Launching Munder Difflin v0.3.0: Selectable Engines, Integrations & Slack-Spawned Workers

for a while now the floor has had a not-so-secret default. Michael — the GOD orchestrator who routes work like a slightly unhinged regional manager — and most of his coworkers were, deep down, Claude-shaped. you could bring Antigravity and Codex to the party (and they were first-class), but the orchestrator at the center of it all was wired to one CLI. the team was multi-provider; the brain wasn’t.

v0.3.0 fixes that. this is the biggest platform release we’ve shipped — the version where the floor stops assuming what runs it.

selectable engines: pick the brain behind every coworker #

the headline is selectable agent engines. the runtime behind each agent is now pluggable:

• 🟣 Claude Code
• 🔵 Antigravity (Gemini, via agy)
• 🟢 OpenAI Codex
• 🖥️ a local provider — a claw/qwen backend proxy, for when you want a model running on your own machine

and crucially: Michael is swappable too. the god orchestrator is no longer hard-wired. there’s an engine picker right in onboarding, and a change-engine flow so you can re-home the orchestrator onto a different engine without tearing down the whole office. don’t love the brain running your floor? change it. it’s a setting now, not a rebuild.

every hire brings its own skills + MCP — with a consent step #

an engine is what runs a hire. but a hire also needs to know what it’s allowed to touch. so each hire now carries its own manifest of allowed skills and MCP servers — a default-deny allowlist over a shared catalog. some skills are bundled right into the app; others come from the catalog.

the part we care about most: a consent UI. before a hire can use a skill or an MCP server, you see exactly what it’s asking for and approve it. a hire you imported from someone else doesn’t get to quietly wire itself into a tool — its requests are reviewed, never auto-granted. (if you want the why-this-matters version, we wrote about MCP and skills in a hive and the tool-poisoning surface separately.)

integrations: connect a service without leaking the key #

the second big thing is an integrations registry plus a loopback secret broker.

before, “connect a service” meant hand-wiring something and hoping the credential stayed put. now there’s a proper registry: a declarative spec drives a Settings UI that renders each integration’s config form, and v0.3.0 ships a first wave of templates so common services are a few clicks instead of a research project.

the bit that matters is how secrets are handled:

secrets are write-only. you set a credential once. it’s reached only through a loopback broker. it is never read back into the renderer.

so the UI can let you configure an integration without the secret ever round-tripping through the part of the app that draws windows. and the connectivity-test path — the “is this hooked up right?” button — is confined, so it can’t be bent into a secret-exfiltration or SSRF trick. write-only in, brokered access out.

Slack → spawn → reply → safe teardown #

the third headline is the one that feels like magic the first time: Michael can spawn a worker straight from Slack.

here’s the loop. a request lands in Slack. Michael spins up an isolated, ephemeral worker to handle it. the worker does the job and posts its reply back into the thread itself. then — and this is the whole trick — it gets torn down safely.

“safely” is doing real work in that sentence:

• 🧷 the teardown gate. a worker is never auto-discarded if it has unintegrated work. nothing of value evaporates because a worker finished its turn.
• 🧮 token caps. every spawned worker runs under a cap, so an ephemeral worker can’t quietly torch your bill.
• 🧹 worktree GC. each worker runs in its own isolated git worktree (the same worktree isolation the floor already uses); abandoned ones are garbage-collected instead of piling up.
• 🪟 the Workers tab. a new tab in the UI shows the ephemeral workers that are live right now, so the loop is watchable instead of invisible.

if you want to wire this up, the Slack setup guide and triggering agents from Slack still apply — v0.3.0 just makes the worker on the other end ephemeral and self-cleaning.

the smaller things that add up #

a release this big has a long tail of quality-of-life wins:

• ⏱️ temporal date-range skills — today, yesterday, thisWeek, lastWeek, thisMonth, thisQuarter, thisYear, lastMonth, last7Days, last30Days, and an arbitrary-range resolver. ask for “last 30 days” and an agent gets concrete ISO dates instead of doing date math by hand.
• 📇 a worker capability catalog — each spawned worker can read exactly which skills and brokered integrations it has, and how to call them. no guessing about its own toolbox.
• 🎛️ a Provider / Hive picker — choosing the engine for a hire is now a visual step (with real provider logos) in onboarding and add-agent, not a free-text command.
• 🎪 the Agent Gallery — The Hiring Fair is rebranded the Agent Gallery, now with six off-the-shelf hires ready to browse, review, and spawn. (import still only pre-fills the Add-Agent modal — the human always clicks spawn.)
• 🧭 feature-aware onboarding — first-run setup adapts to what you actually have available, with an explicit permissions & reliability step.
• 🔧 a visible engine-CLI installer — if the engine binary for your chosen provider is missing, the installer runs visibly and self-heals instead of failing in silence.

