A model-independent coding agent with a terminal UI, local web dashboard and durable background server. Run one task, coordinate agents, or compare patches with a verified workflow—all against the same local Sessions.
Development status: the local core and first web dashboard are implemented. Broader platform, integration and release acceptance is still in progress; see Known gaps.
Coding-quality evidence: GPT-6.1 Sol / ChatGPT pilot, with measured outcomes and comparison limits; stream reliability fixes and follow-up.
Real-repository and solve pilot: original outcomes and separate follow-ups, with setup issues and comparison limits.
Requires Rust 1.89+; SQLite is bundled. On Linux, install bubblewrap for sandboxed commands (sudo apt install bubblewrap).
cargo build --locked -p cyber-cli
export OPENAI_API_KEY="your-key"
./target/debug/cyberOpenAI, Anthropic, GitHub Copilot and OpenAI-compatible endpoints are supported, including local model servers. Instead of an API key, you can sign in with a subscription:
| Subscription | Sign in | Guide |
|---|---|---|
| ChatGPT Plus/Pro | cyber providers login chatgpt, or /login in the TUI |
ChatGPT sign-in |
| GitHub Copilot | cyber providers login copilot (CLI only) |
Copilot sign-in and limits |
| Google Gemini | Not supported: third-party use of Gemini subscriptions has led to account suspensions, so /login google is refused |
— |
There is no native Gemini adapter yet (planned), so GEMINI_API_KEY alone does not enable Gemini models. Use them through a Copilot plan that serves them, or through an OpenAI-compatible gateway.
Examples below use the built binary. If you installed it on your PATH, use cyber instead of ./target/debug/cyber.
No release has been published yet, so the install script below does not work until the first one; build from source as in the Quickstart meanwhile. Tagged releases publish signed archives for macOS, Linux (glibc and musl) and Windows (preview) to GitHub Releases. The install script checks each archive's SHA-256 (and its Sigstore signature when cosign is installed); cyber upgrade always verifies both before replacing the binary.
curl -fsSL https://github.com/CyberdyneCorp/cybercode-cli/releases/latest/download/install.sh | sh
cyber upgrade # verified in-place upgrade
cyber uninstall --dry-run # what uninstall would removePackage-manager manifests (Homebrew, npm, Scoop, WinGet, AUR, Nix) are rendered for each release but not published yet. See install, upgrade and uninstall.
./target/debug/cyber service password # copy the printed login password
./target/debug/cyber web # start/reuse the server and open your browserEnter the password on the login page. The dashboard shares Sessions with the terminal: read transcripts, send prompts, review held agent messages, approve permissions, answer questions, and monitor or control Workflows, Jobs and Goals.
To open it manually, follow the URL printed by:
./target/debug/cyber web --no-openThe password grants owner access and stays in page memory. Off-loopback access requires HTTPS; a user-managed SSH forward is also supported. See the dashboard guide for remote setup, disconnect behavior and desktop/mobile screenshots. Full Settings, general attachment previews and terminal emulation remain unfinished.
./target/debug/cyber exec "fix the failing test" # one task
./target/debug/cyber review main..HEAD # read-only code review
./target/debug/cyber --continue # resume the latest Session
./target/debug/cyber exec --file shot.png "explain this screenshot"
./target/debug/cyber workflows list # bundled and saved Workflows
./target/debug/cyber workflows run audit --wait # run a bundled audit
./target/debug/cyber service status # inspect the shared server
./target/debug/cyber computer status # inspect native support and OS consent
./target/debug/cyber hex fixtures/header.bin # inspect exact bytes without inference
./target/debug/cyber doctor # check your setup
./target/debug/cyber upgrade # verified upgrade of an installed releaseThe server owns execution and stores state in SQLite, so work can outlive a terminal window. After a crash, Sessions recover interrupted state and Workflows can replay; see server lifecycle and recovery.
For independent patch attempts, host-recorded tests and a separate read-only judge, start with the verified solve workflow. Candidate patches remain available for human review; nothing is merged automatically. To combine complementary commits, use the verified integration queue: submit a dependency manifest, test the combined patch, review independent advice, then explicitly publish a branch and PR.
