Orca
A DeepSeek-native terminal coding agent that reads code, edits files, runs commands, and verifies its own work until the task is done or it needs you.
Evidence shows a fairly complete sandbox design: SECURITY.md describes a capability resolver/execution broker, refusal rather than host-shell fallback, non-authoritative project configs, plus landlock/seccompiler dependencies; README offers suggest/auto-edit/full-auto/plan tiers and per-folder /trust. Deductions: a full-auto host-access mode exists with risk boundaries described only at documentation level; no file-edit rollback mechanism is evidenced; copyright is 'blade-deepseek' with unverified publisher, so source attribution is thin.
Centralized workspace dependency versions, some exact pins (=0.10.4 etc.) and rustls support adequate availability management. Deductions: naming inconsistency — crate 'blade-deepseek', npm package '@blade-ai/orca', SECURITY.md links to a blade-deepseek repo — indicating an incomplete rename; no evidence in the provided files of failure-message design or error-path handling.
README clearly covers multi-platform (macOS/Linux/Windows, ARM64/x64), TUI and exec/CI usage, and ACP — enough for full marks on environment fit; /plan read-only mode and a verifier command show capability boundaries. Deductions: trigger precision (when to pick which mode) and audience/scenario depth are thin, with no failure-scenario guidance.
Good information architecture: multilingual README, doc links, SECURITY/CONTRIBUTING/SUPPORT templates; MIT license consistent with Cargo.toml earns full marks. Deductions: no CHANGELOG in the repo (version only as 0.4.24 in Cargo.toml); unstable naming (blade-deepseek vs Orca vs @blade-ai/orca); missing known-limitations documentation; the curl | sh pipe install offers no checksum guidance.
The value proposition is clear (DeepSeek-native, local, prefix-cache-aware prompts, session compression/resume), giving marginal value for the DeepSeek ecosystem. Deductions: no token-cost-control evidence, and output usability is assessable only from documentation; all performance claims lack in-repo benchmarks.
README claims are partially cross-checkable against Cargo.toml (version, license, workspace layout) and SECURITY.md. Deductions: only one language variant of the README could be sampled; key claims (sandbox enforcement, npm publication) lack code-level evidence in the provided files; claims and inference are reasonably separated, with no sign of fabricated tests or provenance.
- curl | sh and PowerShell remote-script installs cannot be checksum- or integrity-verified in static review; prefer a verified installation method.
- The full-auto mode grants full host access and is described by SECURITY.md itself as advisory rather than sandboxed — confirm you understand the risk before enabling it.
- Naming inconsistency (blade-deepseek / Orca / @blade-ai/orca) suggests a complex provenance chain; verify supply-chain attestations (npm provenance, release signing) yourself.
- Publisher identity is unverified and no response SLA is promised; assess your own support needs before production use.
- No in-repo CHANGELOG or known-limitations document; consult the external changelog and release notes before upgrading.
What does this agent do, and when should you use it?
Orca (GitHub repo echoVic/orca-agent) is a DeepSeek-native coding agent built in Rust that runs locally under the MIT license. Given a task, it reads code, edits files, runs shell commands, and verifies results, continuing until the task is done or human input is needed. It speaks DeepSeek's reasoning and tool-use semantics directly, with SSE streaming, prefix-cache-friendly prompts, automatic context management, and retry logic. It exposes an interactive TUI (orca), headless execution (orca exec), plus stable JSONL, app-server, and Agent Client Protocol (ACP) contracts for editors, harnesses, and CI. Sessions support resume and fork, with synchronous/async subagents, a persistent goal mode (/goal), automatic project memory, and four permission modes (suggest, sandboxed auto-edit, full-auto, read-only plan) with per-folder trust.
After installation it connects to DeepSeek models via DEEPSEEK_API_KEY, adopting DeepSeek's reasoning/tool-use protocol directly and fully replaying returned reasoning_content across tool turns. Orca reads, searches, edits, and writes code files and runs shell commands; runtime-owned exec_command sessions can stay alive across tool calls, allocate a PTY, and receive input via write_stdin for editors, REPLs, and TUIs, while a background supervisor settles exited sessions and injects a bounded completion notification before the next model turn. Interactive TUI sessions can ask one to four structured clarification questions with described choices, previews, and multi-select. You can verify results before finishing with orca exec --verifier "cargo test" and restore headless sessions with orca exec resume SESSION_ID. It supports sync/async subagents and JavaScript workflows, with durable child continuation via resume_from; /tasks opens live transcripts and exposes stop/resume/retry controls. It learns a bounded set of durable project facts after committed turns (/remember) and retrieves only prompt-relevant facts later. Model-side, DeepSeek V4 thinking is enabled explicitly (reasoning_effort low/high/max), the deepseek-v4 family (Flash, Vision-exp, Pro) offers a 1M-token context and up to 384K output tokens, and JPEG/PNG/GIF/WebP images are accepted in the TUI and from ACP clients, with Ctrl+V clipboard attach and protocol-aware pixel previews. Configuration priority is environment variables, CLI arguments, config files, then defaults; user config lives at ~/.orca/config.toml and trusted projects can supply .orca/config.toml, AGENTS.md, rules, skills, and workflows.
