◆ Tokencrumbs v0.1 · Public Preview
Local-first Open source MIT licensed
A protocol for AI memory, not a moat

See what your AI sees.
Own what your AI forgets knows.

Tokencrumbs is the user-owned memory layer between AI providers. Export from any. Import to any. You hold the source of truth — not OpenAI, not Anthropic, not Google.

Project setup Database schema Auth questions Deployment ● ↑ live tokencrumb trail
§ 01 — The problem

Memory is the new lock-in.

AI assistants increasingly remember things about you. That's useful. It's also a quiet capture mechanism — and most of it is invisible.

PROBLEM 01

Memory is opaque.

You don't know what your assistant has stored, why a response feels personalized, or which memories shaped which answer. When the model gets something weirdly wrong, there's no trail to inspect.

PROBLEM 02

Memory is siloed.

What ChatGPT learns about you stays in ChatGPT. Switching providers means re-training from zero — even though the underlying facts about you haven't changed.

PROBLEM 03

Transfer is lossy.

Where memory portability exists today, it's manual, one-way, and incomplete. No version history. No diff. No canonical source of truth. Things drop. Things contradict. You can't tell which.

PROBLEM 04

Trust is concentrated.

The provider that holds your memory holds the keys to your AI experience. Policy changes, deprecated features, raised prices, breaches — you're exposed to all of it, and you can't take your context with you.

§ 02 — Live demo

Watch it work.

Hover the highlighted text to see which memories the model drew on, where each one came from, and how Tokencrumbs holds the canonical version.

claude.ai · with tokencrumbs overlay
○○○
You
Help me write a deploy script for my side project.
Claude
Sure — since you're working on your Rust CLI tool and deploying to your Hetzner box, I'll write a bash script that builds with cargo, scp's the binary over, and restarts the systemd unit. Want me to use your usual concise style with no excessive comments?
You
Yes please. And remember it's the new project — not the old gRPC one.
Memories used
3 of 47 in canonical ledger
Rust CLI project
Building a markdown-to-pdf converter
↳ from ChatGPT · Mar 12
Hetzner deployment
Single VPS, systemd, no Docker
↳ from Claude · Apr 03
Style preference
Concise, no over-commenting
↳ from Tokencrumbs profile
§ 03 — Three ways in

Pick your surface.

Tokencrumbs is a protocol with multiple front doors. Choose the one that fits — they all share the same canonical ledger underneath.

For users

Browser extension.

Drop-in overlay for ChatGPT, Claude, Gemini, and Open WebUI. See your memory trail in real time. Export anywhere. Local-first. No account required. Free forever.

Install for Chrome / Firefox →
For developers

Open-source SDK.

Drop Tokencrumbs into your own AI app in 10 lines. Render the trail, read and write the canonical format, ship transparency by default. MIT licensed.

View on GitHub →
For teams

Self-host MCP.

A Tokencrumbs MCP server you run yourself or have us host. Expose your canonical context to any MCP-compatible model. Audit-ready. Compliance-friendly. SOC 2 in progress.

Read the docs →
§ 04 — How it works

A coordination layer, not a clone.

Tokencrumbs sits between you and your providers. It reads conversations locally, segments them into topical crumbs, and maintains a clean, dated, deduplicated memory ledger on your device.

When you switch providers — or use multiple at once — Tokencrumbs generates the right import payload for each one, formatted to that provider's quirks and limits. The canonical version stays with you.

Nothing is injected into your prompts. Nothing leaves your device unless you turn on encrypted sync. Everything is auditable. Everything is portable.

We never inject into your prompts. We never see your data unless you opt in. The architecture enforces this — not a privacy policy.
01. Conversations across providers
02. Tokencrumbs reads them locally
03. Canonical ledger on your device
04. Encrypted sync — optional
05. Generate import for any provider
§ 05 — Twelve-month plan

Each surface feeds the next.

The extension creates awareness. The SDK builds developer mindshare. The MCP captures the technical early adopters. The hosted app converts the highest intent. Enterprise lands when the brand is established and regulation creates pull.

Month Phase What ships
M0 Domain + landing Live demo on the homepage. Pitch before product.
M1–3 Browser extension Local-only canonical ledger. ChatGPT, Claude, Gemini support. Goal: 10k installs.
M2–5 Open-source SDK JS/TS library, MIT license. GitHub stars become the developer pipeline.
M4–6 MCP + plugins Self-hostable MCP server. Open WebUI and LibreChat plugins. Capture prosumer/homelab.
M6–9 Hosted sync (paid) First paid tier — encrypted cross-device sync. Convert highest-intent free users.
M9–12 Tokencrumbs Chat Standalone BYOK app with the full vision. Begin enterprise conversations.
M12+ Enterprise & compliance SSO, audit export, on-prem. EU AI Act enforcement creates forced demand.
§ 06 — Monetization

Four streams, layered.

Not picked — stacked. Each tier serves a different audience and unlocks as the surfaces come online. The free tier is permanent.

Free
$0
forever
  • Browser extension
  • Local canonical ledger
  • SDK access
  • Community support
  • Self-host the MCP
Team
$20
per seat / month
  • Everything in Pro
  • Shared workspaces
  • Admin controls
  • Retention policies
  • Slack & email support
Enterprise
Custom
talk to us
  • SSO / SAML
  • Audit log export
  • On-prem deployment
  • SLA + dedicated CSM
  • SOC 2 / EU AI Act ready
◆ Manifesto ◆

AI assistants are becoming infrastructure.

Infrastructure you can't inspect is infrastructure you can't trust.

Infrastructure you can't take with you is infrastructure that owns you.

Tokencrumbs is the inspection layer · and the portability layer