The situation
Most business AI arrives as a subscription to somebody else's cloud. You send your operational data — payments, pipeline, resident records, client conversations — to a service you don't control, running a model you can't pin, on infrastructure you can't inspect. For a mid-market business that's often not a compliance decision so much as a shrug: there wasn't another option.
There's a second problem underneath it. Business teams don't work in apps. They work in WhatsApp and Telegram groups. Any assistant that requires them to be somewhere else has already lost, no matter how good the model is.
BizClaw started as the answer to both, for our own operation, and became the substrate under everything else on this page.
What it is
A business AI assistant that runs on the machine you own, reachable from the chat apps your team already lives in. Built on NanoClaw (MIT) and extended with what a business needs on day one rather than as an afternoon of configuration.
Apple Container as the default runtime. Each agent runs in a lightweight Linux VM on Apple's Virtualization framework — roughly a second to boot, no Docker daemon, hardware-enforced isolation. Docker still works on Linux via a conversion skill.
Multi-model by task, not by loyalty. Claude orchestrates; another model handles analysis and synthesis; a third summarises. Cost gets optimised per task instead of per vendor.
Preference fine-tuning. Thumbs-up and thumbs-down on responses become a training set. A LoRA-fine-tuned checkpoint is trained on that preference data, served behind an inference proxy, and production traffic routes to it — with a hosted fallback when it's unavailable. Monthly retraining cycle. The assistant gets measurably more like the way this business wants to be answered, rather than generically better.
A real skills engine. Not a prompt folder — apply, merge, rebase, replay, migrate, backup, uninstall, with a manifest, a lock file and path remapping. Capabilities install and update like packages, which is the difference between a system that accretes and one that rots.
Sixty-plus skills, from channel adapters to a one-command sales CRM, a daily collections report, and the apartment-specific skills that make Smriti work.
What runs on it
- Andy — the personal assistant on WhatsApp.
- Andy as chief of staff — on Telegram, orchestrating three specialists: content, revenue and personal intelligence. Delivers a morning brief, maintains a delegation log, and runs a weekly review of its own delegation decisions.
- Movisvami — the cinema content pipeline.
- Smriti — the apartment community agent, live in a real building.
- Sales group agent — a daily collections report generated each evening and emailed to stakeholders.
- Global memory — a shared brain across agents: status, observations, proposals, world context, growth log, delegation log.
Six agents, one substrate, isolated from each other by construction.
Run the script, don't reason
Where this system can be wrong
BizClaw has more of these than anything else we've built, because the platform is where discipline has to live if it's going to hold.
The agent must run the script, not reason. Every skill ships against a ten-step checklist: a written contract, deterministic code, unit tests, and evaluation. The rule is that anything code can do deterministically must not be done by the model. Calendar recall greps local files rather than guessing dates. Collections summary parses message structure with a regex rather than reading and estimating. The line in our own documentation says it best: the agent had the right tool and chose cleverness instead of discipline.
A reliability scorecard. There's an installable skill whose only job is to measure whether the agent is actually calling those deterministic scripts or quietly hallucinating the answer. Compliance is tracked, not assumed. Most AI deployments have no idea how often their assistant improvises — this one keeps score.
A documented failure, fixed in code and left in the docs. The morning brief once went out headed Saturday, June 7, 2026. That date was a Sunday. The fix wasn't a sterner prompt; it was a rule that the date header must be produced by shelling out to the system clock, never computed by the model — and the reason is written into the instructions, wrong date and all, so nobody removes the guardrail later without understanding what it's for.
Security by mounting, not by permission checks. The trust model is written down: the main group is trusted, every other group is untrusted, all messages are treated as potential prompt injection. The project root mounts read-only so an agent cannot rewrite the host application and escape the sandbox on the next restart. The mount allowlist lives outside the project root and is never mounted, so agents can't edit their own permissions. Symlinks resolve before validation. Credentials and private keys are blocked by default. Sessions are isolated per group.
That last one is the general form of Smriti's read-only database handle. The pattern is the same at every level: make the dangerous thing structurally impossible rather than instructing the model to avoid it.
What happened
BizClaw runs the firm. The morning brief, the collections report, the content pipeline, the sales group, and a client deployment in a real apartment building all run on the same install, with a changelog that tracks upstream releases and merges them.
A brain-migration skill exports an agent's memory and specialist team for import into another install — which is what makes a client deployment a delivery rather than a dependency on our hardware.
There's also a forward-deployed engineer onboarding kit: a master checklist, a scoped repo manifest, a core sync process, a build guide, a code review, and a plain-language confidentiality note. Built because handing an agent platform to someone else is a process, not a zip file.
What you can turn on
Capabilities install as skills rather than arriving as a rebuild. A representative set:
| Capability | What it does |
|---|---|
| Collections report | Daily receivables summary from a chat group, emailed to stakeholders. Case study |
| Sales CRM | Lightweight pipeline in any group — leads, stages, deals — with a weekly report by email |
| Content calendar | Plan, approve, produce and publish across platforms. Case study |
| Channel adapters | Thirty-plus: Slack, Teams, Signal, Discord, Matrix, iMessage, WeChat, Webex, Telegram, WhatsApp |
| Voice transcription | Send a voice note; the agent transcribes and acts on it |
| Email & calendar | Send, receive and schedule from the same assistant |
| Reliability scorecard | Measures whether the agent calls deterministic scripts or improvises |
| Brain migration | Export memory and specialist team; import into another install |
That last one is the one to read twice. It is the difference between a system you own and a system you rent.
What made it repeatable
Everything about it. The skills engine makes capabilities installable and updatable. Brain migration makes an agent portable between installs. Setup runs as a guided skill rather than a README. The onboarding kit makes bringing a person up to speed repeatable, not just the software.
The engagement shape
For a business that wants its own assistant rather than a seat on someone else's:
- Install on your infrastructure, connected to the groups your team already uses.
- One high-value automation first — the daily report someone currently assembles by hand.
- Harden it — mounts, trust boundaries, deterministic scripts for anything factual.
- Hand over the brain, so the agent is yours and moves with you.
At the end you own the machine, the data and the memory.
What's still open
Named, not hidden. A case study that admits open items is worth more than one that doesn't.
- macOS-first. Apple Container is the default runtime, so the best experience is on a Mac. Docker conversion exists for Linux but is the secondary path.
- Upstream merge debt. Tracking an active upstream project means periodic reconciliation, and there's a deferred-merge document acknowledging what hasn't been pulled in yet.
- Preference fine-tuning is a quality bet, not a proven uplift. The pipeline works end to end and retrains monthly. Claiming it makes answers better would need evaluation numbers we haven't published.
- Checklist coverage is partial. Several skills still lack the full ten-step treatment. The gap is tracked rather than hidden — which is the whole point of having a checklist.
Next case study
Smriti →