User-owned shared memory for AI agents. A persistent, curated knowledge layer that multiple AI tools can read and write, backed by Postgres + pgvector.
Cerefox — user-owned shared memory for AI agents. CLI + stdio MCP server + web UI + ingestion for a knowledge base on your own Supabase project (or fully self-hosted with Cerefox Local).
2026-11-13
ดาว
—
เครื่องมือ
—
อัปไทม์
—
What is Cerefox?
Cerefox is a user-owned knowledge memory layer: a persistent, curated knowledge base that sits between you and the AI tools you use.
The primary use case is shared memory across AI agents: knowledge written by one tool (Claude, ChatGPT, Cursor, or a custom agent) becomes immediately available to all others. This prevents context fragmentation, so the same information doesn't have to be re-explained in every session.
Cerefox is asynchronous shared memory, not a message bus. It solves the persistent context problem: knowledge written in one context is findable in any other. A user curates project documents and an AI agent discovers them through search without being told they exist. An agent writes a decision during a coding session and a different agent, on a different machine, running a different model, finds it days later. A user switches from one AI tool to another and the accumulated knowledge carries over without manual transfer. The boundaries that Cerefox dissolves are between agents, between sessions, between human and machine, and across time.
For the full project vision, principles, and roadmap direction, see docs/research/vision.md.
Agent-first, not human-first: AI agents are first-class citizens on both sides: they read and write; humans curate and validate
Own your data: everything lives in a Postgres database you control (Supabase free tier or self-hosted)
Cross-agent coordination: agents on separate machines and runtimes coordinate through persistent shared context (see docs/guides/agent-coordination.md)
Not a note-taking app: Cerefox is knowledge infrastructure, not a replacement for Obsidian, Notion, or Bear; those tools handle authoring, Cerefox handles indexing and agent access
Hybrid search: full-text + semantic search finds relevant knowledge even with fuzzy or conceptual queries
Any agent, anywhere: local MCP (Claude Code, Codex, Cursor, Gemini, Claude Desktop); remote MCP via Supabase Edge Functions; cloud Claude over OAuth; ChatGPT via Custom GPT + GPT Actions until OpenAI retires Custom GPTs on December 11, 2026 (see Connecting AI agents)
Keep it cheap: Supabase free tier + low-cost cloud embeddings; see docs/guides/operational-cost.md
Example use cases
Cerefox isn't bound to one tool or one workflow — anything that can run a shell command or speak MCP can read and write your memory, pointed at either a shared cloud deployment or a private local one:
See Connecting AI agents for the how-to per client.
Community
Questions, ideas, or want to follow development? Join the Cerefox Discord — setup & usage help for Cerefox and cf², release news, and a place to share what you're building.
Features
Feature
Details
Hybrid search
Combines full-text (BM25) + semantic (vector) search with a configurable alpha weight
Metadata-filtered search
JSONB containment filter (@>) on document metadata; server-side, GIN-indexed; composable with project filter and all search modes; available across all access paths (MCP, CLI, web UI, GPT Actions)
Metadata search
Standalone metadata-only search (no text query needed); find documents by key-value criteria, project, and date range; optional content inclusion with byte budget; dedicated MCP tool, CLI command, and web UI page
Project discovery
cerefox_list_projects MCP tool for agents to discover available projects; all search results include human-readable project_names alongside UUIDs
Heading-aware chunking
Greedy section accumulation — H1/H2/H3 sections accumulate until MAX_CHUNK_CHARS; heading breadcrumb preserved per chunk
Embeddings
OpenAI text-embedding-3-small (768-dim) via API — or, on Cerefox Local, an optional fully-offline local model (nomic-embed-text-v1.5, in-container ONNX): no API key, text never leaves your machine
Remote MCP endpoint
cerefox-mcp Supabase Edge Function — MCP Streamable HTTP; connect Claude Desktop, Claude Code, or Cursor with just a URL and a Cerefox access token (cerefox token generate); no Python install needed
Local MCP server
cerefox mcp stdio server (TypeScript, from @cerefox/memory) -- local alternative with zero Edge Function usage, lower latency, and offline support; npm install -g @cerefox/memory.
