Ken
Pushes your open decision cards to an ntfy topic (or any URL that accepts a text body) so you can answer unsure writes from your phone each morning; configured via push.url and run with ken push, with ken push --dry-run for testing.
Supports a local Ollama model as the LLM backend for Ken's judging and parsing jobs, configured with base_url: http://localhost:11434/v1, api_key: ollama and api_style: chat under llm: in the config, so no cloud model provider is required.
Uses the OpenAI API as its default model backend: the Decisions API (gpt-6-luna) to decide whether a write is new, a duplicate or a conflict and to rank undated tasks, gpt-5.6-luna as fallback judge, and gpt-5.6-terra to turn free-text tell reports into structured writes. Any OpenAI-compatible API, including local servers, can be substituted via base_url/tiers settings.
This server does not currently have a deployable release on Glama. Follow the repository's setup instructions to install and configure it for your MCP client, including any required credentials and local dependencies.
What your agent receives
Over a few chats you told Claude about the kitchen renovation; on Monday, that the plumber sent a quote. On Tuesday, you ask any agent, in any app, to work on "approve the quote from Tomas for the kitchen renovation". The task names only Tomas and the renovation. Ken adds the rest through the links (trimmed):
{
"task": "approve the quote from Tomas for the kitchen renovation",
"related": [
{"kind": "person", "name": "tomas",
"why": "named in the task",
"note": "The plumber; sent a quote on Monday."},
{"kind": "project", "name": "kitchen renovation",
"why": "named in the task",
"note": "New kitchen by December; budget 18k EUR."},
{"kind": "person", "name": "maya",
"why": "linked: maya is a member of kitchen renovation",
"note": "Your partner; owns the plumber question."},
{"kind": "asset", "name": "floor plan",
"why": "linked: floor plan is part of kitchen renovation",
"note": "The kitchen floor plan."}
],
"last_report": "Maya asked Tomas to keep the old sink"
}The agent starts with the budget, who owns the plumber question, the floor plan and the last thing that happened. You typed none of it into this chat. Each item also carries the evidence it rests on (left out above to keep it short).
Related MCP server: Memxus
What Ken stores
Built-in chat memory | Ken | |
What it keeps | notes from past chats | the current state: open, done, who owns what, what waits on what |
Old facts | can sit next to new ones | replaced; a judge catches duplicates and contradictions |
Structure | flat notes | people, projects, documents ( |
Who decides | the app saves what it wants | a judge checks every write; you decide the unsure ones |
Which apps | that one app | every MCP app reads and writes the same world |
Where | the vendor's cloud | one SQLite file on your machine; write text goes to the model you choose |
What your own apps send (Claude Code on your machine, or any client with your own token) is saved after the judge checks it for duplicates and conflicts. Ken asks you with a card, in the chat you are already in, when a write would change an existing record, when the judge is unsure, when an app saves its own guess rather than your words, or when the write comes from an agent with its own token.
Get started
Rather not install anything? See Ken Cloud.
brew install uv # skip if you have uv. Linux: curl -LsSf https://astral.sh/uv/install.sh | sh
uv tool install 'ken-engine[mcp] @ git+https://github.com/kenning-labs/ken#subdirectory=engine'
ken initken init asks three things: which model Ken uses, whether to add Ken to Claude Code, and whether
to run it in the background. Then it tells you what is ready.
Paste this into Claude Code, Codex or another coding agent:
Install Ken, a local MCP server, from https://github.com/kenning-labs/ken (README, "Get started"):
1. If `uv --version` fails, install uv: `brew install uv` on macOS with Homebrew, otherwise
`curl -LsSf https://astral.sh/uv/install.sh | sh`.
2. uv tool install 'ken-engine[mcp] @ git+https://github.com/kenning-labs/ken#subdirectory=engine'
3. Ask me which model Ken should use: my Claude Code login, or an OpenAI API key (never print a key).
Then run `ken init --claude-code --add-to-claude --service`, or `ken init --api-key <key>
--add-to-claude --service`.
4. Run `ken status` and `claude mcp list`. Tell me whether the engine is live and which server name
reaches it, then what to say to start and how to undo the install.If your agent may not change Claude Code's own settings, ken init prints the claude mcp add
command for you to run.
uv installs Ken and fetches Python 3.11 or newer if you don't have it. After installing uv, open a new terminal. After the first PyPI release, the install becomes
uv tool install 'ken-engine[mcp]'.The model checks what your apps write and reads plain-text notes: an OpenAI key, your Claude Code login, another OpenAI-compatible API, or none for now. Without a terminal, pass it as a flag:
ken init --claude-code,--api-key KEY(with--base-url URL), or--no-model.Your data is in
~/.ken, private to you.Claude Code: if it already has a server named
ken(for example Ken Cloud),ken initkeeps it and adds this engine asken-local; turn the other one off in/mcp.The background service starts at login (launchd on macOS, a systemd user service on Linux). Without it, run
kenin a terminal and keep it open. Later:ken service status,ken service uninstall, orken service install.ken statusshows whether the engine is live;ken --versionshows the version.Apps talk to the engine through
ken-mcp. Writes wait for the engine: with the engine stopped, a write returns outcomequeuedand applies when it starts, so do not send it again.
