Projects

Work mode

Work gives a model project context and a small set of scoped tools. The model can read, search and edit only what the project allows, and every permission is explicit.

What Work is

  • A project folder with its own system prompt, memory, skills and runs.
  • Context locations with read, write and delete permissions.
  • A web permission that is off until you enable it.
  • A send gate: if the tool probe fails, the project cannot quietly run tools.
  • Works with a local model or a remote router (OpenRouter, EuroRouter, OpenCode Go). A Compute host is inference-only, so it cannot run Work tools yet.

Create a project

  1. Work → New project. Give it a name and a folder. The folder is the project's boundary. New chat sits next to it and starts a blank session in the open project.
  2. Write the system prompt. Say what the project is for, what the model may change and what it must not touch.
  3. Pick a provider. A local model is best for private files and reliable tool calls; a remote router works when the local model is too small. A Compute host is inference-only for Chat and cannot run Work tools yet.
  4. Add context. Point at the folders and files the project actually needs. Start read-only.
  5. Enable write/delete only after a read run. The permission is per location, not global.

Sessions and names

Sessions belong to the project that stores them. Chat lists the default project's conversations. Work lists every project that has work sessions, each as a collapsible group with a count. Opening a session from another group switches to that project first. Rename, delete and clear stay on the current project because they edit the live session store. A new session keeps the provider and model of the last chat, so switching projects does not quietly drop you back to a local model. If the saved remote key was deleted, Studio falls back to local.

The assistant's name follows the CUI convention. The sampling temperature selects a Greek tier: Alpha (cool) through Gamma to Epsilon (hot), and a session with no explicit temperature is Delta. The name is shown above each reply and given to the model in its system prompt. Temporary sub-agents keep the {Deity}-{Tier} form, such as Athena-Alpha.

Context and permissions

PermissionWhat it allows
ReadList, search and open files under the context location.
WriteCreate and edit files under that location.
DeleteRemove files under that location. Keep this off unless the task needs it.
WebFetch pages for the project. Off by default.

Paths outside the declared context are not reachable through the project tools. The project folder itself is not a substitute for an explicit context location.

Tool permissions

Every tool has four states. Off removes it from the prompt, so the model never sees it. Never keeps it visible but refuses every call without asking. Ask prompts, and Always permits. Each group header carries the same control and applies one value to every tool in that group, which is the quickest way to stop all writes at once.

A tool inherits from the model, family, project and Studio defaults, and each layer can override the one above it. A folder set to Never is a hard boundary and refuses the call whatever the tool says; otherwise the tool's own state decides, so Always is not silently downgraded by a folder that only asks. A delegated agent may be configured more permissively than the session, but that difference always raises a prompt and is never remembered.

Project settings save as you change them, and sending a prompt hides the panel so the conversation is what you see. A new project starts with a default context at its root, so the file tools work before the settings have ever been opened.

Skills

Skills are reusable instructions with a name, description and body. They can be global or overridden per project. Use them for repeatable procedures: review a dataset, prepare a model card, run a release checklist.

  • The skill list shows the effective source (global or project).
  • A skill can be previewed before it is used.
  • Skills do not grant permissions; the project's context locations still decide what tools can touch.

The work loop

  1. Ask for the outcome, not the exact keystrokes. Mention constraints and files.
  2. The model proposes tool calls; Studio shows them and applies the permission gate.
  3. Review file changes in the session before accepting destructive work.
  4. Keep the useful decisions in the project memory so the next session starts informed.
  5. Use runs and the project's models folder for artifacts you want to keep.
Files, then programs. Work can list, read, search, create, write, append, edit and delete files, and fetch a page when the project allows it. Edit replaces one exact piece of text, so it can insert, change or remove a section without rewriting the whole file; it refuses ambiguous text instead of guessing. RunProcess can start a program in an allowed folder once that folder's Run permission is Ask or Always; it is Never by default, passes arguments without a shell, and kills the tree on timeout. It runs real code with your account's permissions, so Studio gates the command, not everything the program touches.

Model providers

Local modelOpenRouterCompute host
PrivacyFiles stay on this PCPrompt and tool results go to the providerData stays on the host you control
Tool reliabilityDepends on the model; smaller models are weakerStronger models available, still model-dependentInference-only; Work tools not supported
CostFreePer tokenYour host
SetupDownload a model and pick itAPI key and a modelSharpMind Server image, URL and token

A SharpMind Compute host is currently inference-only for Chat. It does not expose Work's tool calling, so Work should use a local model or a remote router. Chat can use any of the three. OpenRouter, EuroRouter and OpenCode Go keep separate keys, base URLs and model lists, so Work can switch between them without re-entering a key. A new session starts on the router used last.

Tool-call recovery. Studio can be strict, lenient or loose about malformed tool calls. A small local model may need Lenient; a hosted model is usually fine with Strict.

Safety

  • Start with read-only context. Add write after the model has shown it understands the project.
  • Leave delete off for source trees, model folders and anything under version control.
  • Keep secrets out of context locations; the model can read every file it is given.
  • The send gate exists to stop a broken tool probe from becoming a silent failure. Fix the probe rather than bypassing it.
  • Review remote-model prompts: OpenRouter sees the conversation and any tool output you send.

Problems

SymptomFix
Send is disabledThe tool probe failed. Fix the provider or model, then retry; the gate is intentional.
Tool says permission deniedThe path is outside the declared context, or the location lacks that permission.
Model ignores toolsUse a tool-capable model, lower the recovery strictness, or restate the task.
File change missingCheck the context path and whether the model wrote to a different project folder.
Remote model is slow or refusesSwitch to a local model for private or tool-heavy work; OpenRouter models vary in tool support.