The runtime for
production agent fleets.
Deploy autonomous agents inside your hardware perimeter or in our managed cloud. Self-hosted, Apache 2.0, powered by Trellis OTP supervision and Dynamic Cognitive Tiering across 19 AI providers. Every tool call gated, every turn replayable, zero context lost on crash.
The 11 Scenes of Sovereign Autonomy.
Click any capability below to inspect how Helix runs, recovers, schedules, and audits agents inside your perimeter.
Supervised BEAM process tree.
Not disconnected shell scripts.
Every agent in the fleet is a supervised OTP fiber managed by the Trellis Kernel. State is strictly isolated per process: if a tool hangs or memory runs out, supervision restarts the fiber in under 2ms with its goal and CloudEvents inbox intact.
Inside your perimeter.
Turn by turn. On the record.
Every other harness watches the agent loop from outside — hooks, logs, or a dashboard tailing stdout. Helix makes deterministic replay a property of the execution model: a run is a supervised BEAM process tree, every tool call passes an ordered gate pipeline, and the event log is the session.
Clicking a signal above updates this schema inspector...
Trellis OTP Kernel & Supervised Fibers
In Helix, subagents are not separate shell processes or unbounded scripts. A subagent is a child fiber spawned via %Directive.SpawnAgent.
BEAM processes isolate state natively: if a subtask crashes, OTP supervision restarts it in < 2ms with its goal intact and zero memory corruption.
Reactive Coeffects & Hot Reloading
Components declare required capabilities via inject: [:llm, :sandbox, :vector_store].
Trellis resolves providers via pub/sub and automatically cycles dependents when models or tools reload — solving the VSCode/Cordis "restart tax" with zero host downtime.
Trajectory Branching & Replay
Ported from Google Agent Executor (AX): the session log is an append-only event stream (Ash Postgres mirrored to ClickHouse). Any run can fork copy-on-write at any turn boundary to evaluate counterfactual reasoning paths or test alternative prompts without losing prior context.
Pluggable Sandboxes: Local, Docker, Sprites & WASM
Unlike harnesses hardcoding one OS sandbox, Helix exposes a unified
Helix.Capability.Sandbox
behaviour: Local
namespaces for dev, Docker
for ephemeral tools, Fly.io Sprites
(Firecracker microVMs with sub-second checkpoint & hibernation), WASM
fuel-metered plugins, and live Browser CDP Screencast.
The Task Contract (Helix.Contract)
Before an agent acts, Helix compiles natural language into a typed, bounded contract. An agent cannot silently solve an easier problem and declare victory.
EvidenceGates & Autonomy Ladder
Autonomy is earned by evidence, not assertion. Deterministic invariant checks run first and cheaply; independent model reviewers verify judgment with fresh context before tool release.
Watch Helix in motion.
Architecture walkthroughs, sandbox live demos, and multi-agent org structures.
Helix Platform — Architecture Explained
Deep dive · Trellis OTP Fibers, Reactive Coeffects & CloudEvents
Production-Grade Hosting for Existing Agents
Demo · Running Claude Code, Codex, and Gemini CLI as-is
Helix v1.0: Multi-Agent Org Structures & Gates
Release · A2A Signal Mesh, Workspace Twins & Human Approval
Just shipped: the latest Trellis Kernel release notes — everything in the core engine release, in writing.
Prefer reading? The Sovereign AI Agent Fleet Architecture Guide (Helix vs. Trinity) — the comprehensive technical comparison whitepaper.
Helix runs the body.
Synapse is the mind.
A company epistemic context graph you can see inside. Synapse is the self-improving cognitive core that lives on Helix — the living graph you met in the scenes above. Pictured below is not a mockup: it is Synapse rendering its own epistemic topology, a live 3D graph of entities, decisions, and temporal intervals that reshapes itself as the system learns your business.
How the mind compounds
The 4-stage autonomous epistemic lifecycle
Capture
Transcripts, PRs, decisions, and enterprise EDW schemas flow into Synapse in atomic epistemic graph nodes.
