v0.1.2 Released — Model Context Protocol (MCP) Server live on npm

One Project.
One Brain.
Unlimited AI Agents.

AgentHelm gives every AI coding agent a shared Project Brain so they remember architecture, APIs, decisions, and project knowledge instead of starting from scratch.

Free forever for up to 3 agents · No credit card required

Claude Code
Cursor
Codex
OpenAI SDK
CrewAI
Continuous sync
THE PROBLEM

AI Agents Forget Everything

Every time your agent starts a new session, it begins with a blank slate. No memory of past decisions, configuration patterns, or API specifications.

1

Context Amnesia

Agents don't remember project architecture, API choices, or database migrations. Every run starts from scratch.

Duplicate engineering overhead
Contradictory design paths
2

Knowledge Silos

Engineering context is locked in single agent sessions. What one agent learns is lost to all others on the codebase.

No shared model memory
Inconsistent styling implementations
3

Governance Drifts

Without shared project standards, agents write code using deprecated functions or insecure library wrappers.

API billing resource leaks
Outdated pattern integration
THE SOLUTION

One Project Brain for All Agents

AgentHelm builds a shared memory repository that persists across agents, sessions, and platforms.

Project Brain

A centralized, version-controlled knowledge base containing repo styling schemas, API declarations, and design rules.

Context Injection

Rank and select knowledge segments dynamically to supply relevant context rules to agents before they start writing.

Continuous Sync Loop

Agents propose newly discovered project decisions back into the ingestion validation queue, keeping the brain up-to-date.

Ecosystem Integration Benefits

  • Prevent duplicate engineering setup loops
  • Enforce design patterns consistently
  • Retain critical architecture decisions
  • Simplify agent context windows token usage
  • Audit code generation choices in timeline logs
  • Catch invalid dependencies at ingestion validation
Compatible out of the box with Cursor, Windsurf, Claude Code, and custom autonomous LLM setups.
HOW IT WORKS

The Brain Pipeline

Knowledge flows through a secure pipeline: proposal → noise filter → validation → verification → analysis → merge planning → publishing → Project Brain → context injection.

Noise Filter

RunningActive

Sarvam-30B classifies raw agent observations as worth compiling or routine noise, before they're considered as a formal proposal

Proposal Ingestion

Queued

Agent submits knowledge proposal containing decisions, files, APIs, DB changes

Sanitization

Queued

Strips secrets, PII, normalizes markdown formatting, validates schema structure

Permissions Validation

Queued

Check tool scopes allowlist (@read, @side_effect, @irreversible)

Authenticity Check

Queued

Validate agent registration state and active JWT token signatures

Confidence Scoring

Queued

Heuristics-based trust scoring and project configuration validation

Conflict Analysis

Queued

Inspect active brain entries for logical overrides and structural conflicts

Merge Execution

Queued

Execute logical additions, mark supersessions, and commit changes

Brain Versioning

Queued

Commit update to active database state and publish version tag

Project Brain Ingestion

Queued

Update shared brain ready for semantic context injection to active agents

Pipeline Execution Log
LIVE
Waiting for pipeline to start...
Model Context Protocol

Zero-Plugin setup with MCP

Expose the Project Brain directly to your IDE agents. AgentHelm is natively compatible with Claude Code, Cursor, Windsurf, and Claude Desktop using the Model Context Protocol.

Connection Config

Configure Claude Desktop to read/write to the Project Brain automatically.

{
  "mcpServers": {
    "agenthelm": {
      "command": "npx",
      "args": ["-y", "agenthelm-mcp"],
      "env": {
        "AGENTHELM_CONNECT_KEY": "ahe_live_...",
        "AGENTHELM_PROJECT": "my-production-app",
        "AGENTHELM_BASE_URL": "https://agenthelm.online"
      }
    }
  }
}
agenthelm-history --blame
project-brain:master
SECURITY ARCHITECTURE

Built-In Security

Every layer hardened. Fail-closed by default. JWT agent authentication. Scoped tool permissions. Complete audit logs.

JWT Authentication

Every SDK request runs authenticated via JSON Web Tokens. Agent tokens are generated during handshakes, expire in 12 hours, and are validated using secure HSM key hashes.

12-hour token expiration lifecycle
Dynamic handshake protocol validation
Cryptographic HS256 JWT signature verification
Per-agent token isolation safeguards

Scoped Tool Permissions

Enforce safety bounds on agent execution. Declare agent actions using decorators like @read, @side_effect, or @irreversible. Force human approval on irreversible actions.

