saki-thinker

How it works

A brain and soul your agents can rent

Persistent memory and a persona, rented by the agent — not rebuilt per runtime.

saki-thinker is a hosted brain over MCP: describe a soul, issue a scoped key, and every runtime you ship reads standing context, deposits what it learns, and recalls it fenced to your workspace.

How it works
  • Scoped API keys
  • Soul + memory in one brain
  • One MCP endpoint

01 · Why

Memory that outlives one runtime

Deposits persist across sessions and survive a runtime swap — the same memory comes back fenced to your workspace, not scattered in one process.

A persona, not a prompt

Describe a soul once — identity, policies, beliefs — and every agent you connect inherits it. No re-prompting the character into existence.

Fenced by default

Every read and write is stamped with its workspace and soul. Nothing your agent learns leaks sideways into another tenant's brain.

02 · Who it is for

  • Agent developersstop rebuilding memory infrastructure for every runtime you ship
  • MCP hostsconnect a described brain with scoped keys instead of hand-managed secrets
  • Operatorssee usage, policies and keys in one console instead of curling an admin API

03 · Setup

  1. 01

    Describe a soul

    Give the agent its identity, policies and beliefs in the console.

  2. 02

    Issue a scoped key

    A per-runtime credential with the exact read and write scopes it needs.

  3. 03

    Point the runtime at the endpoint

    One MCP config line in your host — no SDK, no pasted secrets by hand.

  4. 04

    Deposit and recall

    The agent remembers what it learns and recalls it, workspace-fenced.

04 · How it works

  1. 1. Agent calls

    get_context reads the soul's standing identity and policies.

  2. 2. Deposits a memory

    remember writes one entry stamped with soul and workspace.

  3. 3. Recalls fenced

    recall returns the right memories — nothing crossed the fence.

  4. 4. Runs on MCP

    A hosted brain your runtime plugs into over one endpoint.

05 · Compatible with

Claude Code

One MCP config line connects Claude Code to the brain — scoped key, no SDK.

Hermes Agent

Hermes reads standing context and deposits memories, fenced to your workspace.

OpenClaw

OpenClaw speaks to the hosted brain over the same single endpoint.

Any MCP client

Anything that speaks MCP connects — the endpoint is the interface, not a vendor SDK.

06 · Start

Give your agent a brain it keeps — not one you rebuild every time.