Your agent's model of you should be something you edited, not something that accreted while you weren't looking.
Every agent harness right now is racing in the same direction: automate the human out. Longer autonomous runs, more parallel agents, memory the agent curates for itself, instructions it rewrites on its own. It sounds great, until you actually try it. An agent that doesn't really know you produces the average answer: the code everyone writes, the prose nobody wanted. Take yourself out of the loop and you're not scaling your judgment. You're amplifying slop.
Here's the uncomfortable part: staying in the loop is work. Reviewing what ten sessions did overnight, correcting the same taste mistake twice, deciding what your agent should remember about you — at scale it's exhausting, and that exhaustion is exactly why everyone else concluded the human has to go.
Multiplai makes the opposite bet. The human is the value to amplify, so the tooling should make being the human easier, and honestly, more pleasant. Capture everything automatically; spend your attention only where your judgment matters; make the reviewing itself fast enough that you'll actually do it.
It starts with context, because an agent that doesn't know you can't do anything else well.
- Memory that compounds, with you as editor. Sessions keep a diary and capture learnings on their own — that's the taxing part, automated. But nothing is promoted to long-term memory without your sign-off: a review step ("dreams") proposes consolidations and you decide which ones make it in, because that curation is what shapes everything the agent does next. Relevant context is then routed into every new session automatically. Week 40 knows what week 1 learned, and you know why.
- No black boxes. Every routing decision, memory write, and hook run leaves a standardized, human-readable log line (with a structured trace mirror). When the agent did something surprising, you can see what happened, verify it, and steer — instead of guessing at a vibe.
- Autonomy that's actually safe to grant. Sessions run
--dangerously-skip-permissionsinside a Docker container that is the permission boundary. A key-restricted SSH bridge exists for the few tools that genuinely need your Mac (Xcode, Whisper, your real Chrome), and every bridge skill you enable widens that boundary: you enable them deliberately, knowing what each opens up, never by default. The browser is the sharpest example and gets its own switch — reaching your real logged-in Chrome means reaching every cookie on the machine, so enabling the bridge does not enable it; a file only you can create on the Mac does. You stop babysitting prompts because the walls are real and you know exactly where the doors are. - An environment per project, not one blessed image. Because sessions run in Docker images, the sandbox doubles as something more useful: a fully custom environment for Claude Code. A small overlay Dockerfile in a project's own repo — a Django or FastAPI stack with its database server and the locale its tests need, a pinned Node.js runtime, a data-science image with Jupyter and pandas — builds on top of the generic base, is rebuilt automatically by setup when the base or the overlay changes, and is selected per launch with a profile. Your agent gets exactly the environment each task needs, and switching is one flag. See overlay images.
- 45 skills — the memory engine plus seven skill packs. Autonomous spec-driven builds, deep research, transcription, Slack and Gmail, PM artefacts, long-form writing, Apple-platform builds. Each one encodes how the work should be done, so your standards travel with the task. The authoritative list is the compatibility matrix in the multiplai-cc-mktplace README.
On the roadmap, not yet released: multiplai-gui, a native macOS/iOS cockpit for observing, driving, and triaging many sessions at once (a session is a file, so watching is free). Everyone else ports the same chat box to a smaller screen; what you want to do on your phone is different in kind — glance at what your agents did, swipe through what they learned, approve or redirect in seconds, put it back in your pocket. Nobody else is designing for that, because nobody else is optimizing for the human doing the reviewing.
Same species, opposite bet. The systems getting attention today put the agent in charge of its own evolution:
- Companion agents in the OpenClaw / Hermes mold let the agent curate its own memory and grow its own skills — OpenClaw goes further and rewrites its own operating instructions. Run one for a month and its picture of you is whatever it decided to keep.
- Other memory pipelines compile knowledge through an autonomous LLM librarian. Well-engineered, provenance-tracked, and still: the trust sits with the compiler, not with you.
Multiplai automates the capture and keeps the judgment. Everything gets recorded; nothing becomes memory without your approval; every decision the system makes is logged where you can read it; and the review is engineered to cost you minutes, not evenings.
Add the plugin marketplace from inside Claude Code:
/plugin marketplace add spikelab/multiplai-cc-mktplace
Then /plugin install multiplai-context@multiplai for the memory engine. Want
the sandboxed full environment instead? Clone
multiplai-kit and run ./setup.sh.
Either way, read
GETTING-STARTED.md
next. It walks a first install end to end: what to run after setup.sh, what
your first session should look like, how the work → dream → review loop fits
together, which memory router to pick and what each one costs, and what the
first few failures mean.
| Repo | Role |
|---|---|
| multiplai-container | The sandbox — Docker image + macOS SSH bridge; overlay images build per-project environments on top. Usable standalone. |
| multiplai-cc-mktplace | The features — plugin marketplace: memory engine + seven skill packs. Works on vanilla Claude Code. |
| multiplai-kit | Distribution — setup.sh + claude.sh; the full environment, without touching your ~/.claude. |
| multiplai-core | Shared library — the typed Python plumbing every plugin script imports. |
| multiplai-gui (coming soon) | The cockpit — a FastAPI hub + SwiftUI app for session orchestration. Not yet released; the repo is private until it is. |
Start small — move down the table as you want more of the environment.
| You want | Get | Requires |
|---|---|---|
| Memory + skills on your existing Claude Code | multiplai-cc-mktplace: /plugin marketplace add spikelab/multiplai-cc-mktplace |
uv |
| Safe YOLO mode for the Claude Code you already have | multiplai-container standalone: clone, ./build.sh, docker run |
Docker |
| The full environment — sandbox, plugins, workspace, memory | multiplai-kit: clone, ./setup.sh, ./claude.sh |
Docker/OrbStack; macOS for bridge skills |
| Many sessions, one cockpit | multiplai-gui — hub + app on top of the kit. Not yet released. | macOS |
See ARCHITECTURE.md — the component map, the interlock diagram, and the delivery contracts (everything ships as immutable tags; merging alone delivers nothing).
The docs are the READMEs. Each repo's README is the deepest source for its own component; ARCHITECTURE.md is the map across them; the changelogs below are the release record. A dedicated docs site may follow.
- GETTING-STARTED.md — read this first: a first install end to end, the daily loop, choosing a memory router, and what the early failures mean
- multiplai-container README — image contents, the SSH bridge, tag/release contract
- multiplai-cc-mktplace README — every plugin and skill, the compatibility matrix, what runs unattended
- multiplai-kit README — install, configuration, credentials blast radius
- multiplai-core README — the library API and how the marketplace resolves it
- CHANGELOG.md — changes to this repo's documentation
What shipped recently, per component:
- multiplai-container CHANGELOG — the tag consumed by the kit's
CONTAINER_REFpin - multiplai-core CHANGELOG — the library the marketplace's
uv.lockresolves - multiplai-cc-mktplace CHANGELOG — an index across the per-plugin changelogs (
plugins/<plugin>/CHANGELOG.md), tagged<plugin>@<version> - multiplai-kit CHANGELOG — untagged so far; the kit is consumed by
git pull && ./setup.sh
Young, personal, built in public. I run my whole working life on it — that's
both the pitch and the disclaimer. Expect fast movement on main everywhere;
what's released is what's tagged.
Questions, ideas, show-and-tell: Discussions on this repo — the suite's one community home. Bug reports belong on the issues of the component repo they concern (see The repos).