// For developersMCP-nativeApache 2.0 · open-coreMulti-providerTypeScript

Build agents
your way. Locally.

The control plane for local intelligence — built like a developer tool. MCP-native cognitive layer, multi-provider routing, Apache 2.0 source. Workspace packages you can read, fork, and ship against.

npm install @onenomad/pyre · pyre start · auto-detects your GPU and picks a model that fits.
6
Workspace packages
pnpm + Turborepo
MCP
Native protocol
Engram · Persona · Cortex
Apache 2.0
License
Open source from day one
// The moat

The local cognitive stack.
Free in Core.

Ollama and LM Studio run models. Pyre runs an orchestration system on top — four engineered pieces that turn 16 GB of consumer VRAM into something that feels like a 200K-context cloud model. Co-designed across Engram, Persona, and the Pyre engine, so the components compound instead of fighting each other.

01@onenomad/pyre-context-budget

Context Budget Engine

One tested module owns the watermark math: clamp(floor(ctx × 0.65), 2k, 24k). Slot-aware so the runtime decides what to compact instead of dropping the conversation. Renderer and engine read from the same source — no drift between client and server.

02@onenomad/pyre-tool-vault

Tool Output Vault

Pyre exposes ~70 tools — that's 70 ways to blow your context. The Vault catches tool returns, stores raw output on disk, and hands the agent a structured summary plus a get_tool_output(id) escape hatch. 40–60% token reduction on agentic sessions.

03@onenomad/pyre-compaction-sidecar

Compaction Sidecar

A small model on a parallel slot summarizes scrollback without touching your main inference loop. OpenAI-compatible — works with llama-server, Ollama, vLLM, OpenRouter. Drop-in upgrade for the Vault summarizer; falls through gracefully on outage.

04@onenomad/persona-mcp

Persona minimal-context mode

Soul + persona at three sizes — minimal (~400 tok), standard (~1-2k), full (~3-16k). Pick the budget; keep the personality. Drops VOICE & STYLE on minimal so a 14B model on a 16 GB GPU still feels like itself at 30K visible context.

// Receipts
92%
LoCoMo R@10
Engram, today — beats Mem0 / Zep / Letta
40–60%
Token reduction
Tool Output Vault on agentic runs
8 hr
Agent run, no blowup
Side-by-side video vs vanilla auto-truncation
Soon
LongBench-v2 · HELMET · RULER
In the eval queue · receipts coming
// For developers

Built like a developer tool.

Pyre exposes its insides. Workspace packages with real test suites, an MCP-native cognitive layer, OpenClaw-compatible skills, and an Apache 2.0 license that lets you read and fork everything.

MCP-native end-to-end

Engram, Persona, and Cortex are independent MCP servers. Plug them into Claude Desktop, Cursor, or any MCP client — Pyre is the orchestrator, not the gatekeeper.

Multi-provider routing

Switch between OpenAI, Anthropic, OpenRouter, Ollama, llama.cpp, MLX, and LM Studio per conversation. BYO keys; per-conversation override.

Apache 2.0 · open-core

Pyre Core, Engram, and Persona are Apache 2.0 — open source from day one. Read, fork, contribute. The code is the spec. Cortex is the one exception: source-available under a separate commercial license.

Capability detection

Auto-detects CPU, Mac Metal, CUDA, ROCm, Vulkan on first launch. Recommends quants and context lengths that fit. No config files, no hardcoded model lists.

Workspace packages

pyre-context-budget · pyre-tool-vault · pyre-compaction-sidecar · pyre-vram-calc · persona-mcp · engram-memory. Real packages, real test suites, real npm releases.

Skills + plugins

OpenClaw-compatible skill format (SKILL.md). Plugin catalog ships in Pro. Build once, run across desktop, CLI, and web.

// Open source

Read it. Fork it. Ship against it.

Apache 2.0 from day one. Workspace packages with real test suites, a SKILL.md format that any agent runtime can load, and architecture docs that aren't marketing in disguise.

How the code is organized
pyre/                    # monorepo (pnpm + Turborepo)
├── apps/
│   ├── desktop/           # Electron + native runtime
│   └── web/               # getpyre.ai dashboard
├── packages/
│   ├── pyre-engine/             # orchestrator + provider routing
│   ├── pyre-context-budget/     # context-window allocator
│   ├── pyre-tool-vault/         # vetted tool catalogue
│   ├── pyre-compaction-sidecar/ # async scrollback summarizer
│   └── pyre-vram-calc/          # hardware capability detection
└── turbo.json             # Turborepo pipeline

# MCP servers live in their own repos (independently installable):
@onenomad/engram-mcp     # github.com/OneNomad-LLC/engram-mcp
@onenomad/persona-mcp    # github.com/OneNomad-LLC/persona-mcp
What a plugin looks like — SKILL.md
---
name: pr-reviewer
description: Reviews pull requests for design, security, and correctness
license: Apache-2.0
---

# PR Reviewer

You review pull requests with the rigor of a senior engineer. You
read every file before commenting. You cite line numbers. You
prioritize findings.

## Triggers

- The user asks for a code review, second opinion, or pre-merge check.
- A PR URL or diff is in the request.

## Tools

- read_files
- run_bash      # for git operations
- search_web    # for security advisories

## Output format

Group findings by severity (blocker / major / minor / nit). For
each: file:line, what's wrong, suggested fix. No prose preamble.

OpenClaw-compatible. Drop the file in your project; Pyre loads it. Same format runs in Claude Code, Cursor, and any MCP client that speaks skills.

Architecture deep-dives

Cognitive stack, multi-provider router, capability detection, persona evolution model — all written down. No paywall, no "contact sales."

// Tiers

Free locally.
Pay for continuity.

Open-core funnel: free local runtime → paid cloud features → enterprise contracts. Same product, three deployment shapes.

Most popular
Pyre · Core
Personal · Free forever
$0Apache 2.0 · open-core
Install
  • Full local runtime — every model, every provider
  • Single-machine personal memory (Engram) + personality (Persona)
  • Multi-agent orchestration up to your hardware limit
  • Desktop · CLI · Web server · Chrome companion
  • Local cognitive stack — 200K effective context on 16 GB GPU
Pyre · Pro
Cloud · Compounding personal AI
$20per month · or $200/yr
Start free trial
  • Everything in Core
  • Your Persona, personal memory, and project context follow you across every device — E2E encrypted
  • Always-on background agents — long sessions keep running when your laptop sleeps
  • Nightly E2E-encrypted backups — restore your AI on any new machine in minutes
  • Full memory portability — export your Persona and Engram any time. Your data, your call.
  • Curated plugin & SKILL.md catalog
Pyre · Enterprise
Licensed · On-prem
CustomAnnual contracts
Talk to us
  • Everything in Pro
  • On-prem deployment — Terraform module, Helm chart, air-gap installer
  • SSO / SAML — Okta, Azure AD, Google Workspace
  • Audit logging, RBAC, configurable retention
  • Cortex Enterprise — managed company-memory ingestion
  • 99.9% SLA + implementation services
// Get Pyre

One command. Any machine.

Pyre auto-detects your hardware on first launch — CPU, Mac Metal, CUDA, ROCm, or Vulkan — and picks a model that fits. No config files, no ceremony.

terminal
# macOS / Linux
curl -fsSL https://getpyre.dev/install.sh | sh

# Or via npm
npm install -g @onenomad/pyre

# Run
pyre start
// Native installersAll releases →
Windows
Pyre Desktop
Installer.exe
Linux
Pyre Desktop
AppImagex86_64