Web UI
React + TypeScript SPA (Mantine UI) at /app/; Hono (TypeScript) JSON API backend served by cerefox web; Markdown viewer, search with 4 modes, document editing, project management
Markdown-first ingest
.md / .txt / .docx (Markdown is the storage format; .docx is converted via mammoth on ingest, fidelity varies. PDF is not supported — convert upstream)
Batch ingest
cerefox document ingest-dir recurses directories
Deduplication
SHA-256 content hash; re-ingesting the same file is a no-op
Partial document edits
cerefox_insert / cerefox_edit change part of a document without resending it — append, replace a section, delete one, or rename a heading. Anchors are heading paths; an ambiguous anchor is refused with the options rather than guessed. Multiple operations apply atomically.
Read one section
cerefox_get_document(section: "## Heading") returns exactly the text a replace_section would overwrite, resolved by the same code as the write — so an edit need not start with a full read.
Metadata writes
cerefox_set_document_metadata changes tags without touching content. Merges by default (a null value removes a key), so concurrent agents setting different keys do not clobber each other. No re-chunk, no re-embed, no new version.
Agent delete + restore
cerefox_delete_document (soft, requires the caller's read-hash, reason recorded in audit) and cerefox_restore_document (its audited inverse) — v1.7.0. Permanent purge stays web-UI-only with human confirmation.
Link integrity
Every write validates [Text](uuid) document links and rejects links to nonexistent ids (mangled-UUID protection — LLMs regenerating text corrupt long random ids at a structural rate); examples escape via code formatting; [[wikilinks]] may dangle; cerefox document dead-links sweeps for legacy dead links. See docs/guides/linking.md — v1.7.x
Concurrency-safe updates
Optimistic locking on content updates (v0.11+): writers pass the content_hash they read; a concurrent change fails with a conflict (re-read → merge → retry) instead of silently overwriting another agent's work. Explicit last_write_wins opt-out for file re-sync flows
Backup and restore
JSON snapshots, optional git commit
Small-to-big retrieval
cerefox_context_expand RPC returns chunk neighbours for richer context
Audit log
Immutable, append-only log of all write operations (create, update, delete, restore, status change). Author attribution with author_type ('user' or 'agent'). Browsable via web UI, queryable via MCP tool and Edge Function
Review workflow (optional)
Store-level switch review_workflow_enabled (off on a fresh install). On: agent writes land pending_review for a person to approve; filterable on search. Off: no surface shows or enforces a review status (the status is still recorded, so turning it back on shows what was there). Never gates retrieval
Version governance
Version archival (protect specific versions from cleanup), configurable retention (cerefox config set version_cleanup_enabled), version diff viewer
Lean version storage
Archived chunks keep their content (the safety copy) but carry no embeddings or FTS (v1.8.0) — typically 30–45% of chunk storage reclaimed
Usage tracking
Opt-in logging of all operations (reads and writes) across all access paths. Tracks operation type, access path (remote-mcp, local-mcp, edge-function, webapp, cli), requestor identity, query text, and result count. Controlled via cerefox config set usage_tracking_enabled true/false -- no redeploy needed
Analytics dashboard
/app/analytics -- 8 interactive charts: calls per day, access path breakdown, top documents, top readers, operations donut, requestor word cloud, requestor→document access patterns (HEB), and requestor→operation patterns (HEB). Date range + project + path filters. CSV export.
Project status
As of v0.10.0 Cerefox runs two ways: against a hosted Supabase project,
or fully local / self-hosted in a single Docker container (no cloud, no
account). The whole runtime — CLI, MCP server, web UI, ingestion, and
server-side deploy — ships in the
@cerefox/memory npm package
(no repo clone); the local backend bundles that same runtime plus Postgres +
pgvector into one image.
Until v1.0.0 the SemVer policy in CONTRIBUTING.md is
aspirational — breaking changes can land in minor versions when there's a good
reason; after v1.0.0 it's binding. Full release history is in
CHANGELOG.md; the roadmap and iteration log live in
docs/plan.md.