Reviewed rewind previews code, conversation or combined rewind with cyber sessions rewind <session> --to <message>. Applying requires the reviewed --confirm <fingerprint>; changed context is refused, unrelated edits are preserved and later messages remain on an inactive branch. Code rewind uses Git snapshots, with a private file-copy fallback for admitted Linux/macOS non-Git Locations. The TUI offers /rewind, idle Esc-Esc and reviewed /undo//redo with guarded composer restoration. File-copy restores retain recovery journals; managed recovery resolution and broader acceptance remain in progress under #93.
Experimental native computer use adds macOS 14+ application/display screenshots and input with expiring owner grants, CLI/browser owner controls, private evidence and emergency Stop. Off by default; controlled macOS capture, input and Stop acceptance passed. Other layouts and signed distribution remain release checks.
The binary workspace provides paged hex/ASCII viewing, literal hex/UTF-8 search, exact goto, bookmarks, byte copying and an LE/BE inspector. Open it with cyber hex <path>, /hex <path>, or x on an explorer file, without starting inference. Session read permissions apply. On admitted Linux/macOS workspaces, F2 prepares private staged edits with undo/redo; Ctrl+S reviews the exact byte changes, then explicitly saves. Reviewed Restore Last Save and retained draft recovery are available through the owner API/SDK; cyber hex <path> --session <id> --recover <document> reopens a kept manual draft in view mode with fresh F2 edit admission. Authorized Sessions can review shared saved history, and Session diffs show saved binary references independently of Git. The model-facing binary tool supports bounded inspect/read/search and exact reviewed patches in isolated invocation drafts. Separate owner routes and SDK helpers discover and reconcile model-operation receipts. Checkpoint/retention/rewind adapters and remaining acceptance work are tracked in #94.
Actual TUI capture of a synthetic 512-byte fixture, opened with cyber hex fixtures/firmware.bin --offset 0x20 --read-only.
Actual TUI review of a staged one-byte overwrite. Opening the review leaves the source unchanged; a second Ctrl+S applies it.
Cyber supports communication between Sessions and between different people's agents working on the same Git project.
The TUI sidebar previews named local Sessions, team/child agents and remote Sessions with status and source labels. Open /agents for the full directory and its message, watch and rename actions.
Actual TUI capture with isolated demo data. See the screenshot gallery and capture instructions.
- On one server: Sessions and agents can discover, message and watch each other. See Session messaging.
- Across your machines: direct peers connect Sessions without an account or relay.
- With teammates or friends: a self-hosted project hub connects verified members and their explicitly published agents, with conversations and shared tasks. Each person binds their own checkout to the project; tools, working trees and execution approvals stay local. Incoming agent messages are held for receiver consent by default. The hub is experimental and off by default.
# One server: message another Session by title or ses_ ID (agents use the send_message tool)
./target/debug/cyber messages list --reachable
./target/debug/cyber messages send "API refactor" "main changed; rebase and rerun the tests"
# Two of your machines: pair once, then address Sessions as <session>@<machine>
desk$ cyber peers invite laptop # prints a one-time cyberpeer:// code
laptop$ cyber peers add desk '<code>'
laptop$ cyber messages send --from reviewer "API refactor@desk" "review is ready" # held on desk until released
# Different people and accounts: bind each checkout to a hub project, then publish an agent
cyber features enable collaboration && cyber login
cyber collab bind --hub https://hub.example.com --project cprj_PROJECT --confirm
cyber collab publish --project cprj_PROJECT --session ses_SESSION --name reviewer # then /collab in the TUIMessages always arrive as untrusted text, never as instructions or approvals.