Connect
Each entity is indexed with bitemporal coordinates (Valid Time vs. System Time) and openCypher typed relationships.
Converge
Justification-Based Truth Maintenance (JTMS) cascades belief invalidation when upstream premises change, while W3C PROV-O records cryptographic lineages.
Consult
Every specialist agent queries Synapse and BigQuery/Snowflake/Databricks EDW connectors for live context, prior post-mortems, and architectural invariants.
The 6-Layer Memory Hierarchy
Confidence-gated writes, character injection budgets, and dual-scoped twins
Coding style, tone, language, default tools, and personal formatting quirks injected into every prompt.
Active repos, dependencies, ongoing milestones, architecture boundaries, and active team initiatives.
Recent conversation history scored with exponential time decay: high-utility interactions stay hot; stale turns fade.
Confidence-gated key-value assertions about customer environments, infrastructure IPs, and tech stacks.
Institutional rules synthesized from human corrections and verified runs: "Always route customer escalations to #on-call".
Immutable event-sourced session transcripts, tool observations, and approval tokens retained for audit and time-travel replay.
System-One Decision Substrate
Sub-500ms non-generative decisions evaluated directly on model logits, plus < 15ms dynamic routing across 19 AI providers
Evaluates typed questions (noul
binary, choice
discrete, score
rubrics) directly from model logits without generating text.
fast-jev-compaction
prunes stale tool dumps while preserving code diffs and reasoning 100% verbatim. Zero lossy summarization.
Classifies prompt complexity in < 15ms across ⚡ Fast, ⚖️ Standard, 🧠 Frontier, and 🏠 Local tiers with automatic resilient fallback (Frontier ↷ Standard).
Replaces high-variance LLM judges with deterministic trace evaluators boasting 92–913x lower variance and 100% oracle agreement.
The Paradigm Shift
Why static RAG document stores and single-agent scratchpads fail production fleets
| Dimension | Second Brain | RAG Knowledge Base | Single-Agent Vector Store | Helix Synapse Epistemic Core |
|---|---|---|---|---|
| Who it's for | One individual person | An app or query endpoint | Single-box agent fleet | Enterprise multi-tenant orgs & fleets (Dual Scope) |
| What it holds | Personal raw notes | Document chunks & vectors | Vector embeddings & flat notes | Bitemporal graph, 6-layer memory, PROV-O & Beacon skills |
| Substrate | Local Markdown | Vector DB (Pinecone/Milvus) | Git + raw Docker containers | openCypher + BEAM OTP processes + ClickHouse + TypeSafe Jev |
| Decision Latency | Manual human reading | 1.5s – 5.0s (LLM calls) | 1.5s – 4.0s (LLM calls) | < 500ms (TypeSafe Jev Logits · 0 Token) |
| Context Pruning | Manual deletion | Top-K cosine similarity | Full transcripts in context | Verbatim Compaction (100% diff fidelity) |
| Where intelligence lives | In user's head | In retrieval + black-box LLM | In flat vector distances | In bitemporal graph topology, contracts & calibrated logits |
Domain Watch
Monitors competitor repos, arXiv preprints, regulatory changes, and incident feeds. Reads new inputs through the lens of existing company knowledge and files new nodes automatically.
Incubation Loop
Allows unresolved questions, trade-offs, and design tensions to simmer across multiple runs. Re-evaluates hypotheses as new evidence arrives and converges them into hardened rules.
Personal intelligence.
Shared team brain.
Every other harness models one person working alone on one machine. Helix is the first runtime where every personal primitive gets a workspace-scoped twin from day one — team memory, shared artifacts, and automations that outlive departing employees.
Shared 6-Layer Memory
Team conventions, shared project state, and institutional lessons ("Always route escalations through #on-call") inherited by every team member's agent.
Curated Once · Shared EverywhereWorkspace-Owned Automations
Standing goals, schedules, and webhooks are owned by the Workspace, not the individual. A human steward is assigned; if they leave, the automation persists without disruption.
