A local, auditable, model-independent software-engineering agent harness.
Plan · Orchestrate · Execute
ARSY is the core platform; this repository contains ARSY CODE, its terminal interface.
Note
ARSY CODE is under active development. The CLI ships as a signed release for six Tier 1 targets and installs from npm, Homebrew, Scoop, or the release archives, while some commands and integrations remain roadmap-gated.
A capable coding agent needs more than a model connection and shell access. It must understand repository instructions, preserve long-running context, select the right capability, request approval at the correct boundary, coordinate parallel work, and leave an auditable trail.
- Provider-neutral — select models per task without changing the runtime.
- Policy-first — every effect passes through permission evaluation, sandboxing, and audit.
- Durable — sessions, events, operation results, and checkpoints survive restarts.
- Composable — support operations, MCP, skills, hooks, and
AGENTS.md. - Multi-agent — represent work as a dependency graph with explicit budgets and concurrency limits.
- Observable — keep tokens, cost, latency, edits, approvals, and verification traceable.
$ arsy
ARSY CODE · workspace: ~/code/payments · model: auto
› find the cause of the checkout timeout, make the smallest fix, then run the relevant tests
✓ read AGENTS.md and Git status
✓ mapped the checkout request flow
! approval required: run integration tests with local network access
→ approve once / approve rule / denyarsy # interactive TUI
arsy run "fix bug #42" # non-interactive execution
arsy resume <session-id> # resume a session
arsy session list # find a session to resume
arsy session show <session-id> --turns # inspect recorded turns
arsy review # review local changes
arsy auth set anthropic # store a provider credential as a handle
arsy provider list # list allowed providers
arsy model list # list allowed models
arsy config explain policy # show effective config and its source
arsy policy explain fs.write # ask whether an operation would be allowed
arsy code symbol Workspace # find a declaration
arsy mcp add docs --transport stdio --command ... # define an MCP connection
arsy serve --protocol mcp # expose operations over stdio
arsy doctor # diagnose the environmentThe full surface — session, configuration, policy, credentials, connections, extensions, evidence,
and maintenance commands — is specified in CLI and TUI surface. Run
arsy --help for the command surface available in the current build.
npm install --global @suiflex/arsy-code
arsy --versionInstalling downloads the binary for the host platform from the matching GitHub
Release and verifies its SHA-256, so the install needs network access. macOS,
glibc Linux, and Windows are supported on x86-64 and ARM64. Run arsy doctor
afterwards to inspect the active build and environment.
brew install suiflex/tap/arsy-codescoop bucket add suiflex https://github.com/suiflex/scoop-bucket
scoop install arsy-codecurl -fsSL https://github.com/suiflex/arsy-code/releases/latest/download/install.sh | shirm https://github.com/suiflex/arsy-code/releases/latest/download/install.ps1 | iexBoth scripts verify the downloaded archive's SHA-256 checksum before
installing, but not the Sigstore signature. Every release publishes a
SHA256SUMS signed through keyless Sigstore, verifiable with the bundle beside
it — see Install and verify when
that stronger guarantee matters.
arsy
arsy run <TASK> [--workspace <PATH>] [--output human|json|ci]
arsy resume <SESSION_ID> [--follow]
arsy review [REVISION] [--base <REVISION>] [--strict]
arsy doctor [--strict]
arsy update [--check] # reports the running version; ARSY does not self-update
arsy session list [--workspace-only] [--limit <N>]
arsy session show <SESSION_ID> [--turns] [--evidence]
arsy session export <SESSION_ID> [--out <PATH>] [--include-artifacts]
arsy session rewind <SESSION_ID> --to <EVENT_ID>
arsy session fork <SESSION_ID> [--at <EVENT_ID>]
arsy config explain [KEY] [--source-only]
arsy compat explain <claude|codex|omp> [--loss-only]
arsy policy explain <OPERATION> [--resource <REF>] [--actor <ID>]
arsy code symbol <NAME> [--tier auto|text] [--limit <N>]
arsy code explain <SYMBOL_ID>
arsy code references <SYMBOL_ID>
arsy code diagnostics <PATH>
arsy auth set <PROVIDER> [--handle <NAME>]
arsy auth login <PROVIDER>
arsy auth list
arsy auth remove <HANDLE> [--force]
arsy provider list [--all]
arsy model list [--provider <ID>] [--capability <NAME>]
arsy mcp list [--source <KIND>]
arsy mcp show <NAME> [--source <KIND>]
arsy mcp add <NAME> --command <CMD> [-- ARGS...]
arsy mcp add <NAME> --transport http --url <URL>
arsy mcp remove <NAME> [--scope user|workspace]
arsy mcp enable <NAME> [--scope user|workspace]
arsy mcp disable <NAME> [--scope user|workspace]
arsy mcp test <NAME> [--timeout <SECONDS>]
arsy mcp reconnect <NAME> [--all] [--timeout <SECONDS>]
arsy mcp refresh <NAME> [--all]
arsy skill list [--source <ECOSYSTEM>]
arsy hook list [--event <NAME>]
arsy plugin list [--capabilities]
arsy plugin install <SOURCE> [--scope user|workspace] [--force]
arsy plugin inspect <ID>
arsy plugin remove <ID> [--force]
arsy plugin run <ID> [--to <INPUT>]
arsy plugin refresh [ID] [--dry-run]
arsy artifact show <REF> [--max-bytes <N>]
arsy artifact export <REF> --out <PATH>
arsy memory list [--scope <SCOPE>] [--all]
arsy memory remember <CLAIM> [--scope <SCOPE>]
arsy memory forget <ID> [--to <REASON>]
arsy gc [--apply] [--retention <DURATION>]
arsy migrate [--apply] [--backup <PATH>]
arsy eval <SUITE> [--trials <N>] [--strict] [--out <PATH>]
arsy serve [--protocol mcp|acp] [--transport stdio]
arsy completions <bash|zsh|fish|powershell> # roadmap-gated: reports a diagnostic and exits 1Global options can be used with commands that support them:
--workspace <PATH>, --config <PATH>, --provider <ID>, --model <ID>,
--output human|json|ci, --no-color, --debug, --help, and --version.
Run arsy --help for the exact syntax and availability of the current build. Commands that are
roadmap-gated report a diagnostic instead of silently behaving differently.
- Complete architecture specification
- Product requirements
- System architecture
- Rust workspace and technology choices
- Competitive research and claim ledger
- Accepted ADRs
- Interactive TUI and non-interactive execution.
- Official Anthropic and OpenAI API adapters.
- File, search, patch, shell, and read-only Git operations.
- Hierarchical instructions through
AGENTS.md. - Approval policy, workspace sandbox, and audit log.
- Persistent sessions, resume, context compaction, and token/cost summaries.
- MCP client support for stdio and Streamable HTTP.
These paths exist at different maturity levels; the source-audited status is in the engineering roadmap. Durable asynchronous agents and isolated writer integration are the next P0 runtime work. Remote runners, plugin marketplaces, and a default daemon remain deferred.
- Model and operation output is always untrusted.
- Read and write are separate capabilities.
- Network, filesystem, process, and credentials have separate policies.
- Destructive commands cannot rely on generic approval.
- Secrets never enter prompts or event logs.
- Every external effect has an operation ID, status, and result evidence.
See the threat model.
| Area | Status |
|---|---|
| Product and architecture specification | Complete |
| CLI implementation | Active development |
| Distribution | Released — npm, Homebrew, Scoop, and signed release archives |
| Stable API | Not available |
ARSY CODE is an independent open-source project licensed under the MIT License. It is not affiliated with Anthropic or OpenAI.