Getting Started
Upgrading to v1.0.0? Two changes need action (a rotatable access token
replaces the legacy anon JWT for Edge Functions, and a schema redeploy) and
Python is fully removed. One guide covers it:
docs/guides/migration-1.0.md.
Cerefox runs two ways — pick your backend. Both expose the same features,
web UI, and MCP tools; they differ only in where your data lives and how you
install. (Contributors who want to run from source: see Run from source below.)
Option 1 — Cloud (Supabase)
Your data lives in your own Supabase project (free tier is enough). You use
the cerefox command from the @cerefox/memory
npm package. No git clone, no Python, no build.
# 1. Install (one-liner; detects Bun, falls back to npm):
curl -fsSL https://github.com/fstamatelopoulos/cerefox/releases/latest/download/install.sh | sh
# or: npm install -g @cerefox/memory (Node ≥ 24)
# 2. Configure + stand up the server side (against your own Supabase project):
cerefox init # interactive setup: Supabase URL/keys, embedding key
cerefox server deploy # schema + RPCs + all 9 Edge Functions, from the npm bundle
cerefox token generate # mint the Edge Function access token; sets it on Supabase
# and writes CEREFOX_ACCESS_TOKEN to your .env. Needed for
# GPT Actions / remote MCP + a fully-green doctor; the local
# MCP, cloud Claude (OAuth) and CLI/web don't use it.
cerefox doctor # verify everything is wired up
# 3. Wire up your AI agent(s) — run the ones that apply:
cerefox configure-agent --tool claude-code # local MCP (preferred) — also: claude-desktop | cursor | codex | gemini
# Cloud Claude (claude.ai/app): connect over OAuth (setup-supabase Step 7).
# Custom GPT (ChatGPT): paste the token from step 2 into the Action's
# Authentication → API Key (Bearer). See docs/guides/connect-agents.md.
# Works until 2026-12-11, when OpenAI retires Custom GPTs (see below).
# 4. Use it:
cerefox document ingest my-notes.md --title "My notes"
cerefox search "what did I decide about auth?"
cerefox web # web UI → http://localhost:8000/app/
Prerequisites: Node 24+ or Bun 1.0+ · a Supabase account (free tier) · an
embedding API key (OpenAI text-embedding-3-small).
Everything runs in one Docker container on your machine — Postgres + pgvector
the Cerefox server. No Supabase account, no Node/Bun on the host, just
Docker. You get a cerefox-local command (same KB verbs as cerefox).
# 1. Install (one-liner; pulls the all-in-one image, adds a `cerefox-local` command):
curl -fsSL https://github.com/fstamatelopoulos/cerefox/releases/latest/download/install-local.sh | sh
# Fully offline instead? add `-s -- --local-embedder` — embeddings run
# in-container (no OpenAI key; text never leaves your machine).
# 2. Set your OpenAI key (or pick the local embedder) + wire up an AI agent:
cerefox-local init # OpenAI key or [2] Local embedder (re-creates the container)
cerefox-local configure-agent # wire an MCP client (e.g. Claude Code)
# 3. Use it:
cerefox-local document ingest my-notes.md --title "My notes"
cerefox-local search "what did I decide about auth?"
# web UI → http://localhost:8000/app/ (or the port the installer chose — it auto-steps
# to 8010/… if 8000 is busy; `cerefox-local status` shows the URL. Manage: status | upgrade | stop)
Prerequisites: Docker (Docker Desktop or Colima)
· an OpenAI API key (embeddings still use the OpenAI API).
Full walkthrough:docs/guides/setup-local.md.
Cloud and local are independent — different installer, different command name —
so they never collide if you happen to run both.