reliability: the floor survives a closed lid #

if you run agents overnight, two fixes are for you:

• 💤 auto-revive wedged terminals on wake. a terminal that wedged while your machine slept is now detected and revived when the machine wakes — instead of sitting dead until you notice and restart it.
• ⏰ catch up missed schedules on wake. a scheduled mission whose fire time elapsed while the laptop was asleep is now caught up on wake, not silently skipped. (we wrote a verifier for this so it stays fixed.)

close the lid with confidence. the floor picks itself back up.

everything from 0.2.8 is still in the box #

v0.3.0 is purely additive. nothing got removed. you still get shareable hires (now spawning from the Agent Gallery), and everything from v0.2.7 and earlier:

• 🎙️ voice dictation — hold Option to talk to your floor
• 🕸️ the enterprise Knowledge Graph — answers from your own documents and policies
• 🪟 multi-window floors — spread the office across monitors
• ✍️ the rich composer, ⏯️ session resume, plus observability, the circuit breaker, durable persistence, the Command Center, task kanban, GitHub/CI integration, and the Schedules tab

the floor you already love — now engine-agnostic, with a front door for your tools and a self-cleaning worker on the other end of Slack.

get v0.3.0 #

Munder Difflin is free, open source, and local-first on macOS, Windows, and Linux. no account, no cloud — your machine, your subscriptions, your floor.

Download v0.3.0, then pick an engine for your first hire, browse the Agent Gallery, and — if you’re feeling brave — point a Slack channel at Michael and watch him spawn a worker, answer, and tidy up after himself.

curious how the orchestrator decides any of this? read how the god orchestrator works. want the local-first philosophy? that’s why local-first matters. missed the last launch? v0.2.8 shareable hires is right here.

full release notes live in the CHANGELOG.

that’s it. go pick a brain, plug in a tool, and let Michael run the floor. (he’d like you to know the engine is now his choice too.)

FAQ

What's new in Munder Difflin v0.3.0?

The floor stops being Claude-shaped. Three big things: selectable agent engines (every hire — and Michael, the god orchestrator — runs on a pluggable engine: Claude Code, Antigravity, Codex, or a local provider), each carrying its own consented skills + MCP catalog; an integrations registry with a loopback secret broker that keeps secrets write-only; and a god-triggered ephemeral Slack worker loop where Michael spawns an isolated worker straight from a Slack message, replies in-thread, and tears it down safely. Plus temporal date-range skills, a worker capability catalog, a visual Provider/Hive picker, the Agent Gallery (the rebranded Hiring Fair) with six off-the-shelf hires, feature-aware onboarding, and wake-reliability hardening.

What does 'selectable agent engines' mean?

Until now the runtime behind an agent was effectively fixed. In v0.3.0 the engine is pluggable per hire: you choose Claude Code, Antigravity, OpenAI Codex, or a local provider (a claw/qwen backend proxy) from a visual picker. Even Michael — the god orchestrator you talk to — is swappable, with an engine picker in onboarding and a change-engine flow. Each hire also carries a manifest of allowed skills + MCP servers, surfaced through a consent UI before anything can use them.

How does the integrations registry keep my secrets safe?

Secrets are write-only. You set a credential once through a registry-driven Settings form, and it's reached only through a loopback secret broker — it's never read back into the renderer. The connectivity-test path is confined so it can't be turned into a secret-exfiltration or SSRF primitive. v0.3.0 ships a first wave of declarative integration templates.

Michael can spawn a worker from Slack — is that safe?

Yes, and safety is the point. Michael spawns an isolated, ephemeral worker in response to a Slack request; the worker does the job, posts its reply back into the thread, and is then torn down. Teardown is gated so it never auto-discards a worker's unintegrated work, every worker runs under a token cap, and abandoned worktrees are garbage-collected. Live workers show up in a new Workers tab.

Does v0.3.0 still include shareable hires and the gallery?

Yes. Everything from v0.2.8 — shareable hires and the community gallery — is included. The Hiring Fair is rebranded the Agent Gallery and now ships six off-the-shelf hires, with a visual Provider/Hive picker in onboarding and add-agent. Import still only pre-fills the Add-Agent modal; the human always clicks spawn.

Do I still get everything from v0.2.7 and earlier?

Yes. v0.3.0 is additive. Free Flow voice dictation, the enterprise Knowledge Graph, multi-window floors, the rich composer, agent session resume, observability, the circuit breaker, durable persistence, the Command Center, task kanban, GitHub/CI integration, and the Schedules tab all remain functional and shipping.

---