π¦ Sparrow
A local-first Rust agent cockpit β route, run, replay, rewind.
One event stream. Terminal UI, WebView cockpit, JSON output, or gateway β your choice.
Quick Start Β· Why Sparrow Β· Commands Β· Architecture Β· Docs Β· Releases
Sparrow is a single-binary CLI agent written in Rust. It routes each task to the cheapest capable model, keeps you in control with Git-backed checkpoints, and makes every run replayable. Local models (Ollama) are always the first hop; cloud providers are explicit fallbacks.
The project focuses on a narrow product promise: a Rust-native local cockpit where every run is visible, replayable, budgeted, and checkpointed.
Why Sparrow vs Claude Code / Codex / Aider
| Capability | Claude Code | OpenAI Codex CLI | Aider | Sparrow |
|---|---|---|---|---|
| Single static binary, no Node/Python runtime | β | β | β | β |
| Choose any provider, any model | β Anthropic | β OpenAI | β | β 38 providers |
| Local-first (Ollama OOTB) | β | β | β οΈ | β |
Git checkpoints + rewind per run |
β | β | β οΈ | β |
Budget caps (--max-cost-usd / --max-wall-secs) |
β | β | β | β |
| WebView cockpit + TUI + JSON stream | β οΈ TUI only | β οΈ TUI only | β οΈ TUI only | β all three |
| MCP server host + client | β | β οΈ | β | β |
Drop-in import of ~/.claude/ config |
n/a | β | β | β |
| Multi-agent swarm (Planner β Coder β Verifier) | β | β | β | β |
| Telegram / Discord / Slack gateways | β | β | β | β |
| Pre-commit secret scanner bundled | β | β | β | β |
Voice (speak, transcribe) |
β | β | β | β |
| Replay & share session as URL/Gist | β | β | β | β |
| Source open, MIT | β οΈ closed | β οΈ closed | β | β |
| Zero telemetry by default | β οΈ | β οΈ | β | β |
See docs/comparison/vs-competitors.md for the long form (incl. OpenCode, Hermes, Continue, Cursor).
β¨ What's New β v0.5.8
Reliability and surface-polish release β the agentic loop now completes correctly against thinking-mode providers, the Windows TUI renders cleanly, the WebView no longer leaks selector metadata into prompts, and DeepSeek-style inline tool markup is recovered instead of leaking as text.
Engine
- Multi-tool turns survive thinking-mode β a single model turn that emits N tool calls is now replayed as one assistant message with
reasoning_content+ alltool_calls, instead of N separate messages droppingreasoning_content. Fixes the HTTP 400 loop against DeepSeek / Qwen / Moonshot (and routes likeopencode-go) that left multi-file tasks half-done. - Inline tool-call markup recovery β providers that emit
<ο½ο½DSMLο½ο½invoke β¦>or<invoke name="β¦">blocks incontent(DeepSeek native format, Anthropic XML-style) now have those calls parsed and executed instead of leaked as raw text.
Windows TUI
- UTF-8 console β
SetConsoleOutputCP(65001)is set before alternate-screen entry. The mojibake (Γ’,ΓΒ·, truncated words) is gone and box-drawing/middle-dot/arrows render correctly. - Clean initial buffer β
terminal.clear()runs once at startup so stray dots from the parent shell prompt never bleed into empty panels.
WebView
- Protocol-prefix isolation β the
__agent:NAME__ROLE__BASE64__and__model:M__prefixes the console wraps around your input are now stripped at the very top of the dispatch loop. They no longer reach the LLM, the persisted conversation history, or the user-visible transcript. GET /healthzβ lightweight 200 OK liveness probe for IDE extensions and external orchestrators (no more 404).
Install & distribution
cargo install sparrow-cliv0.5.8 on crates.io.- Homebrew tap, Scoop bucket, winget manifests with real SHA256 against the v0.5.4+ release artifacts.
sparrow launchβ first-run wizard auto-detects 20+ provider keys, ranks by cost tier, offers Ollama as a free fallback.