Continuous ExecutionTightest-Wins Governance
Workspace governance sets the ceiling: Policy ∩ Profile.
Individual agents can narrow permissions further, but can never breach the team ceiling.
Pre-trained, specialized packs.
All 7 packs on one invariant substrate.
Every department needs distinct tools, invariants, and sandbox virtualization. Under Tenet H12.12, domain packs own tools, skills, artifact types, and Gen-UI components; the agent loop, memory, tool pipeline, and session GenServer remain strictly invariant.
Engineering Pack
Automates git worktrees, Sourceror AST refactors, test synthesis, and compiler invariant checks with EvidenceGate guarantees.
Science & Research
Structured literature review across PubMed, arXiv, and ChEMBL. Molecular property lookups, protein structure search, and BLAST workflows.
Sales & RevOps
Deep lead research, ICP scoring, account brief synthesis, and automated outreach pipelines held by human authorization gates.
Marketing & Growth
Campaign synthesis, brand voice compliance checking, competitor telemetry watch, and multi-channel asset distribution.
Healthcare & Life Sciences
Clinical documentation support, MEWS acuity scoring, FHIR R4 resources, and zero-egress HIPAA Safe Harbor 18-identifier PHI scrubbing.
Finance & Trading
SOX/FINRA audit logging, EDGAR 10-K/10-Q extraction, double-entry ledger reconciliation, and deterministic integer-cent arithmetic.
Legal & Compliance
Attorney-client privilege walls, contract clause risk matrices, redlines, and multi-framework compliance mapping (EU AI Act, GDPR, HIPAA, SOC 2).
Custom Hex Registry
Package custom Jido actions, Ash domain resources, and task contracts with @behaviour Helix.Pack.
Cross-Harness Self-Improving Memory
Ingest traces from Claude Code, Cursor, Codex CLI, and Cline. Distill skills via Jev and broadcast via MCP.
Normalizes session traces from Claude Code, Cursor, Codex CLI, Cline, and Helix into a unified trace schema in ClickHouse.
Scores runs on evidence quality ($0.0-1.0$), reuse potential ($0.0-1.0$), and human correction signals ($0.0-1.0$).
Promotes high-signal runs to
SKILL.md
and Layer 5 Lessons broadcast instantly to all developer tools via MCP.
Meet your agents where your team lives
One identity gate on your perimeter. Deploy agents across chat channels, native SDKs, and the 14-module local workstation cockpit.
Connect Cursor, Claude Code, or Codex directly to your Helix fleet via 90+ native MCP tools. Orchestrate supervised subagents in plain English.
Zero-dependency bi-directional WebSocket & HTTP SDKs for Elixir (Helix.SDK), TypeScript (@helix/sdk), and Python (helix-sdk).
Install via Homebrew, npm/npx, Winget, or Scoop. Includes an embedded 14-module Bandit/LiveView Cockpit with Worktrees, Pods, Tunnels, and CI Gates.
Start a session in Slack, approve a held gate on your phone via WhatsApp, and inspect the deterministic replay trace in the LiveView workbench.
Helix & Trinity Parity.
Architectural comparison at a glance.
Helix builds a sovereign, perimeter-isolated runtime with the Synapse epistemic core and Trellis OTP kernel. Here is how the underlying substrate, execution engine, and memory models compare with conventional harnesses.