This is a map of available capabilities, including partial and experimental areas. The feature reference records their limits and acceptance status.
mindmap
root((Cyber Code))
Clients
Terminal UI
Exec and public API
Local web dashboard
TypeScript SDK
Signed releases and cyber upgrade
Agents and automation
Subagents and worktrees partial
workflows["Workflows"]
solve["Verified solve"]
integrations["Integration queue"]
Goals and loops
Agent teams experimental
Safety and recovery
Permission modes
macOS and Linux sandbox
Git snapshots
Durable Sessions and replay
OpenTelemetry export
Tools and integrations
Providers and local endpoints
MCP client and server
Hooks and skills
Image and PDF input
Binary hex/ASCII viewing
Output styles and model-family prompts
Collaboration
Local Session messages
Direct peers
Project hub experimental
Native computer use
macOS 14 plus
Owner grants and Stop
Cyber Code has working features in each area below, but it is not at full feature parity with every alternative. Checked 2026-10-24 against official documentation; this compares capabilities, not coding quality or benchmark performance. Related desktop, IDE and hosted products are distinguished from each terminal CLI.
| Compared with | What their documentation provides | What remains in Cyber Code |
|---|---|---|
| OpenCode | A broad provider catalog, IDE integration and web UI. | Native provider coverage is narrower; dedicated IDE/desktop clients are planned. The web dashboard is implemented, but full TUI-equivalent browser controls remain unfinished. |
| Claude Code CLI | The CLI is part of a terminal/IDE/desktop/web offering, with plugin marketplaces. | IDE/desktop clients, plugin marketplaces and complete cross-tool configuration/plugin compatibility remain unfinished. A plugin host, SDK, skills and hooks already exist. |
| Codex CLI | A terminal coding agent, related IDE/cloud surfaces, and a native Windows sandbox. | Windows confinement remains partial. Dedicated editor/desktop clients, hosted runners and paired-device Relay access are planned. Private HTTPS/SSH dashboard access already works. |
| Pi | A deliberately small terminal harness with TypeScript extensions, packages and an SDK. | Cyber ships its own plugin host and SDK; package marketplaces and broader extension compatibility remain unfinished. Pi intentionally leaves features such as subagents and plan mode to extensions/packages, so its smaller core is a design choice. |
Other limits to consider:
- Acceptance is ongoing: the full local-model baseline, Windows enforcement and several P1 migration, memory and code-intelligence paths remain open.
- Automation has edges: reserved budgets currently support bounded official OpenAI Responses requests; ChatGPT/Copilot subscription models use soft token/time controls. Finished-agent transcript retention and persistent scheduling are separate roadmap items; agent teams and the project hub are experimental.
- Native computer use is experimental: macOS 14+ only, off by default; controlled capture, input and Stop passed. Other layouts, live OS revocation and release signing remain unverified.
- Remote access is private: physical-device certificate trust is an operator check. Relay, paired Devices and mobile apps remain roadmap work.
Use the feature reference, P1 evidence, P2 evidence and roadmap for exact scope. A planned OpenSpec requirement is not proof of implementation. See the documentation reconciliation for the current inventory, source references and verification limits.
The TUI, web dashboard, CLI, SDK and MCP clients call one server. That server enforces permissions, schedules agents, streams events and owns durable state. See the architecture diagrams and storage decision.
| I want to… | Read |
|---|---|
| Configure providers, permissions, MCP or Sessions | Local core guide |
| Send traces, metrics and logs to my OpenTelemetry collector | OpenTelemetry export · cyber telemetry show |
| Use the browser dashboard and remote approvals | Dashboard guide |
| Inspect binary file bytes without inference | Binary workspace |
| Enable native screenshots and input | Native computer use |
| Run agents across machines or a group project | Direct peers · Project hub |
| Build or customize a workflow | Workflow guide · Verified solve · Integration queue |
| See the interface | Screenshots · Dashboard screenshots |
| Integrate another program | TypeScript SDK · Plugin SDK · OpenAPI |
| Find a command or planned capability | CLI reference · Roadmap · OpenSpec |
| Check builds, tests or evaluations | Development guide · Evaluation guide |
Run just for available development tasks and just ci for local checks. Browse the documentation index for the full guide set.
Contributions are welcome: see CONTRIBUTING.md, and report security problems privately as described in SECURITY.md. Notable changes are listed in the CHANGELOG.
Cyber Code is licensed under the Apache License 2.0.