Then, in Claude (Code or Desktop), say "set up my world". Claude reads what it can reach (connected mail, calendar, files) and proposes your people, projects and documents in batches. Ken saves nothing until you say yes.
Connecting an agent. Any MCP client works: Claude Code, Codex, a script on a schedule. Give each
agent its own token under http.agent_tokens, and every write from it becomes a card for you. Only
your own token can answer a card.
The agent that calls Ken does most of the work: it reads your mail or your code, decides what matters, and sends Ken structured updates. Ken itself calls a model for a few jobs:
Job | Model (OpenAI default) |
Decide whether a write is new, a duplicate or a conflict; rank your undated tasks | OpenAI Decisions API ( |
Judge a write when Decisions cannot |
|
Turn a free-text report ( |
|
You have | What Ken does |
An OpenAI key | everything above; most calls cost a fraction of a cent, a |
A Claude Code login | the same jobs through |
No model | stores what agents send with no duplicate or conflict check, and cannot read |
Turn Decisions off with decisions: {enabled: false}; other models are under Models below.
Add Ken to claude_desktop_config.json (Settings → Developer → Edit Config) and restart the app.
Use the full path that which ken-mcp prints:
{
"mcpServers": {
"ken": { "command": "/full/path/to/ken-mcp" }
}
}Set a token in ~/.ken/config.yaml, then serve Ken over HTTP:
http:
host: 127.0.0.1
port: 8765
token: "a-long-random-secret"ken-mcp --httpPut it behind a tunnel or reverse proxy you control (ngrok, Cloudflare Tunnel, Caddy). Then
connect with https://<your-host>/mcp and the header Authorization: Bearer <token>:
claude mcp add --transport http ken https://<your-host>/mcp --header "Authorization: Bearer <token>"An app that cannot send a header can use https://<your-host>/t/<token>/mcp. The token is the
only lock, so treat the URL like a password. Ken refuses to listen beyond localhost without a token.
All settings sit under llm: in ~/.ken/config.yaml.
Setup | Under |
OpenAI (default) |
|
Any OpenAI-compatible API |
|
Local model (Ollama) |
|
Your Claude Code login |
|
api_style: chat is for a server with no /v1/responses endpoint. The Claude Code provider runs
claude -p for each call, so answers take a few seconds longer. Use it on your own machine or your
own household server, not for a public service. More options are in
engine/config.example.yaml.
Your AI app shows each card in the chat. You can also answer from the terminal:
ken cards # the open cards
ken decide <id> yes # or: noken push sends your open cards to any URL that takes a text body, for example an
ntfy topic. Set push.url in the config and run ken push each morning from
cron or launchd. Nothing waiting means nothing sent. Try it with ken push --dry-run.
Tool | What it does |
| open (or closed) tasks; everything that matters for one task, with the evidence for each item |
| one person or project, a search, your whole world, what happened in a date range (with |
| task writes |
| who does a task; which task waits on which |
| events, people, projects, preferences, facts |
| a free-text report; Ken turns it into the right writes |
| day zero: build your world in batches you confirm |
| the open cards, and your answer to one |
| what the Dreamer found overnight, and your answer |
Every write returns the real outcome and the ids of any cards it raised, so an agent has the facts to tell you correctly what was saved.
Ken Cloud
Rather not run it yourself? We run Ken for you: the same engine as this repo, on a small virtual machine and disk of your own in Frankfurt (EU). Sign in with Google, paste one URL into Claude (or ChatGPT, where custom connectors are on), and see your day and your world in a web app. Model calls are included, so you need no API key. Get early access →
Self-hosted | Ken Cloud | |
Where Ken runs | your machine | your own virtual machine and disk, Frankfurt (EU) |
Setup | install, | sign in with Google |
Model | your OpenAI key, Claude Code login or a local model | included |
Connect claude.ai or ChatGPT | a tunnel or reverse proxy you run | one URL |
Web app (your day, your world) | not included; build your own | included |
Updates |
| we deploy them |
What changes when we host it:
Where your world lives: a 1 GB disk attached only to your machine. Accounts share the sign-in and the URL gateway, not the engine or the database.
Who can reach it: your own token, and us as the operators of the machine. We do not read it unless you ask us to debug something.
Model calls: the text of a write goes to OpenAI, under our account, so it can be judged.
Calendar: if you link Google Calendar, the gateway keeps its token encrypted and reads today's events live for the web app. It stores no events.
Moving to self-hosting: ask us, and we hand you your SQLite file to run with
ken.
Benchmark (preview)
A preview, and our own benchmark. We wrote ken-bench for the job Ken is built for: keeping one person's current state right while it changes. Read it as a check on that job, not as a neutral comparison.