// System Parity Matrix: Helix vs. Trinity (Ability.ai)
Side-by-Side| Architecture Layer | Trinity (Ability.ai) | Helix (Trellis + Jido) | Parity Status |
|---|---|---|---|
| Runtime Substrate | Docker container orchestration + Node/Python runtime | Trellis OTP Kernel on BEAM (native actor processes) | Superset |
| Fault Recovery | Container restart on crash (~5s – 15s) | < 2ms supervisor fiber restart, zero context loss | Superset |
| Dependency Injection | Static tool & MCP configuration |
Reactive Coeffects (inject: [:llm, :sandbox])
|
Superset |
| Sandboxing Matrix | Docker containers only | 4 Pluggable: Local dev, Docker, Fly.io Sprites microVMs & WASM | Superset |
| Cognitive Model Routing | Static per-agent model selection | Dynamic Tiering (✨ Auto) across 19 providers + resilient fallback & offline local models | Superset |
| Epistemic Context Graph (Synapse) | Flat Zettelkasten knowledge graph + 3D Brain Orb + LLM | Bitemporal openCypher Graph + JTMS + PROV-O + TypeSafe Jev (< 500ms) | Full Parity + Jev |
| Memory Model | Periodic 15-minute git commits to agent repository | Formal 6-Layer Hierarchy + continuous ClickHouse event stream | Superset |
| Session Branching | Linear session log with git commit history | Trajectory Branching (Google AX): turn copy-on-write forks | Superset |
| Multi-Tenant Tenancy | Single-box / perimeter-isolated machine | Dual Scope: Personal account + Workspace twin from day one | Superset |
| Verification & Gates | Static tool & budget caps, manual pause before actions |
Task Contracts (done_when) + EvidenceGates (Maker != Checker)
|
Superset |
| Cross-Harness Learning | Runs Claude Code / Codex CLI in separate containers | Beacon Pattern: trace ingestion, Jev scoring, MCP skill synthesis | Superset |
Open source is the whole platform.
Enterprise adds governance.
Open core — Enterprise includes everything in open source.
- Agent platform & Cockpit fleet, chat, workflow canvas, local cockpit
- Trellis OTP Orchestration scheduling, monitoring, < 2ms recovery
- Channels, CLI & DASP 1.0 SDKs Slack, WhatsApp, MCP, Elixir/TS/Python SDKs
- Security core & BYOK Vault 8-gate pipeline, AES-256-GCM encrypted keys
- Sovereign deployment your hardware, zero phone-home
- User & org management invite, deactivate, activity view
- SSO, 2FA & SCIM SAML/OIDC, automated provisioning
- Advanced RBAC & Dual Scope workspaces, granular roles, twin sync
- Regulated Domain Packs Healthcare HIPAA, Finance SOX, Legal Privilege
- Compliance & GDPR/DPDP Suite ClickHouse SIEM export, SHA-256 erasure receipts
- Cognitive Policy Ceilings tier caps, budget enforcement & air-gap local models
- License & SLA management offline-verified, dedicatedVPC & 99.95% SLA
Run your first agent tonight.
Install the CLI or deploy on your own hardware — the whole platform, free to self-host. Or launch a managed workspace with us and scale your fleet with zero infrastructure overhead.
Start here — self-host.
- ✓ Free forever — Apache 2.0 licensed, self-hosted on your hardware.
- ✓ Bring the agents you've built — Claude Code, Codex, and Gemini CLI agents run as-is.
- ✓ Cross-platform CLI & SDKs — Homebrew, npm/npx, Winget, Scoop, and DASP 1.0 SDKs.
Run it with us.
- ✓ Run by our engineers — a dedicated, single-tenant instance in your perimeter; uptime, upgrades, and monitoring around the clock.
- ✓ Clears security & compliance — 8-Gate zero-trust pipeline, AES-256-GCM BYOK vault, HIPAA/SOX/Legal packs, and signed DPA.
- ✓ Still yours — same open core as public release; self-serve portable export and leave anytime.
Run your fleet in 4 commands
Docker Compose spins up backing infrastructure; BEAM runs the host. Or use the Helix CLI (brew install codemintlabs/tap/helix).
git clone https://helixloop.ai/helix.git && cd helix# Start Postgres 18, Mailpit, Redis, and ClickHouse via Docker Compose:docker compose up -d --wait# Install dependencies, setup database, and compile assets:cd apps/helix_app && mix setup# Launch the Phoenix LiveView supervision tree at http://localhost:4000:mix phx.server