StateArk

Never explain your project to an AI twice.

Your chat is a workspace, not an archive. StateArk turns a working session into a versioned, portable project state that any new chat can pick up. Open terminal and write:

npx stateark

Then quit Claude Desktop completely (Cmd+Q) and reopen it. That is the whole install.

(Prefer to install from source? npm install && npm run build && npm run setup — see INSTALL.md. Undo any time with npx stateark remove.)

Here what we had in mind developing stateark - we address a problem millions of users have...

You:  ...three hours of work...
You:  Savepoint
      -> my-project v0.3 written to ~/StateArk

      (next day, new chat, empty context)

You:  Resume my-project
      -> decisions, constraints, rejected approaches, open questions,
         and the actual files - all current, none of the detours

Local by default. Savepoints are ordinary folders of Markdown, JSON and your real files. Nothing is uploaded. If StateArk disappears tomorrow, you still have everything.

~/StateArk/projects/my-project/v0.3/
  state.md        <- the canonical state, readable in any editor
  state.json
  manifest.json   <- every artifact with its SHA-256
  artifacts/
    app.py
    schema.sql

Commands

In the chat	What happens
Savepoint	consolidate the session into a new version
Resume <project>	load the latest state into a fresh chat
History <project>	list the versions
Diff <project>	what actually changed between two savepoints

In the terminal
npx stateark	register with Claude Desktop
npx stateark report	anonymised local usage summary, printed for you only
npx stateark remove	unregister; savepoints are kept

What makes it more than a summary

• Carry-forward. Files you do not re-submit are copied into the new version with their original hash, so a forgetful model cannot silently lose your schema.
• Integrity checks. StateArk cannot see your chat - so it checks what it can: truncation markers (// ... rest unchanged), files that collapsed in size, files the state describes but never handed over, savepoints identical to their predecessor. It warns, it never blocks: the savepoint is always written.
• Journal. Between savepoints the model quietly records turning points, so a Savepoint at the end of a long session does not depend on a degraded context. Only for projects you already saved once - scratch conversations are never touched.

Requirements

Node 20+. Claude Desktop, or Claude Code (claude mcp add --transport http ...).

Custom-connector UIs (Claude web, Cowork, ChatGPT web) dial your server from the vendor's cloud and cannot reach localhost. See INSTALL.md.

Licence

Elastic License 2.0. Use it for anything including commercially, read and modify the source, share the package. You may not offer it to third parties as a hosted or managed service. Your savepoints are yours and are not covered by this licence.

---

Architecture

ChatGPT / Claude / Gemini-capable MCP client
                 |
                 v
          StateArk Local Agent      <- 127.0.0.1 only, by default
                 |
                 v
        ~/StateArk/projects         <- SOURCE OF TRUTH
                 |
                 | optional
                 v
             Supabase               <- mirror / backup / transport

A Savepoint is an ordinary directory, not a proprietary database:

StateArk/projects/my-project/
  journal.ndjson      # entries recorded since the last savepoint
  project.json        # index (self-healing: rebuilt from disk if corrupt)
  v0.3/
    state.md          # human- and LLM-readable canonical state + integrity warnings
    state.json        # the same state, structured
    meta.json         # version, lineage, platform, sync status, warnings
    manifest.json     # artifacts with SHA-256, stored/pending, carried_forward_from
    journal.ndjson    # the entries this savepoint consolidated
    artifacts/
      app.py
      schema.sql
      prototype.zip

If StateArk disappears, those files remain usable.

Install

Requires Node 20+.

npm install
npm run typecheck
npm test          # 89 checks against a throwaway store — run this first
npm start

No .env is needed for local-only mode. Copy .env.example to .env only if you want to change the port, the store location, or enable the Supabase mirror.

On first run StateArk generates a random access key and stores it in ~/StateArk/.access-key (mode 0600). The startup banner prints your endpoints:

MCP:         http://localhost:8787/mcp/<key>
Upload page: http://localhost:8787/upload/<key>
Health:      http://localhost:8787/health

Security model

The HTTP entrypoint is a local server on your own machine, so StateArk assumes any web page you visit is hostile:

Control	Default
Bind address	127.0.0.1 (STATEARK_BIND to change)
Access key	random per install, in <root>/.access-key, constant-time compared
Origin header	loopback only, plus STATEARK_ALLOWED_ORIGINS
Host header	loopback only, plus STATEARK_ALLOWED_HOSTS (DNS-rebinding guard)
CORS	echoes the validated origin, never *
Upload form	single-use CSRF token, capped body size
Non-loopback bind	refuses to start unless the access key is ≥ 24 chars

Before exposing the agent over HTTPS: set a long STATEARK_ACCESS_KEY, set STATEARK_ALLOWED_HOSTS to your tunnel hostname, and put a real reverse proxy in front.

Platform reality

Local-first and hosted LLMs are different networking environments. Claude Desktop/Code and other local MCP clients can talk to the local agent directly. A hosted ChatGPT/Gemini client generally cannot reach localhost on your computer. For those you would need a secure HTTPS route to your running local agent — remote access is on the roadmap and does not exist yet.