Explicit tool validation allowlists
Throw typed errors on unauthorized tool requests
Hold irreversible tool calls in pending queue
Human-in-the-loop validation via dashboard

Audit Trail Logging

Maintain complete accountability for agent choices. Every context injection, knowledge proposal, and reasoning decision is tracked with permanent timestamp logs.

Immutable reasoning steps history
Proposal submissions & merge status tracking
Context injection requests correlation
Complete agent lifecycle audit trail

Anomaly Alerts

Actively monitor for runaway agents. Detect token consumption spikes, silent crashes, and high error rate thresholds instantly to protect backend billing APIs.

Silent agent detection (10 min idle)
High error rate trigger notifications (>20%)
Sudden token spike checks (hourly avg multiplier)
Real-time alerts via Telegram bot

Fail-Closed Design Paradigm

Agent Silent Connection

Monitored at ingest. Auto-transitions to 'stopped' state if liveness heartbeat fails.

Validation Error

Rejected immediately. Never merges into target Project Brain. Emits warning code.

Permissions Defect

Blocked at runtime. Throws immediate SDK exception. Logs event to security alerts.

OBSERVABILITY ENGINE

Integrated Observability

Correlate logs, track percentiles, and check agent health metrics directly inside your control panel.

In-Memory Metrics

Calculate processing percentiles directly inside the backend engine. Buffer stage events and group metrics dynamically by step.

Pipeline stage duration statistics
Proposal throughput tracking
p95 latency calculation benchmarks
Success vs failure count rates
Custom in-memory collector bounds
Error categories grouping

Structured Logs

Zero-dependency structured logs. Capture agent activity, duration timings, status codes, and error trace identifiers.

Standard JSON-structured output format
Categorized trace ID linkage
Log severity levels: info, warn, error
Proposal/Project context metadata tags
Console stream output redirection
Queryable records via timeline dashboard

Trace Correlation

Trace requests from agent script to DB. Associate every reasoning decision and context injection with unique transaction IDs.

UUID trace ID correlation propagation
Link SDK calls to dashboard events
Trace lifecycle correlation in logging
Step-by-step reasoning steps tracking
Proposal status correlation states
API request tracking hashes

Health Checks

Proactive service checks. Monitor database readiness, ping response speeds, and active memory queue bounds.

System ready/live liveness indicators
Ping heartbeat latency monitoring
Supabase connection integrity check
Redis lock queue connectivity checks
System warning detection logs
Token budget consumption alerts
PRODUCTION SDKs

Unified SDK Integrations

Choose your runtime language. Same core capabilities, security boundaries, and telemetry correlation.

Python SDK

v1.1.0
pip install agenthelm-sdk
import agenthelm
dock = agenthelm.connect("ahe_live_...", name="My Agent")
Async/await native handlers
Type hinting support built-in
Pydantic-based payload models
Context managers for checkpointing
Decorator safety (@read, @side_effect)
Reconnections with backoff logic

Node.js SDK

v1.0.1
npm install agenthelm-node-sdk
const { AgentHelm } = require('agenthelm-node-sdk')
const dock = new AgentHelm("ahe_live_...", { name: "My Agent" })
First-class TypeScript support
ESM + CommonJS dual-package bundling
Zod schema validation structures
Structured output event emitters
Lifecycle graceful shutdown hooks
Local logging adapters correlation

REST API

vv1
curl -H 'Authorization: Bearer ahe_live_...'
Language agnostic HTTP client requests
POST /api/sdk/proposals POST /api/sdk/inject POST /api/sdk/ping
Standard HTTP POST JSON payloads
JWT Bearer authentication header validation
Detailed validation error codes returned
Stateless event timeline reporting endpoints
Idempotency proposal hash verification
CORS-ready REST endpoint bindings

Core Agent Capabilities

Context Injection

Token-budgeted relevance scoring. Injects only the most relevant brain entries for the current task hint.

State Checkpoints

Automatically persist agent state snapshots. Resume from precise step index on failures.

Safety Decorators

@read, @side_effect, @irreversible declarations. Hold dangerous operations for approval.

Telegram Bot Alerting

Receive immediate notifications on high errors, token consumption spikes, or liveness failures.

Structured Logging

Trace events correlated with UUID transaction identifiers. Clean JSON logs emitted to console.

Metrics Collector

Compute duration averages and p95 latency statistics per pipeline stage in memory.

ECOSYSTEM

Works With Your Stack

AgentHelm sits between your agents and the brain. Any framework. Any model. Any cloud.

