The problem
Every AI forgets you when the session ends. Your preferences, your stack, the mistakes it already made — gone. Next session starts from zero, in every tool.
SEED fixes this. Your AI gets one persistent soul: it loads who you are at the start of each conversation, learns from corrections, writes a daily log, and builds a knowledge base — the same soul in every MCP client you use.
Quick start
-
Sign in at seed.abhibots.com/login, create your soul and copy your API key from the dashboard.
-
Run:
npx seed-protocol-mcp setup YOUR_API_KEY
This adds SEED to every MCP client it finds (Claude Desktop, Claude Code, Cursor, Windsurf, VS Code/Cline).
-
Restart your editor and start a chat. On a new soul the AI's first job is build_soul: it reads what already exists on your machine — Claude Code / Cursor history, or an OpenClaw agent's notes — shows you a preview, and uploads only if you say yes. After that it calls load_soul at the start of every conversation.
Prefer the terminal? Same thing, same consent question:
npx seed-protocol-mcp import claude-code
Hybrid mode — one soul on every device, synced on your terms
npx seed-protocol-mcp setup --hybrid YOUR_API_KEY
The soul lives on this machine, and only the parts you tick in the dashboard (Privacy → Devices & sync: memories, corrections, traits, knowledge, daily logs) are copied to your cloud soul — so your laptop, a VPS agent and every AI tool share them. Reads merge both. Offline is fine: writes wait locally and go up at the next start or with npx seed-protocol-mcp sync.
Local-only mode — no account, nothing sent
npx seed-protocol-mcp setup --local
Every tool works the same, but the soul lives in ~/.seed/soul.sqlite on your machine and the connector never calls the SEED server. No API key, no dashboard. Switch to online later with setup YOUR_API_KEY. (Same thing by hand: set SEED_MODE=local in the MCP config's env.)
{
"mcpServers": {
"seed-protocol": {
"command": "npx",
"args": ["-y", "seed-protocol-mcp@latest"],
"env": {
"SEED_API_KEY": "YOUR_API_KEY",
"SEED_API_URL": "https://seed.abhibots.com"
}
}
}
}
MCP tools
| Tool | When the AI uses it |
|---|
load_soul | First thing in every conversation — who the AI is (it speaks as the soul, not as a generic assistant), who you are, the rules (corrections it must never repeat), your strongest and newest memories, the last log and your workspace. On a new soul it returns first_run |
build_soul | First job on a new soul (or when you ask to import) — builds it from this machine's history: preview first, upload only after you agree, re-runs send only what changed |
get_soul_state | Current emotional state, learned traits and counts |
evolve_soul | Record a stable behaviour pattern as a trait (e.g. likes-concise-answers) |
remember | Save a durable fact, preference or decision |
recall | Search saved memories (hybrid keyword + semantic) |
correct | The user corrected the AI — saved forever, escalates if repeated |
get_skills | List the skills attached to the soul, by category |
daily_log | End of a meaningful session — what was done, decisions, next steps. Saved under the app's own name (claude-code, cursor…), so two apps never overwrite each other |
kb_write | Save reusable know-how as a knowledge-base article |
kb_search | Read knowledge-base articles |
history_search | Search your past AI sessions in the local brain — never calls the network |
search | One search across the whole soul — rules, memories, workspace documents and the sessions you brought into SEED (local-only mode: this machine's memories and notes) |
ws_tree · ws_read · ws_write · ws_mkdir · ws_move · ws_archive | Your workspace — Markdown documents in folders that you and your AIs grow together. Edits carry the version they read, so the AI never overwrites your own change; an AI never deletes, it archives |
projects · |
The workspace, search over the cloud and reports need online or hybrid mode; in local-only mode those tools say so.
Prompt: seed_bootstrap — asks the AI to mine your recent local history for durable preferences (remember) and past corrections (correct), then write today's daily_log.
CLI
npx seed-protocol-mcp setup <KEY> # configure MCP clients + save key for the CLI
npx seed-protocol-mcp import claude-code # ~/.claude/projects/*/*.jsonl → local brain (+ your git repos → Projects cards)
npx seed-protocol-mcp import cursor # Cursor state.vscdb → local brain (best effort)
npx seed-protocol-mcp import openclaw # OpenClaw workspace notes → daily logs, memories, articles
npx seed-protocol-mcp status # key, API reachability, soul, local brain counts
npx -y seed-protocol-mcp # start the MCP server on stdio (what clients run)
Import options:
| Option | Meaning |
|---|
--yes | Upload without asking (after the local import) |
--local-only | Build the local brain only — never upload |
--days N | Upload digests for the last N days (default 30, 0 = all) |
--path P | Custom Claude Code projects dir, Cursor state.vscdb path or OpenClaw workspace |
--full | Re-read Claude Code files even if unchanged |
--key K | API key (default: SEED_API_KEY, else the key saved by setup) |
Re-running an import is safe: messages are stored once, and digests, memories and articles that were already uploaded are not sent again — only new or changed items go. That makes import openclaw --yes fine on a cron, so an always-on agent keeps its dashboard current:
*/15 * * * * npx -y seed-protocol-mcp import openclaw --yes >> ~/.seed/openclaw-sync.log 2>&1
Privacy
Plain-words policy: seed.abhibots.com/privacy · security: seed.abhibots.com/security. On the server, everything you keep is encrypted at rest with a key derived from your API key.
-
Raw history stays local. Imports write to ~/.seed/brain.sqlite (owner-only permissions). history_search reads only that file and makes no network request. Delete the file to wipe it.
-
Nothing is uploaded without consent. After an import you see a summary and the question Upload N daily logs and M preference memories to your SEED dashboard? (y/N). Only y (or --yes) uploads. --local-only never does. In chat, build_soul shows the same preview and the AI must ask you before calling it with upload=true.
-
What gets uploaded (only after yes):
- one digest per day — date, session and message counts, project folder names, and the first line of each of your prompts (truncated);
- up to 50 preference sentences you wrote yourself (lines with always / never / don't / prefer / instead / hamesha / mat karo …), as
preference memories;
- one import record with counts only.
Assistant replies, tool output, code blocks and full conversations are never uploaded.
-
OpenClaw is different on purpose: those files are the agent's own notes, so they go as written (scrubbed): memory/YYYY-MM-DD.md as daily logs (first ~19k chars), MEMORY.md bullets as memories, other memory/*.md notes and the skills committed to the workspace git as knowledge articles. Installed (untracked) skill packs, SOUL.md, USER.md and session transcripts are not read.
-
Secrets are scrubbed before anything is stored locally and again before any upload: OpenAI/Anthropic keys, SEED keys, AWS access keys, GitHub/Slack/Google/Stripe/Telegram tokens, JWTs, bearer tokens, PEM private keys, scheme://user:pass@ credentials and api_key=… / password: … style assignments become [REDACTED:<type>]. The server scrubs once more on write.
Configuration
| Env var | Default | Purpose |
|---|
SEED_API_KEY | — | Your key from the dashboard |
SEED_API_URL | https://seed.abhibots.com | API base URL |
SEED_HOME | ~/.seed | Local brain, local soul (local-only mode) and CLI config |
SEED_MODE | — | local = nothing sent · hybrid = this device + what you sync · online overrides a saved setup |
CLAUDE_CONFIG_DIR | ~/.claude | Where import claude-code looks for projects/ |
OPENCLAW_PROFILE | — | import openclaw reads ~/.openclaw-<profile>/workspace instead of ~/.openclaw/workspace |
Requires Node.js 18+. No native build — the local brain uses sql.js (SQLite compiled to WebAssembly).
Links
License
MIT