If you are setting this up, in this order:

1. run StateArk locally;
2. test Savepoint and Resume from a local MCP client;
3. enable Supabase sync only if you actually want a mirror;
4. expose the agent over HTTPS only after the three steps above work.

Local-only mode — the default

Leave SUPABASE_URL, SUPABASE_SECRET_KEY and STATEARK_OWNER_ID unset. Cloud sync stays off and files never leave the machine. Default store: ~/StateArk (override with STATEARK_LOCAL_ROOT).

This is what you get out of the box. The Supabase section below is opt-in.

Local + Supabase sync

Run the migrations in the Supabase SQL editor, in order:

1. supabase/migrations/001_stateark.sql
2. supabase/migrations/002_artifacts.sql
3. supabase/migrations/003_hardening.sql

Then set SUPABASE_URL, SUPABASE_SECRET_KEY (server-side Secret key, never a publishable key), STATEARK_OWNER_ID, and STATEARK_STORAGE_BUCKET.

Cloud is a mirror, not the master:

• the local savepoint is committed first and a cloud failure cannot invalidate it;
• cloud deletion does not delete local data;
• sync status lives in meta.json (disabled / pending / synced / failed);
• retry a failed push with the sync_savepoint tool.

Text/code artifacts up to 2 MB are mirrored inline in Postgres; everything else goes to the private Storage bucket. The mirror is not end-to-end encrypted — the Supabase project can read what it stores. Sync is off by default for exactly that reason; turn it on only for a project you would be comfortable putting in any hosted database.

Savepoint behaviour

When you say Savepoint, the host LLM should:

• reconstruct the latest valid state rather than summarise chronology;
• retain governing decisions, requirements and constraints;
• keep rejected approaches only when the reason matters;
• send exact text/code as transfer=text;
• send binary bytes as transfer=base64 only when truly available;
• otherwise mark the artifact transfer=pending instead of fabricating it;
• omit files that have not changed — they are carried forward automatically;
• never abbreviate a file with ... rest unchanged.

Local creation is atomic: StateArk writes a temporary bundle and renames it only when complete. Artifact names are sanitised; if a name had to be changed, manifest.json records the original under original_name.

Integrity checks

Every savepoint is compared against its predecessor. Findings are returned in the tool result and rendered at the top of state.md. The savepoint is always written — a check never costs you work, it only tells the model to come clean.

Code	Meaning
artifact_carried_forward	file was not re-submitted and not declared deleted, so it was copied from the previous version
artifact_became_pending	content that was stored is now only pending
artifact_shrank	file collapsed to under 50% of its previous size
truncation_marker	stored text contains ... rest unchanged, [...], … gekürzt and similar
no_change	this savepoint is identical to the previous one
journal_not_reflected	decisions were journalled but the submitted state lists none
artifacts_mentioned_but_missing	the state text names a file (schema.sql, app.py, …) that was never handed over as an artifact

The last one is the check that matters most in practice. StateArk cannot read your chat, so a filename appearing in the prose is the only available clue that the model described a file instead of preserving it. The warning is phrased as a question, because a named file may legitimately be planned or live outside the conversation.

Journal

note_event records one short line per turning point, silently, and only for projects that already exist in StateArk. Untracked projects return tracked: false and nothing is written — that is the filter that keeps scratch conversations out.

At the next Savepoint the journal is consolidated into the new version and cleared. It is never cleared before the savepoint is durably on disk.

Inspect pending entries with the journal tool. Nudging is based on entry count (STATEARK_JOURNAL_NUDGE_AT, default 20), never on a timer.

Diff

diff_savepoints with no version arguments compares the latest savepoint against its predecessor: LCS line diff on the text fields, set diff on the lists, SHA-256 comparison on the artifacts, plus an attention block for anything that looks like silent loss.

Artifacts

Every stored artifact has a SHA-256. Pending files are explicit in manifest.json and in the resume_project output, so the LLM cannot quietly pretend a file exists.

If a binary cannot travel through MCP, open the local upload page printed at startup and attach it to the existing project/version. This resolves the pending manifest entry.

Resume

Resume My Project returns state.md plus the artifact manifest. The LLM calls get_artifact for exact text only when needed. For binaries it gets the verified local path and hash rather than invented content.

MCP clients

• Claude Desktop — npx stateark registers the stdio entrypoint for you.
• Claude Code — point it at the local Streamable HTTP endpoint: claude mcp add --transport http ...
• ChatGPT web/mobile, Gemini, and other hosted clients — these dial your server from the vendor's cloud and cannot reach localhost on your machine. That is a networking fact, not a missing feature. Reaching your savepoints from a hosted client needs an authenticated HTTPS route to your own machine; it is on the roadmap and does not exist yet.

Roadmap

Nothing here exists yet. In roughly this order:

• Optional cloud storage — your savepoints mirrored so a dead laptop is not a dead project. Off by default, and only ever a copy: your disk stays the original.
• Remote access — reaching your state from web and mobile clients, not just the desktop app on the one machine that holds it.
• Branching — two chats working the same project currently produce one linear chain of versions. Documented, not solved.
• State pruning — over many savepoints state.md grows without limit. It needs a size ceiling and a rule for what ages out.

StateArk is free to use, including commercially. See the licence above for the one thing it does not allow.