- Developers who want a terminal coding agent powered by DeepSeek that reads and edits local code and runs tests to verify its own work
- CI/pipeline maintainers running tasks headlessly (e.g., "fix the failing test") with orca exec, integrating via JSONL output and typed exit codes
- Engineers using ACP-compatible editors, connecting via orca --mode=acp for an in-editor agent experience
- Teams tackling long or parallel work, using subagents and JavaScript workflows with /goal for persistent objectives
- Multi-platform users (Windows/macOS/Linux) who rely on unified sandbox permission modes and per-folder trust controls
- Workflows needing resumability, using --resume/--fork or orca exec resume to continue conversations up to a message boundary
What are this agent's strengths and limitations?
- DeepSeek-native: speaks DeepSeek reasoning/tool-use semantics directly, keeps the Chat Completions transport, and fully replays reasoning_content across tool turns, exploiting 1M context and up to 384K output windows
- Complete interface surface: one runtime host backs TUI, headless, JSONL, app-server, and ACP contracts, fitting editors, harnesses, and CI
- Detailed reliability engineering: session switches start replacements before closing runtimes, escape-driven cancellation commits terminal child state, fail-closed continuation recovery restores only digest-verified checkpoints, and budget stops settle tools and exit 4 with a typed terminal
- Granular permission model: suggest, sandboxed auto-edit, full-auto, and read-only plan modes plus per-folder trust, with AppContainer isolation and restricted sessions on Windows
- Strong multi-platform support: the npm package covers macOS/Linux/Windows on ARM64 and x64, with explicit Windows handling of PowerShell 7/5.1/cmd.exe
- Model lock-in to DeepSeek: core operation depends on DEEPSEEK_API_KEY and DeepSeek semantics (e.g., reasoning_content replay); no alternative model-provider adapter path is documented
- The README references deepseek-v4-family models (Flash, Vision-exp, Pro) and V4 thinking; availability and pricing of those models to you should be confirmed before adopting
- Advanced capabilities (sandbox, subagent continuation, workflows, memory, MCP) require a trusted project with .orca/config.toml and related configuration, creating initial setup cost
- Windows behavior branches: PowerShell 7 is preferred, restricted sessions fall back to cmd.exe when it is missing, and 5.1 is limited to modes without AppContainer isolation
- Image attachments share a 5 MiB total limit; SSH sessions without a graphical clipboard must paste or reference remote paths, and terminals without image protocols degrade to low-resolution cell previews
How do you install or deploy this agent?
Install globally via npm (macOS/Linux/Windows on ARM64 and x64): npm install -g @blade-ai/orca. Or install the native binary directly: on macOS/Linux run curl -fsSL https://orcaagent.dev/install.sh | sh; on Windows PowerShell run irm https://orcaagent.dev/install.ps1 | iex. From a Windows project directory, provision the restricted sandbox with & ([scriptblock]::Create((irm https://orcaagent.dev/install.ps1))) -SetupSandbox. Prebuilt archives are also available on GitHub Releases.
How do you use this agent?
1) Set the key: export DEEPSEEK_API_KEY=sk-... (Windows PowerShell: $env:DEEPSEEK_API_KEY = "sk-..."). 2) Run orca to open the TUI; or orca exec "fix the failing test" for headless runs; or orca exec --verifier "cargo test" "fix it" to verify before finishing. 3) Resume with orca --resume [SESSION_ID], fork with --fork SESSION_ID, or continue headless sessions with orca exec resume SESSION_ID "continue" (use --last for the most recent, --resume-at MID to stop at a message boundary). 4) In the TUI, @ searches files/skills/plugins/MCP resources; /plan for read-only planning, /goal for a persistent objective, /tasks for subagents and background commands, /trust for sandbox permissions, /config for session settings. 5) Configuration lives in ~/.orca/config.toml; tune thinking depth via reasoning_effort in config.toml or ORCA_REASONING_EFFORT.