Visible execution boundaries
Commands, files, and network access run inside explicit boundaries: AppContainer on Windows and namespaces on Linux.
OPEN SOURCE · DURABLE BY DESIGN
Work survives the interruption.Proof closes the task.
Sigma Code is an open-source coding agent built for long-running work. Sessions live in a durable event stream, commands run inside native sandboxes, and a task completes only when tests, validation, and review evidence agree.
Linux x64 · StableWindows x64 · PreviewMIT licensed
01 · DURABLE SESSIONResume from the latest checkpoint
02 · NATIVE SANDBOXNative Windows / Linux isolation
03 · EVIDENCE GATE4 tests passed · sealed
57.303%Terminal-Bench 2.1
51 / 89Full run · same setup
NativeWindows / Linux sandbox
MITFree, open, auditable
THREE OPERATING PRINCIPLES
Reliability is more than saying “done.”
Model capability is only the starting point. A dependable agent needs explicit execution boundaries, recoverable state, and a completion protocol that words alone cannot bypass.
State outlives the process
Plans, tool calls, results, and checkpoints enter a durable event stream. Close the terminal or restart the process—the task keeps its place.
Completion requires evidence
Tests, validation, and review results are part of the completion protocol. Without passing evidence, the task stays open.
THE DURABLE LOOP
Every step leaves state the next run can pick up.
Sigma Code turns agent work from a one-off conversation into a recoverable execution chain. The process may stop; the working context remains.
Read the runtime design- 01PLAN
Understand
Read the code, constraints, and goal; produce a traceable plan.
- 02EXECUTE
Execute
Edit files and run commands inside a native sandbox while recording results.
- 03RESUME
Resume
Continue from the latest checkpoint without reconstructing the task.
- 04PROVE
Prove
Complete only after tests, validation, and review all agree.
ONE RUNTIME · TWO SURFACES
Focus in the terminal. See the whole system on desktop.
The same recoverable runtime powers a keyboard-first TUI and an ACP v1 desktop workflow.
TRANSPARENT EVALUATION
Publish the score. Publish the boundary.
In a diagnostic comparison across Terminal-Bench 2.1, the same DeepSeek model, and all 89 tasks, Sigma Code completed 51 tasks.
This is a reproducible engineering diagnostic, not a claim of universal superiority. The run list, failure categories, and method are auditable.
Read the evaluation methodTerminal-Bench 2.1DeepSeek · 89 tasks · same setup
AUDITABLESigma Code51 / 89
57.303%
OpenCode49 / 89
55.056%
Δ +2 completed tasksFull 89-task run
FAQ
A few important questions.
What makes Sigma Code different?
Recovery and verification are runtime protocols, not prompt conventions. State persists, and completion needs evidence.
Can a task resume after a restart?
Yes. Sessions are written to a durable event stream and resume from the latest checkpoint; granularity depends on the task stage.
Which models are supported?
Sigma Code supports multiple providers, including DeepSeek and the unified Pi provider gateway. See the project docs for authentication details.
Is there a Windows build?
v0.1.4 includes an unsigned Windows x64 preview. Linux x64 is the current stable release.
START WITH A REAL TASK
Let the next long task continue where it left off.
Free, open source, and MIT licensed. Choose a platform and start a task that can prove its result.