Tier 2 hardening (carried forward from v0.5.x)
- CLI rich rendering, streaming + chat composer, TTS/STT/file_search, Memory CLI + FTS5 session search, voice commands.
- Encrypted credential fallback store; honest hardened-sandbox reporting.
- Pre-commit secret scanner; typed knowledge graph; Playwright e2e for browser/computer-use.
- Humanized French errors for HTTP 401/429/connection/config failures.
Why Explore It
| Model routing | Budget-aware fallback chains across Ollama, NVIDIA, Anthropic, OpenAI-compatible APIs, and 30+ registry entries |
| WebView cockpit | Live route/token/cost/context at http://127.0.0.1:9339/ with drawer panels, slash palette, and agent picker |
| Terminal-native | Animated TUI cockpit, sparrow run, sparrow chat, --json output, replay, memory, gateway |
| Rollback safety | Auto-checkpoint before any mutating action; sparrow rewind <id> to restore |
| Persistent context | SQLite facts + knowledge graph, SOUL-style .agent.md files, guarded skill registry, full transcripts |
| Browser/computer-use | Playwright-backed browser tool and gated screenshot/click/type computer primitive |
| Gateway | Telegram, Discord, Slack, WebSocket API β wired with honest errors, not silent failures |
Status
Sparrow is public beta with a green cross-platform CI baseline. The kernel, routing core, console surfaces, replay, checkpoints, and memory are wired and tested; external transports are being validated by early adopters.
| Area | Status | Evidence |
|---|---|---|
| CI / Rust build | β Stable | Ubuntu Β· macOS Β· Windows; fmt, clippy -D warnings, check, release builds |
| Test suite | β Stable | Full cargo test green on current master |
| Security audit | β Stable | rustsec/audit-check on all three platforms |
| Engine loop | β Stable | Event stream, task classification, fallback execution, auto-checkpoint, auto-compaction |
| WebView console | β Stable | Full cockpit β rail/drawer, typed event stream, compact highlighted code cards, themes, composer, approval modal |
| TUI cockpit | β Stable | Animated cockpit, swarm lanes, checkpoint/diff/cost panels, @ picker, history |
| Plan mode / slash | β Stable | sparrow plan, /plan, built-in commands, user/project Markdown discovery |
| Permissions / hooks | β Stable | 6 permission modes; Pre/Post lifecycle hooks for run/tool/checkpoint/compact |
| Declarative agents | β Stable | SOUL TOML + Markdown frontmatter; agent run, agent mention, CRUD |
| Skills / plugins | β Stable | Progressive references + templates; plugin manifests; CLI install/list/remove |
| Toolsets | β Stable | Toolset/risk/auth/mutation/network/exec metadata; surface filtering |
| Browser / computer-use | πΆ Alpha | Playwright driver, screenshot blocks, click/type, Linux bwrap wrapper when available |
| Security audit CLI | β Stable | sparrow security audit [--json], WebView /security |
| Sandbox policy | β Stable | Protected paths, env allowlist; Docker/SSH/Worktree backends; honest vendor errors |
| Media tools | β Stable | vision, image_generate, text_to_speech, transcribe; WebView upload/artifacts |
| GitHub Action | β Stable | action.yml, sample workflow, sparrow github review/status/logs, --dry-run |
| Context / compaction | β Stable | ContextMeter, engine auto-trigger at 120k chars, durable HandoffDoc |
| Gateway | β Stable | /status roundtrip on port 9338; run registry with real abort |
| Replay / memory | β Stable | Recorder, checkpoint, rewind, SQLite facts, knowledge graph, optional Neo4j sync, bounded MEMORY.md, session search |
| Provider routing | πΆ Alpha | Ollama + NVIDIA tested locally; 92 NVIDIA models discovered |
| First-run setup | πΆ Alpha | Conversational setup agent + interactive fallback |
| Telegram / Discord / Slack | πΈ Partial | Transport implementations exist; E2E token validation pending |
| Extra transports | π§ͺ Experimental | WhatsApp, Signal, Email, Feishu, WeCom, QQ, Teams adapters present |
| Cloud sandboxes | π§ͺ Experimental | Modal, Daytona, Vercel, Singularity β placeholder entries |
| Cross-platform release | β Stable | Linux Β· macOS Β· Windows pre-built binaries on every tag |
See docs/AUDIT.md for module-by-module proof.