Run from source (contributors)
Clone the repo and run from source. bun drives everything.
git clone https://github.com/fstamatelopoulos/cerefox.git && cd cerefox
bun install # workspace deps: root + packages/memory + frontend
cp .env.example .env # fill in Supabase URL/keys + embedding key
bun scripts/db_deploy.ts # schema + RPCs (--dry-run to preview · --reset to wipe first)
npx supabase functions deploy # Edge Functions (or just use `cerefox server deploy`)
cd frontend && bun run build && cd .. # build the SPA `cerefox web` serves at /app/
bun test # run the suite (root + packages/memory + _shared)
Python is gone. As of v1.0.0 the entire runtime (CLI, MCP, web,
ingestion) is TypeScript in @cerefox/memory; the Python implementation —
including the uv run cerefox mcp fallback — was removed. Only the SQL schema
assets under src/cerefox/db/ remain (they are not Python). See
docs/guides/upgrading.md.
The fastest path is cerefox configure-agent --tool <client> — it writes the
right config for Claude Code, Claude Desktop, Cursor, Codex, or Gemini. There
are several ways an agent can reach Cerefox:
1 — Local stdio MCP (recommended for local agents).cerefox mcp runs the
same 15 core tools in-process (plus 4 document-relation tools, hidden until you enable them with cerefox config set relations_enabled true) — lower latency, no per-call Edge Function billing, and
no Edge Function token needed (it reaches Supabase over the Data API with its own
.env credential). configure-agent wires it up, or point your client at
command: "cerefox", args: ["mcp"].
2 — Remote HTTP MCP (advanced/fallback). The cerefox-mcp Edge Function
speaks MCP Streamable HTTP. Just a URL + a Cerefox access token
(cerefox token generate — see
setup-supabase.md). No local install:
3 — OpenAI: Codex today, ChatGPT changing.Codex (the CLI, and Codex
threads in the ChatGPT desktop app) works with the local stdio MCP:
cerefox configure-agent --tool codex. ChatGPT chats connect today through a
Custom GPT with GPT Actions pointing at the Edge Functions (requires ChatGPT
Plus): paste the OpenAPI block from
connect-agents.md and set the Action's Bearer
auth to your Cerefox access token. OpenAI retires Custom GPTs on December 11,
2026, and custom actions do not carry over to its replacement (plugins), so
this path stops working then. ChatGPT will need a new connection: an MCP
connector to cerefox-mcp over OAuth, the way cloud Claude connects below. That
is being validated in
#326
(research and plan).
4 — Shell CLI. Local coding agents with a Bash tool (Claude Code, Codex,
opencode, …) can read and write Cerefox by running the installed cerefox
command directly — no MCP config at all. Point the agent at
AGENT_GUIDE.md and let it use cerefox search /
cerefox document ingest.
5 — Cloud & mobile Claude (optional). claude.ai web and the Claude mobile
app connect to cerefox-mcp over OAuth — memory in your browser and on your
phone, with full hybrid search. It's an opt-in extra: it needs a one-time
Supabase OAuth setup plus a hosted consent page (the repo ships a one-command
Cloudflare Worker for that, free). Nothing else
above requires it. See
setup-supabase.md → Step 7.
Full setup for every client — plus a manual per-client config appendix for when
configure-agent can't reach a tool — is in
docs/guides/connect-agents.md.
Cerefox — user-owned shared memory for AI agents. CLI + stdio MCP server + web UI + ingestion for a knowledge base on your own Supabase project (or fully self-hosted with Cerefox Local).
คัดลอกคอนฟิก stdio สำหรับ Cursor, Claude Desktop, Claude Code หรือ VS Code ได้ในหน้านี้ คอนฟิกนี้ใช้การติดตั้งของผู้เผยแพร่ ไม่ใช่พร็อกซีของ PlugMCP
@cerefox/memory ต้องใช้ API key ไหม?
ระเบียนนี้ไม่ได้ระบุตัวแปรสภาพแวดล้อมที่จำเป็น
@cerefox/memory ใช้ทรานสปอร์ตแบบใด?
@cerefox/memory ทำงานแบบโลคัลผ่าน stdio
@cerefox/memory ถูกตรวจสอบล่าสุดเมื่อใด?
ตรวจสอบล่าสุดเมื่อ 2026-11-13 Cerefox — user-owned shared memory for AI agents. CLI + stdio MCP server + web UI + ingestion for a knowledge base on your own Supabase project (or fully self-hosted with Cerefox Local).