Each scenario is 16 weeks of one person's life: 4 projects, about 25 tasks and about 137 changes, written as chats, emails and Slack messages. Dates move, tasks get done in other words, and other people report things that are not true. The same agent (DeepSeek V4.1 Flash) reports every day into each system through that system's own tools. At 4 checkpoints we ask what is true today. 4 scenarios × 3 runs, on seeds we did not use while we built the benchmark.
Mean per run | Ken | to-do file | mem0 |
Planted false values found in the store | 0 | 6.8 | 6.6 |
Answers that repeat a planted false value | 0 | 1.8 | 1.3 |
Answers that give an out-of-date value | 1.0 | 0 | 0 |
Current-state score, all questions (0 to 1) | 0.92 | 0.89 | 0.87 |
· Which tasks are open? | 0.90 | 0.75 | 0.84 |
· Which tasks are due by a date? | 0.93 | 0.76 | 0.76 |
· Is this task done? (asked in other words) | 0.97 | 0.99 | 0.89 |
Yes/no questions put to the user in 16 weeks | 53 | 0 | 0 |
False reports stay out. In 12 runs, no planted false value reached Ken's store. The to-do file and mem0 each stored about 7 and repeated 1 to 3 in answers.
The overall gap is small. Ken leads the to-do file by 0.03, and the to-do file is ahead on 2 of the 4 scenarios. Most questions are easy for all three.
Ken pays with questions: about 3 a week, and the user says no to 4 in 10.
Ken was wrong too. It gave an old value about once per run, and in 6 of 12 runs it took an errand that someone posted to a whole channel as the user's task.
mem0 is the open-source
mem0ai2.2.1. It is built for recall, not for this job.No cost numbers in this preview.
The full table, the method, the limits, the scenarios and every raw answer are in ken-bench.
Privacy
This is self-hosted Ken. For Ken Cloud, see what changes when we host it.
Ken stores everything in
~/.ken/: one config file and one SQLite database, both private to you.The text of a write goes to the model you configure so it can be judged: OpenAI by default, your Claude Code login, or a local model that keeps it all on your machine.
Ken sends no telemetry. Apart from those model calls, only
ken pushsends anything, and only if you turn it on.Ken never reads your mail, chats or calendar itself. Your agents read those with their own tools and send Ken the result: who is who, what was decided, what is due.
Agents you connect run with your permissions. Connect only agents you trust; give each one its own token so its writes come to you as cards.
To remove Ken and everything it stored: uv tool uninstall ken-engine && rm -rf ~/.ken.
FAQ
Their memory lives inside one app and one vendor. Ken is one world that every MCP app shares, it is stored on your machine, and you approve what agents write. You can see and change everything it holds.
mem0 (OpenMemory) and Zep (Graphiti) also serve memory over MCP. mem0 updates and deletes facts as they change; Graphiti keeps typed, time-aware facts in a graph. Ken differs in three ways: it tracks task state (owner, status, deadline, what a task waits on), it holds uncertain and agent-token writes for your approval, and it is one SQLite file with no vector store.
Ken looks things up by the names and links you confirmed: the task names Tomas, Tomas works on the kitchen renovation, the floor plan belongs to it. That is exact and explainable, and it fits in one SQLite file. The trade-off: a lookup by name misses a synonym ("the plumber" does not find Tomas unless a note says so). The agent that calls Ken usually knows the name.
Alpha. We use it every day. Before 1.0, expect breaking changes to config keys and MCP tool arguments; the database migrates forward on upgrade. How it works inside is in ARCHITECTURE.md.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
This server cannot be deployed
Maintenance
Related MCP Connectors
One memory for all your AI tools: Claude, ChatGPT, Cursor and any MCP client share the same notes.
Shared AI memory. ChatGPT, Claude, Cursor and any MCP app read and write the same memory.
Shared brain, one current version of memory for any MCP client: ChatGPT, Claude, Gemini, Cursor
One memory, every AI. A shared, user-owned markdown memory your AI clients read and write over MCP.
Related MCP Servers
- AlicenseBqualityDmaintenanceA shared memory layer for AI agents — one memory.md synced across Claude Desktop, Cursor, Claude Code, OpenAI Codex, and any MCP client.42MIT
- AlicenseAqualityCmaintenanceOne context engine for every AI. Persistent context from GitHub, Notion, and your decisions, delivered to Claude, ChatGPT, Cursor, VS Code, Gemini, and any MCP-compatible client.93AGPL 3.0
- AlicenseAqualityCmaintenanceShared, versioned memory for your agents across Claude Code, Cursor, Kiro and Grok. One MCP server, sourced answers, full history.3MIT
- AlicenseNot gradedqualityCmaintenanceEnables AI apps such as Claude, ChatGPT, Cursor, VS Code and Codex to share one searchable memory, so a fact told to one app can be recalled in another with the exact line, speaker and date it came from. Users sign in per app and can allow read-only or read-and-write access, saving, correcting or forgetting memories and disconnecting apps at any time.MIT