DeepSeek-native
Built around DeepSeek reasoning, max reasoning effort by default, SSE streaming, tool-use semantics, and prefix-cache behavior. One orca exec hands off the task — no context switch.
v0.2.19 · Rust-native
A DeepSeek-native coding agent, in your terminal
Orca is a local terminal coding agent built around DeepSeek: long-context coding, multi-turn tool use, prefix-cache friendly prompts, resumable history, persistent goals, and approval-aware automation in one Rust binary.
orca deepseek-v4-pro · full-autocontext 8%
›
esc backtrack · /goal goal · ^c exit1M · v0.2.19
What you'll notice
Built around DeepSeek, not around a generic agent shell.
1M context, self-managed
A 1M token window with automatic compaction past the 80% threshold, preserving the system prompt and recent turns. Long tasks keep their context.
Persistent goal mode
Set a long-running objective with /goal; it auto-continues after each successful turn, survives restarts, and only lets the model complete or block after a goal audit.
Approval modes
Tool specs declare capabilities; reads run directly, while write, shell, network, and agent actions follow your configured approval and sandbox policy.
Skills and user input
Markdown skills can be explicitly injected with $skill ids, and the TUI can answer structured request_user_input questions without ending the turn.
Resumable history
Local JSONL transcripts support list, search, --resume, --fork, archive, and optional zstd compression.
Prefix cache
Tuned for DeepSeek prefix cache, end-to-end.
Nine rounds of real-API tuning. The wire prompt stays append-only at the byte level — system, tools, history, summary baseline — so multi-turn loops, compaction, and resume hold a stable prefix instead of redoing it each turn.
Quick Start
Install once, then just run orca.
After installation, start Orca from your terminal. The first run guides you through the DeepSeek API key, then drops you straight into the interactive coding TUI.
Install
Use npm for the fastest path, or switch to the native curl installer below.
npm install -g @blade-ai/orca
Run Orca
Launch the TUI. If no API key is configured yet, Orca opens the setup flow and saves it for future sessions.
orca
Start coding
Type the task in the interactive terminal. For automated runs later, add verifier commands from the exec workflow.
› fix the failing auth test
Common dev tasks
The loops you already run, with more memory and fewer handoffs.
Fix failing tests
Trace the failure, edit the right files, run the verifier, and keep the transcript for review.
Long refactors
Use persistent goals and automatic continuation when one turn is not enough to land the change.
Codebase archaeology
Search, read, and summarize a repo without losing the path from evidence to conclusion.
Release checks
Combine command execution, JSONL events, and release notes into an auditable final pass.
Reusable workflows
Run project or user workflows from .orca/workflows/ when a process deserves a repeatable shape.
Approval-aware automation
Let reads flow quickly while edits, shell, network, and agent actions respect policy.
Control surface
Inspect, resume, and gate every turn.
From prompt to tool call to result, Orca keeps coding runs readable, verifiable, and resumable instead of hiding them behind a black box.
read_fileglobeditgrepbashwrite_filegit_statusweb_searchsubagentWorkflowupdate_planget_goalcreate_goalupdate_goalMCPexternal
Sessions & resume
Local JSONL transcripts under ~/.orca/sessions/, with --resume to continue and --fork to branch a new run.
Verifier loop
Pass --verifier "cargo test" to gate a run on a real command; pass/fail is reported with exit code 2 on failure.
Sandbox & hooks
Permission profiles can scope filesystem and network access, while lifecycle hooks return structured JSON to deny, modify, or inject context.
Structured event stream
--output-format jsonl emits readable session, reasoning, approval, tool, workflow, and completion events for automation or audit trails.
Why teams pick Orca
A local DeepSeek workflow, not another chat window.
Regular AI chat
- Context disappears between tasks.
- File edits and test runs stay manual.
- Reasoning settings are detached from the dev loop.
Generic coding agents
- DeepSeek behavior is treated as a generic OpenAI-compatible backend.
- Long-context and prefix-cache behavior are rarely visible.
- Resume, audit, and approval contracts vary by surface.
Orca
- DeepSeek routing, max reasoning effort, and prefix-cache-friendly prompts are first-class.
- Persistent goals, workflows, and verifier gates keep long tasks moving.
- Local JSONL transcripts, resume/fork, approval policies, and sandbox hooks stay inspectable.
Command surface
One set of verbs across the real dev loop.
# Hand it a task orca exec "fix this test" # Refactor in full-auto orca exec --approval-mode full-auto "refactor the auth module" # Pick a model + gate on a verifier orca exec --model deepseek-v4-pro --verifier "cargo test" "fix the failing test" ✓ 3 files edited · cargo test passed · exit 0
Technical specs
1M
Context window, auto-compacted past the 80% threshold.
4×
Native binaries: macOS and Linux, arm64 and x64.
16
Built-in, MCP, and external tools share one spec-driven registry.
100%
Written in Rust, running as a single local binary.
Frequently asked questions
Answers before you install.
Does Orca only work with DeepSeek?
Orca is tuned for DeepSeek V4 and gives the best experience there, but the config can point at an OpenAI-compatible endpoint when your team needs that.
Will my session history be lost?
No. Orca stores local JSONL transcripts under ~/.orca/sessions/ and supports list, search, --resume, --fork, archive, and compression.
Can I keep control over edits and shell commands?
Yes. Built-in, MCP, and external tools declare capabilities, and approval modes plus sandbox profiles decide what can run automatically.
What makes persistent goals different from a long prompt?
A /goal survives restarts, auto-continues successful turns, and requires an explicit complete or blocked audit before Orca stops.
Install
Use npm, or install the native binary directly.
npm install -g @blade-ai/orcaSupported platforms: macOS arm64/x64 and Linux arm64/x64. Downloads are available on GitHub Releases.