...

In Tau, we don't promote. We act.

latest Tau installer
$npx -y @abdoknbgit/tau-installer@latest
providers / 25
AnthropicOpenAICommand CodeAntigravityOpenRouterAgentRouterVercel AI GatewayRequestyOpenCode ZenOpenCode GoFireworks AICloudflare Workers AIMistralNVIDIA NIMDeepSeekGLMMoonshot AIMiniMax AIOllamaLM StudioKiloCodeClineCline PassGitHub CopilotKiro

Why Tau?

01

Tools in every mode

Both /mode cheap and /mode normal include the same standard Tau tools by default: file operations, shell execution, search, and the core agent workflow. Cheap mode keeps that working set without loading extra capability surfaces. Normal mode can add the optional layer through /tools: workflow detection, test pairing, package safety, integration and deploy checks, browser inspection, visual audits, artifacts, diffs, diagrams, AFT, and LSP. It can also load subagents, skills, MCP servers, and plugins when the project needs them.

/tools by task
02

Provider-native lanes

Tau does not force every model through one generic server. Each provider has a lane that normalizes messages, requests, and responses into the endpoint shape that provider expects, so Anthropic, OpenAI-compatible, Antigravity, Cline, Kiro, local, and gateway models keep their native protocol instead of a flattened imitation.

distributed lanes
03

Context optimization

Tau treats cache hit rate as a design target. Stable system and tool bytes stay in the cacheable prefix, volatile session context stays after the dynamic boundary, provider lane affinity keeps repeated requests routed consistently, and each lane protects the byte identity needed for cache hits.

cache hit rate
04

Diagnostics during edits

Optional LSP tools give Tau definitions, references, hover text, symbols, call hierarchy, and diagnostics while code is still active. A broken reference can be caught in the same run instead of after the useful context has already moved on.

optional LSP
05

Signals from long logs

Lint and exception output can be huge. Tau prunes around the useful signals: the important error, the top and tail, and the nearby lines needed to act. The agent reads the failure, not a wall of repeated log noise.

log pruning
06

Snapshots and session control

Tau keeps session state project-local. Resume and branch trees carry the run state, while snapshots save working-tree checkpoints into a shadow git store separate from the real .git. Save, list, diff, and restore give time travel without rewriting project history.

sessions / snapshots
07

Self-learning loop

After meaningful work, /learned can propose one reusable lesson. Approve, edit, or skip it; approved lessons land in project memory with origin: learned, then can be viewed, edited, deleted, or turned off from the same hub.

/learned memory
08

Claude Code compatible extensions

Tau can load MCP servers, LSP integrations, agents, plugins, and skills shaped for Claude Code. Keep them off in cheap mode, or expose them in normal mode when a project benefits from that surface.

MCP / LSP / agents
optional toolsclick rows to expand
[+]Code Intelligence2

Use when code shape, symbols, or diagnostics matter during the run.

AFTRead-only code outline, symbol zoom, AST search, navigation, and diagnostics.
LSPLanguage-server code intelligence for definitions, references, hover, symbols, and call hierarchy.
[+]Project & Workflow6

Use when Tau needs repo-native workflow, testing, packages, deploys, or guidance.

ProjectWorkflowReturn repo-native build, lint, test, dev, preview, and deploy commands.
TestSearchFind likely source and test counterparts.
PackageManagerDetect the package manager and suggest safe package commands.
IntegrationHubScan database, auth, storage, integration, and secret signals.
DeployPreviewInspect deploy, preview, tunnel, hosting, and port readiness.
ToolGuideChoose the right Tau-native workflow and tool sequence.
[+]Testing & Verification4

Use when a change needs page checks, browser evidence, visual audit, or system state.

InspectSiteVerify HTTP pages, expected text, assets, and simple forms.
WebBrowserOpen URLs/local files in the native browser or capture compact HTTP/local HTML snapshots.
VisualDesignAuditScan frontend styling risks and visual verification needs.
NativeSysInfoReturn local CPU, memory, disk, load, and process summary.
[+]Artifacts2

Use when the work needs a reviewable artifact, mockup, report, diff, or patch.

ArtifactCanvasCreate browser-reviewable HTML artifacts for reports, previews, mockups, and canvases.
DiffArtifactCreate shareable browser diff artifacts and unified patch files.
[+]Diagrams1

Use when architecture or flow needs a Mermaid diagram.

MermaidRenderCreate Mermaid diagrams and browser-reviewable HTML previews.