Open source · Free for individuals · macOS

An agent that outlasts
every subscription.

Permagent is an open-source agent OS that runs on your Mac. It chats, executes, browses, and schedules — and it stores everything it learns in a memory that lives on your disk, not someone else’s cloud. It doesn’t reset between sessions. It compounds.

v1.31

runtime shipping today

25+

model providers, incl. local

100%

of memory stays on your disk

0

required subscriptions

What permagent is

One agent. Yours. Permanent.

Most AI tools are rented. They reset every session, live on someone else’s hardware, and answer to someone else’s roadmap. Permagent is the opposite: a single agent that runs on your machine, with your keys, on any model you choose — and a memory that makes it more valuable every week you use it.

The code is open source. The data format is yours. If we disappeared tomorrow, your agent — and everything it knows — keeps working.

01 · The appShipping

A full operating environment

Chat with persistent memory, a real terminal, a native browser, learned skills, and scheduled automations. Free for individuals. Your machine, your keys.

02 · The Mesh2027 · roadmap

A network agents earn their way into

Verified agents share knowledge, then compute — pooling hardware to run open models no single machine could. Identity-gated by Chitin ID. First 300 seats form the Founding Senate.

What it does today

Five things your agent
does right now.

Not a teaser list. Every item below is in the v1.31 runtime.

01

Chat with a permanent memory

Every conversation lands in Spectral — an on-device memory engine with a knowledge graph. Your agent on Monday remembers what it learned on Friday.

02

Built-in terminal + browser

A real PTY terminal and a native webview, side by side with chat. The agent executes, browses, and reports in one window.

03

Skills it learns from you

The runtime watches for repeated work. Do something twice and Permagent proposes a reusable skill — yours to keep, edit, and version.

04

Any model. Your keys.

25+ providers — Anthropic, OpenAI, Bedrock, Vertex — plus fully local inference through Ollama and llama.cpp. Switch any time. No middleman markup.

05

Automations that run while you sleep

A cron scheduler is built into the runtime. Recipes run overnight and present findings, not logs. (Scheduling UI lands in Phase 2 — today it's API and CLI.)

AnthropicOpenAIOllama · localllama.cpp · localAWS BedrockGoogle VertexAzureDatabricksGeminiOpenRouterxAIKimiAnthropicOpenAIOllama · localllama.cpp · localAWS BedrockGoogle VertexAzureDatabricksGeminiOpenRouterxAIKimi

Inside the app

Not a chat wrapper.
An operating environment.

The Workbench

Shipping

Chat. Execute. Browse. One window.

Split panes let you watch the agent work in real time — conversation on one side, the terminal or browser it's driving on the other. No copy-pasting between apps.

Chat. Execute. Browse. One window.

The Brain · Spectral

Engine shipped · view in Phase 2

It never forgets.

Every conversation and decision lands in Spectral memory on your disk. The engine ships today; the Brain view — browse, search, and curate what your agent knows — is the headline of Phase 2.

It never forgets.

Automate · Recipes

Runtime shipped · UI in Phase 2

It works while you sleep.

The cron scheduler and recipe engine are in the runtime now — schedule workspace cleanup, storage insights, project health checks. The visual scheduling UI arrives in Phase 2.

It works while you sleep.

The Lab · World View

In development

Watch your agents at work.

A 3D marble pantheon where your agents appear as characters — the orchestrator directing, the Librarian shelving memories in a mezzanine library. In active development; the render below is from a development build.

Watch your agents at work.

Spectral · the federated brain

One brain per agent today.
One brain per team tomorrow.

Spectral is Permagent’s open-source memory engine — and it was designed for federation from the first commit. The long-term shape: many Permagents contributing memory shards into a shared master brain. A startup’s collective context. A community’s shared knowledge. Every member’s agent, recalling from the same well.

Shipped today

In v1.31

Frequency-domain recall

Deterministic fingerprints instead of embeddings. Sub-millisecond recall, no GPU, no embedding API bill.

A real knowledge graph

Typed entities and relationships with multi-hop traversal — “how does X relate to Y” is a query, not a vibe.