AI Coding Agents

Claude Code

IDE-native agent with full repo context

Cursor

AI-first editor with Composer

Codex

OpenAI's coding agent

GitHub Copilot

Chat & inline completions

Windsurf

Cascade agentic flows

Cline

Autonomous coding assistant

Agent Frameworks

CrewAI

Role-based multi-agent crews

LangGraph

Stateful graph orchestration

AutoGen

Microsoft's multi-agent framework

OpenAI Assistants

Built-in tool use & threads

Semantic Kernel

Microsoft's AI orchestration

Haystack

RAG & agent pipelines

LLM Providers

OpenAI

GPT-4o, o1, embeddings

Anthropic

Claude 3.5 Sonnet, Haiku

Google

Gemini 1.5 Pro, Flash

Mistral

Large, Medium, Small models

Groq

Ultra-fast inference

Ollama

Local model hosting

Infrastructure & Tools

Supabase

Postgres + Auth + Realtime

Vercel

Edge functions & hosting

Docker

Containerized agent deployments

Kubernetes

Orchestrated agent clusters

Redis

Caching & pub/sub for agents

Telegram

Mobile agent control

Unified SDK. Universal Compatibility.

AgentHelm's SDK wraps any agent framework. No vendor lock-in. Switch frameworks without rewriting your governance layer.

Python SDKNode.js SDKREST API
PRICING

Simple, Transparent Pricing

Get started free, unlock advanced failure tracing on Indie, or orchestrate agent swarms on Studio.

Free

$0

Perfect for individuals and small personal projects.

  • 3 Agents limit
  • 100,000 tokens/month
  • 7-day log history
  • Telegram alerts & control
Popular

Indie

$19/mo

Ideal for indie developers building production wrappers.

  • 10 Agents limit
  • 2,000,000 tokens/month
  • 30-day log history
  • AI failure explanations
  • All anomaly alerts

Studio

$99/mo

For orchestrating agent teams and visual cost breakdowns.

  • 1,000 Agents limit
  • Unlimited tokens usage
  • 90-day log history
  • Swarms orchestration
  • Trace Replay (Time-Travel)
  • Visual Cost Breakdown

Feature Comparison

Core FeatureFreeIndieStudio
Monthly Price$0$19$99
Agents Included3101,000
Token Budget / mo100K2.0MUnlimited
Log Retention7 days30 days90 days
Telegram Alerts
AI Failure Explanations
Anomaly Alerts
Swarms Orchestration
Trace Replay (Time-Travel)
Visual Cost Breakdown

Need Custom Scaling or Security?

Get dedicated brain clusters, on-prem deployment, custom SLAs, and dedicated priority support.

Contact Enterprise Sales
DOCUMENTATION

Start Building in Minutes

Explore standard guides, integration specs, and copy-paste examples to wire up your agents.

Copy-Paste Quickstarts

Python
# pip install agenthelm-sdk
import agenthelm

dock = agenthelm.connect(
    "ahe_live_...",
    name="research-agent"
)

# Submit knowledge proposal
dock.propose(
    summary="Added user auth module",
    decisions=["JWT with refresh tokens"],
    files_modified=["auth/jwt.py", "auth/models.py"],
    apis_affected=["POST /auth/login"],
)

# Inject context for a task
context = dock.inject(
    project="my-project",
    task_hint="Implement password reset"
)
Node.js
// npm install agenthelm-node-sdk
const { AgentHelm } = require('agenthelm-node-sdk')

const dock = new AgentHelm("ahe_live_...", {
  name: "payment-agent"
})

// Submit knowledge proposal
await dock.propose({
  summary: "Stripe webhook handler",
  decisions: ["Idempotency keys required"],
  files_modified: ["payments/webhook.js"],
  apis_affected: ["POST /webhooks/stripe"],
})

// Inject context
const context = await dock.inject({
  project: "my-project",
  task_hint: "Handle subscription cancellation"
})
cURL (REST)
# Submit proposal
curl -X POST https://api.agenthelm.online/api/sdk/proposals \
  -H "Authorization: Bearer ahe_live_..." \
  -H "Content-Type: application/json" \
  -d '{
    "project": "my-project",
    "content_hash": "abc123...",
    "payload": {
      "summary": "Database migration v3",
      "decisions": ["Add idx on user_email"],
      "db_changes": ["ALTER TABLE users ADD INDEX..."]
    }
  }'

# Inject context
curl -X POST https://api.agenthelm.online/api/sdk/inject \
  -H "Authorization: Bearer ahe_live_..." \
  -H "Content-Type: application/json" \
  -d '{
    "project": "my-project",
    "task_hint": "Optimize slow query"
  }'

Ready to build your Project Brain?

Start free forever for up to 3 agents · Integrate in 5 minutes