Quick Start
Available today β same sparrow binary either way:
# Universal (Rust toolchain) β published on crates.io
# macOS / Linux β one-liner (pulls the latest GitHub release)
|
# Windows β PowerShell one-liner
|
Or grab a prebuilt binary directly from the latest release (Linux x86_64, macOS arm64, Windows x86_64).
Package managers (manifests ready, publishing in progress):
# macOS β Homebrew
# Windows β Scoop
&&
# Windows 11 β winget
Then:
That's the 60-second tour. No API key required β the first-run wizard offers a free provider (Groq / Gemini / NVIDIA) or local Ollama (auto-installed if missing).
Launch Sparrow:
sparrow launch runs first-launch setup when needed, then opens the WebView cockpit on
http://127.0.0.1:9339/. Use sparrow launch --tui for the terminal cockpit.
Build from source:
Run the WebView cockpit:
# β open http://127.0.0.1:9339/
Routing smoke test:
List detected providers and models:
Force a specific route:
# Local Ollama first
# Explicit NVIDIA route
# Coding / reasoning route
First Configuration
Useful environment variables:
NVIDIA_API_KEY=...
ANTHROPIC_API_KEY=...
OPENAI_API_KEY=...
GROQ_API_KEY=...
OPENROUTER_API_KEY=...
OLLAMA_HOST=http://127.0.0.1:11434
Config lives in the platform config directory (e.g. %APPDATA%\sparrow\config.toml on Windows). Sparrow never needs API keys in the repository.
Provider Routing
Sparrow keeps a static provider registry and expands it with live model discovery when credentials are available. Stored credentials added with sparrow auth add nvidia are used for discovery, so sparrow model --list can populate the NVIDIA catalog even when NVIDIA_API_KEY is not exported.
Default NVIDIA chain:
| Model | Use case |
|---|---|
meta/llama-3.1-8b-instruct |
Fast general chat and cheap smoke tests |
stepfun-ai/step-3.5-flash |
Fast backup route via NVIDIA NIM |
nvidia/nemotron-3-super-120b-a12b |
Stronger fallback for heavier tasks |
sparrow model --set nvidia resets an older pinned config back to this chain.
Common Commands
Custom slash commands can be declared as Markdown files in .sparrow/commands/*.md or %APPDATA%\sparrow\commands\*.md. User-level commands override project and built-in ones by name. Skills are also exposed as slash commands.
Architecture
user task
β
routing-need classifier
β
budget-aware fallback chain
β
Engine
think β tool β observe β emit
β
ββββββββββββΌββββββββββββ
CLI TUI WebView
JSON Gateway Recorder
Load-bearing contracts:
| File | Role |
|---|---|
src/event.rs |
Canonical event stream |
src/provider/mod.rs |
Brain abstraction |
src/router/mod.rs |
Model ranking and fallbacks |
src/engine/mod.rs |
Agent loop |
src/tools/mod.rs |
Tool contracts |
src/gateway/mod.rs |
External message routing |
Docs
| Document | Topic |
|---|---|
| docs/AUDIT.md | Module-by-module proof |
| docs/architecture.md | System architecture |
| docs/cli-reference.md | Full CLI reference |
| docs/routing.md | Routing and provider chains |
| docs/autonomy.md | Permission modes and hooks |
| docs/sandboxing.md | Sandbox policy and backends |
| docs/browser-computer.md | Playwright browser and computer-use tools |
| docs/replay.md | Replay and checkpoints |
| docs/swarm.md | Multi-agent swarm |
| docs/keyboard.md | Keyboard shortcuts |
| docs/configuration.md | Configuration reference |
| assets/brand/ | Brand assets (SVG, HTML, ASCII) |
Contributing
Before opening a PR:
Keep docs honest: mark features as Stable, Alpha, Partial, Experimental, or Planned based on tests and runnable examples. See CONTRIBUTING.md.
License
MIT β see LICENSE.