Local checkpoints, adapters, quantization, inference backends, and three provider-shaped APIs form a replaceable model layer beneath the agent client.
Agent Atlas / Adjacent references
unslothai/unsloth
A local model workshop and serving layer whose Desktop beta joins training, export, OpenAI and Anthropic compatibility routes, harness-specific launch adapters, local subagents, and an opt-in model control plane.
Local model infrastructure
Full dossier
Reported + sourced
Reviewed August 26, 2026
Intended operator
Builders training or serving local models beneath coding-agent and general-agent harnesses.
Primary surfaces
Desktop model workshop / OpenAI and Anthropic compatibility APIs / harness launch and local-subagent adapters / opt-in MCP model control
Tracked repository
Teaching mechanism
The harness is not the model
The launch command configures distinct protocol, context, compaction, state, and permission adapters for Claude Code, Codex, OpenCode, OpenClaw, Hermes, and Pi instead of treating one compatible endpoint as complete agent equivalence.
Evidence limitDesktop is a beta, compatibility does not establish feature or behavioral parity, and TheGitReporter did not run a live model, GPU, tunnel, or agent session for the launch article.
Inspect the source
System Map
Eight inspection boundariesEach harness keeps its own history and compaction behavior while Unsloth's launcher supplies the actual model context limits and isolated session configuration.
Coding-harness launches default to client-side tools; Studio and API server tools can execute web, Python, or terminal work as the host user when enabled.
The launcher maps permissive mode into each harness's own policy, isolates configuration where possible, and requires separate protection for API keys, tunnels, and server-side tools.
Harness-owned session state remains distinct from Unsloth's model lifecycle; compatibility translations and bounded response healing repair selected protocol failures without becoming a universal recovery layer.
Claude Code, Codex, OpenCode, and Pi can register the local model as a bounded subagent while retaining a different parent model and harness.
Desktop exposes model selection, training, export, serving, and model-control operations while harnesses retain approvals and execution policy.
Tagged source, adapter code, route translations, tool-call healing tests, security warnings, and reproducible harness-by-model runs provide the strongest receipts.
Learn Through This Project
8 lessons
Chapter 1
Before the First Model Request
Chapter 3
Tool Calls: Contract, Authorization, Execution
Chapter 4
Loops, Workflows, and Recovery
Chapter 5
Delegation Is a Distributed System
Chapter 7
The Approval Trap
Chapter 8
The Human Control Loop
Chapter 9
Trust Must Make Its Case
Chapter 10
The Future Has Three Branches
Latest Reporting
7 articles
Unsloth Can Reject a New Model to Keep Another Loaded
October 19, 2026 — New Studio code gives chat, image and video models a choice: share GPU memory when possible, make room, or refuse a disruptive load. When media does displace chat, it can also save the primary GGUF model’s cached work. unslothai/unsloth
Unsloth Adds Shared Fine-Tuning for Seven Text-Diffusion FamiliesOctober 11, 2026 — A new trainer connects familiar datasets to distinct learning recipes for models that refine text instead of only predicting the next token. The gain is less custom integration—and a clearer view of the differences that still matter. unslothai/unsloth
Unsloth Adds Chosen MCP Sources to Research ReportsOctober 10, 2026 — Studio’s research mode can draw on search tools you select, alongside the web and its knowledge base. The merged change expands what can reach a report, while leaving a revealing boundary around how the next question is chosen. unslothai/unsloth
Inside TypeSafe’s Model Comparison: What the Adapter Makes You Pay ForSeptember 29, 2026 — A probability-bearing answer buys more than a label. Local decision models and a public routing experiment now expose the next choice: how much accuracy must the cheaper route preserve? typesafe-ai/system-one-adapter-python, unslothai/unsloth, earendil-works/pi, ickma2311/jev-baselines-eval
Unsloth’s Sandbox Is Back. A Working Tool Still Does Not Prove Isolation.September 26, 2026 — The replacement for Studio’s reverted sandbox has landed, followed by a Windows Preview. Its promise is conditional: establish what boundary actually ran, and distinguish an unavailable host from a launch that must be refused. unslothai/unsloth
Unsloth Rolled Back the Sandbox Its Review Had Just ClearedSeptember 18, 2026 — A live review reported working Linux and macOS isolation. The next day, the implementation was reverted. Studio’s remaining protections make the consequence more specific than “the sandbox is gone.” unslothai/unsloth
Unsloth Desktop Turns the Local Model Into an Open-Harness BackendAugust 12, 2026 — The beta is more than a local-model GUI: it puts training, export, three compatibility APIs, harness-specific adapters, and a model-control plane on one machine—creating a serious test bench for separating the model from the agent that drives it. unslothai/unsloth
Primary Sources
Open Questions
- Can reproducible harness-by-model matrices separate model quality, protocol repair, harness policy, tool execution, and compaction costs?
Evidence state describes what this Atlas profile can support today. It is not a product score and it is not evidence that uninspected capabilities are absent.