Sonderr v1.7 Is Coming — The Claude-Grade Workflow

No download link in this one — this is a coming-soon post. Sonderr v1.7 is in active development, and the theme is simple: the workflow. Claude-grade workflow discipline — plans you approve, verification with evidence, memory that spans days — wrapped around everything v1.6 already shipped. Here is exactly what’s coming and why.

For context: Sonderr is the privacy-first AI workspace you install in one command and run entirely from your machine at 127.0.0.1:4173. Free models with no API key, our own local Sonderr-v1 model, real file tools, seven Studios with a two-specialist crew, Swarm rooms and opt-in Web3 tooling — all on the hardened v1.5-lineage local runtime. v1.6 turned lab experiments into the product. v1.7 turns the product into a place you can work in for hours.

The workflow, reworked — plans you approve

The headline of v1.7 is a reworked agentic loop modeled on the workflow discipline that made Claude the gold standard for long engineering sessions. Every substantive request moves through explicit phases: plan (a visible plan with milestone-sized items that you approve before edits start), build (each pass keeping edits at H-tier size and care, with new evidence every turn), and verify (quality checkpoints as the final acceptance check, not an afterthought). The pieces exist today — todo lists 2.0 with evidence notes, marathon U-tier budgets up to 72 hours, checkpointed pauses. v1.7 wires them into one continuous loop that feels like a senior engineer at your desk instead of a slot machine.

Two rules carry the whole design. First, honest clocks: the loop restates real active-work duration, remaining budget and wall-clock span every turn, and treats those numbers as the only trustworthy clock — no hallucinated progress, no manufactured filler. Second, memory that spans days: checkpoints persist goal, milestone, verified evidence and the one exact next action, so a task you park on Tuesday picks up cleanly on Thursday — clean handoffs across sessions, not a demo that dazzles for five minutes and forgets by dinner.

65 self-loading skill playbooks

The runtime already ships 65 skill playbooks — systematic debugging, API design, threat modeling, data visualization, secure integrations, wallet safety, technical writing and more. They self-load: when a task matches a skill, the playbook slides into context on its own, no configuring. In v1.7 they become visible and steerable — the workflow surfaces which skills informed a pass, and the docs gain a full catalog explaining what each one does and when it fires. If you have ever wished your AI workspace came with the accumulated habits of a good engineer instead of just weights — that is this.

Clean agent handoffs

Model Context Protocol support grows from “tools you can configure” to workflows you can run: dedicated MCP workflow guides, guarded mutating calls that require a matching present-turn action, and clean handoffs between the lead and read-only workers in rooms and Studios — the lead synthesizes, workers never edit. The same strict boundary enforcement the hardened runtime is known for: earlier context may resolve file or tool selection, but it can never authorize an edit. Your external tools, same trust model as everything else in Sonderr: local-first, explicit, fail-closed.

Image generation, front and center

The image stack graduates. Generation and editing already run on OpenAI-compatible /images endpoints, gated exactly right — Vision mode, a direct affirmative request in the current message, and a Tools & Access level that permits it. v1.7 polishes the whole path: cleaner prompting, better edit fidelity with source_path carry-over of details you want kept, and finished images delivered as interactive download cards with inline previews. Analysis stays analysis, editing stays explicit — the boundary that keeps the tool honest is not going anywhere.

Docs you can read for hours — and a Gallery

v1.7 ships alongside the biggest documentation expansion in DXN1 history. The docs site grows deep, per-feature guides — image generation, skills, agents & Swarm, MCP, tasks & checkpoints, the CLI, troubleshooting trees — written to the standard of the best product docs on the internet, and then some. There is also something new: DXN1 Gallery, a dark, cinematic showcase of AI-generated key art from across the studio — a different look for one page, same DXN1 colors underneath.

What does not change

The security posture stays exactly where it matters: fail-closed release gate, size-capped release metadata fetch, timestamped update backups, no-follow exclusive file writes, escaped markdown rendering, and workers that stay strictly read-only in Swarm rooms. Free stays free. No API key stays no API key. Local-first stays local-first. v1.7 is a workflow release, not a business-model release — nothing is moving to a subscription, a cloud requirement or an account wall.

When, and how you get it

No date — shipped when it’s ready, same as always. Watch this site, the roadmap and the repo. When it lands, managed installs get it automatically through the required-update flow with timestamped backups, exactly like every release before it. Fresh install is still one command:

Current stable lives on GitHub Releases →

That’s the preview. v1.7 is the update where Sonderr stops being a tool you use and becomes a place you work — built in the open, one stubborn builder, shipped when it’s ready.