Start here5 min readUpdated for Syntheo v0.4 Series

Syntheo User guide

© 2026 Quadropic Pvt Ltd. Proprietary and confidential. Do not copy, redistribute, scrape, adapt, or use this material to develop or train a competing product. Full use restrictions.

Syntheo is a development harness: it gives an AI model a workspace, durable task history, code-search tools, a terminal, a browser, Git workflows, permissions, and optional multi-agent coordination. The model matters, but most of the leverage comes from choosing the right execution mode and giving the harness a crisp target.

This guide documents the behavior currently available in the desktop app.

Start here#

If this is your first session, follow Getting started. For day-to-day work, keep Tasks, prompts, and context nearby.

What you wantBest starting pointWhy
A small, well-defined code changeStandard + SubmitLowest coordination overhead
An explanation with no intended editsAskOptimized for a concise answer
A plan you want to review before implementationArchitectInspects the code and writes a durable plan without changing implementation files
A hard bug with unclear causeDebugPrioritizes reproduction and diagnosis before a focused fix
A broad or ambiguous problemHunter + SubmitGives the run more reasoning capacity
Web research or interaction with a siteBrowser preset or Browser panelRoutes to web-aware behavior and provides a persistent page surface
UI/product design workDesign presetUses the model assignment configured for design work
Risky or experimental changesVirtual EditIsolates the thread in a temporary Git worktree
Several independent workstreamsA named room + ConductorGives workers a shared objective and coordination scope
Recurring work while Syntheo is openScheduled TasksRuns a prompt on a daily, weekly, weekday, or cron schedule
Start work from a shell or scriptsyntheo CLICreates a desktop task without leaving the terminal

The leverage loop#

The highest-signal Syntheo workflow is:

  1. Choose the workspace. A task can only inspect and edit the directory it is launched in (or its associated worktree).
  2. Choose the action. Submit, Architect, Ask, and Debug carry different instructions. Do not use a large prompt to compensate for the wrong action.
  3. State an observable outcome. Name the behavior that should exist when the task is done.
  4. Attach focused evidence. Mention files, paste an error, add an image, or identify a reproduction path.
  5. Define constraints and validation. Say what must not change and what command or behavior proves success.
  6. Let the agent inspect. Avoid prescribing an implementation before the agent has seen the code unless the implementation itself is a requirement.
  7. Steer with deltas. While a run is active, send only new information or a changed priority.
  8. Review the evidence. Read the final summary, inspect File Diff, and run the relevant tests.
  9. Integrate intentionally. Apply a Virtual Edit, stage/commit through Git, or keep the result isolated.

A strong request usually fits this shape:

text
Outcome:
When a signed-in user refreshes /billing, their selected plan is preserved.

Evidence:
The failure starts after commit abc123. Reproduction: ...

Constraints:
Do not change the public API or add a dependency. Preserve the guest flow.

Done when:
- the reproduction no longer fails
- the existing billing tests pass
- a regression test covers refresh after selection

Guide map#

Core terms#

Workspace: A project directory registered with Syntheo. Tasks and tools are scoped to a workspace or a worktree derived from it.

Thread: A durable conversation and execution history for one task. A thread keeps its model choice, messages, tool activity, approvals, and artifacts.

Action: The intent attached to a prompt: Submit, Architect, Ask, or Debug.

Preset: A user-selectable model lane such as Standard, Hunter, Design, or Browser. Each lane maps to a provider, model, and optional thinking level in Settings.

Virtual Edit / layer: A temporary Git worktree and branch associated with a thread. It isolates changes until you apply the layer.

Autonomy: The board for staging, running, reviewing, and organizing agents.

Room: A coordination boundary with a shared objective. A named room may also have a shared worktree inherited by its agents.

Conductor: An orchestration agent that can create rooms, decompose work, create worker tasks, start them, steer them, and stop them.

Code Atlas: A local semantic index used for code-aware search and navigation. It complements exact text search; it does not replace normal file inspection.

Skill: Reusable instructions that teach an agent a specialized workflow.

MCP server: A local or remote tool server whose discovered tools can be exposed to agents.

Safety defaults#

  • Keep write-capable and network-capable tools on Ask until you understand their scope.
  • Inspect the arguments in every approval prompt. Allow Once is the safest temporary choice.
  • Use a Virtual Edit for broad refactors, migrations, generated changes, or any task where rollback matters.
  • Treat Add Current Changes as git add -A: it can stage unrelated working-tree changes.
  • Commit only commits files that are already staged.
  • Pull is fast-forward only. Sync fast-forward pulls and then pushes the current branch.
  • A room worktree is shared by its room’s agents. Assign non-overlapping ownership before running several writers.
  • Scheduled tasks only run while Syntheo is open.
  • Privacy Veil hides the app from supported screen capture; it is not a substitute for removing secrets from prompts or logs.