Federation primitives, shipped

Every brain has an Ed25519 identity. Every fact carries provenance — which brain asserted it — and a visibility level. The annotations federation needs are already on your data.

Concept — shards converging into a master brain

On the roadmap

Federation protocol

Memory shards

Mark memories shareable and they become signed shards — portable, attributable units of knowledge.

The master brain

A team, a community group, a research collective: many Permagents contribute shards into one shared brain that every member's agent can recall from.

Provenance all the way down

A shared fact always knows which agent asserted it. Trust, dispute, and supersede — at the level of individual memories.

Straight talk: the federation protocol — sync, conflict resolution, merge — is not built yet. The identity, provenance, and visibility primitives it needs shipped in v1.31, so the data you accumulate today is already annotated for the brain your team shares tomorrow.

The Mesh

Vision · published design, not yet in build

Your hardware decides what you contribute — not what you can access.

A 16GB MacBook can’t run a 70B open-source model. Four Mesh members pooling their Apple Silicon can. The Mesh turns local-first from an individual constraint into a collective advantage: members contribute GPU time to a shared pool and get access to models no single machine could hold — with no per-token API bill and no cloud middleman reading your data.

Identity comes from Chitin ID — one agent, one identity, reputation on-chain. Coordination assigns model shards to peers and verifies results. Privacy is the hard layer, and we’ve published exactly how we plan to climb it.

Concept — devices pooling compute around one shared model

What never leaves your machine

  • Your Brain (Spectral DB) — never leaves
  • Your documents — never leave
  • Your memories — never leave
  • Raw prompt text — encrypted before leaving, or never leaves

The only thing that travels is the inference computation itself — and each phase of the roadmap hardens how that computation is protected, from trusted circles to hardware attestation to fully homomorphic encryption.

Phase 2Next up

Knowledge Mesh

The Mesh launches as a knowledge network: verified agents share skills and memories under Chitin identity. Text, not compute.

Phase 3Designed

Inference routing

One interface for “where does inference run”: local, Mesh peer, or cloud fallback. You set the policy.

Phase 4Vision

Split inference

Your Mac runs the first and last model layers; peers run the middle on abstract activations. Raw text never leaves your machine.

Phase 5Vision

TEE-attested compute

Hardware-isolated inference on peers' machines. You trust the silicon, not the stranger.

Phase 6Horizon

Encrypted inference

Fully homomorphic encryption: math no party can peek inside. The endgame — privacy that is proven, not promised.

Where this stands today: an 80-page published design (identity, coordination, split inference, TEE, FHE) and reserved fields in the shipped schema. Zero Mesh code in the runtime yet — and we won’t pretend otherwise. The Knowledge Mesh comes first.

The full Mesh plan

Build log

Shipping in the open.

Exactly where we are: what’s done, what’s being built, what’s planned. No vaporware sold as features.

Shipped

Runtime v1.31 on Apple Silicon

Spectral memory + knowledge graph

Auto-skill detection

Terminal, browser, chat workspaces

25+ model providers, local inference

Cron scheduler + recipes (API)

In progress · Phase 2

Brain view — see + search your memory

Scheduling and recipes UI

Rich chat renderer (charts, tables, files)

Voice loop (Whisper, push-to-talk)

Chitin ID v1

Founding Senate (private rollout)

Planned & vision

Public app ships · Q3 2026

Knowledge Mesh private beta · Q4 2026

Mesh general availability · 2027

Spectral federation — shared team brains

Mesh distributed inference (vision)

Governance by election · 2027+

Who it is for

Builders who want to compound.

A

Solo founders

You need an operator that works while you sleep and remembers the company by Monday.

B

Indie developers

You want tools that learn your codebase once and keep learning. No resetting every session.

C

Teams that compound

You're tired of renting intelligence. You want agents whose knowledge accrues to you — and eventually, to a brain your whole team shares.

Get in before the door closes.

Join the waitlist to hear when the Mac app ships. If you want permanent standing in the record, Founding Senator seats are opening soon.

Founding Senate

300 seats. $250 deposit, fully refundable until the Mesh launches. Opening soon through private outreach.