the number of fragments detected and the joined version. Use when a user types stream-of-consciousness or fragmented input.
Inputs: text (required).
recommend_agent_configurationRecommend AI agent configurationARead-onlyIdempotentInspect
Diagnose why a deployed AI agent may fail. Takes environmental entropy, predictability, stakes, context horizon, and commitment style, then returns receiver profile values (TI, SG, FT, UE, AR), platform parameters (temperature, top_p, strategy), regime prediction, reasoning, and warnings. Optionally pass target_model to attach MEASURED per-model receiver posture (E-LIT table). Deterministic, stateless, read-only — does not store past recommendations.
Inputs: task, environment, target_model, target_platform. Returns: warnings, reasoning, confidence, predicted_regime, +3 more
rewriteRewrite text for a target audienceARead-onlyIdempotentInspect
Get rewrite instructions for adapting text to a specific audience style: technical, plain, socially_gentle, concise, detailed, or direct. Pass the result to an LLM with the rewrite_instructions as the system prompt. Use when communication needs tone adjustment.
Inputs: text (required), style.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 4 tools
Each tool has a clearly distinct purpose: recommend_agent_configuration handles diagnostics, interpret detects ambiguity, normalize cleans text, and rewrite adjusts tone. No overlap in functionality.
All tool names follow a consistent verb_noun pattern (recommend_agent_configuration, interpret, normalize, rewrite), with descriptive suffixes where needed. No mixing of conventions.
Four tools is a well-scoped set for an agent tuner and translation bridge, covering diagnostics, ambiguity resolution, text normalization, and style adaptation without excess or deficiency.
The tool set lacks a direct translation feature despite the server name. Additionally, the rewrite tool only provides instructions, not actual output, creating a dependency on an external LLM. Core operations are present but notable gaps exist.
Maintenance
Related MCP Connectors
Token guard and rate limiter preventing runaway API cost spikes for OpenAI and Anthropic.
Stops runaway AI agent loops and token burn. Checks each tool call before it runs. No signup.
Runtime gate for AI agent claims. Catches sycophantic drift, circular self-validation, causal errors
Cryptographic 563-leaf LLM rate oracle, context-cliff auditor, and prompt cache optimizer.
Related MCP Servers
- AlicenseNot gradedqualityCmaintenance5-voter Byzantine council halts agentic loops when no real progress is happening, preventing token waste on infinite spins.MIT
- AlicenseNot gradedqualityBmaintenanceProvides deterministic context compaction and safety gating for AI agents, preserving original messages verbatim while dropping irrelevant ones and flagging destructive commands.2MIT
- AlicenseAqualityBmaintenanceZero-dependency token optimizer, test-failure triage gate, and System 1.5 semantic guardrail for AI coding agents. Intercepts compiler errors, missing dependencies, and doom loops in < 500µs.612MIT
- AlicenseNot gradedqualityBmaintenanceEnables detection of autonomous agent infinite tool loops, planning deadlocks, and liveloss cycles via trajectory state entropy analysis. Also provides safety guardrails for prompt injection, destructive commands, budget overruns, and data leakage.7MIT