# aid (ai-dispatch) > aid delegates execution to separate agent contexts and worktrees, returning conclusions and evidence for your decisions. Version: 11.1.0 Release date: 2026-10-14 ## Documentation - [Quick start](https://aid.agent-tools.org/quick-start/): install, research, code, review and integrate. - [Commands](https://aid.agent-tools.org/commands/): top-level commands and key run flags. - [Agents](https://aid.agent-tools.org/agents/): built-in adapters and their descriptions. - [Operating guide](https://github.com/agent-tools-org/ai-dispatch/tree/main/default-skills/aid-guide) - [Releases](https://github.com/agent-tools-org/ai-dispatch/releases) ## Install curl -fsSL https://aid.agent-tools.org/install.sh | sh Install the agent CLIs separately and authenticate them, then run aid setup. Choose an installed agent explicitly after aid advise; there is no auto agent. Routing uses third-party leaderboard evidence and observed quota; missing evidence remains unknown. Use --worktree for code changes, --verify for checks, and aid show to inspect delivery and outcome. Done or Merged is not principal acceptance. Only Verified and Delivered are successful outcomes. aid accept records explicit approval; aid merge integrates code; aid gc requires acceptance and durability. Failed work is preserved in task worktrees or shared-checkout recovery refs. ## Security Agents can access files, run shell commands and send data to providers. Use a dedicated development machine or VM and review output before merging. --dry-run previews execution. --read-only is a post-run Git snapshot check, not a process sandbox. Ignored files, paths outside the run directory and its own Git refs are outside that check. Container and sandbox support depends on the adapter. Review permissions before dispatch. ## Agents - gemini: Research, coding, web search, file editing - agy: Research, coding, web search, file editing with Antigravity CLI - qwen: Research, coding with Qwen3-Coder models - codex: Complex implementation, multi-file refactors, test suites - copilot: General coding, repo navigation, tool-assisted implementation - opencode: Simple edits, renames, type annotations, quick fixes - commandcode: General coding, multi-file implementation, model-routing across providers - kilo: Simple edits, renames, type annotations (free tier) - mimocode: Coding via Xiaomi MiMo Code CLI (opencode-family) - cursor: General coding, strong model selection, frontend - droid: Complex implementation, multi-file refactors, debugging - oz: Complex implementation, multi-file refactors (Warp) - claude: General coding, review, refactoring, research - grok: General coding via Grok CLI ## Commands - run: Dispatch one agent task with optional worktree, verification, retry, audit, context, skills, or background execution; only a successful TaskOutcome exits 0 in the foreground - errors: Inspect recent CLI errors, including requests rejected before task creation - advise: Show agent routing advice without dispatching - batch: Dispatch a dependency-aware TOML task graph - benchmark: Run the same task through multiple agents and compare results - watch: Follow live task or group progress in text mode, the TUI, or a JSONL stream - wait: Block until selected tasks or a group finish worker settlement, including delivery and verification checks; returns non-zero when any task did not succeed - board: Show the current task board (default: current project only; --all shows every project) - notifications: Print recent notifications - changelog: Read release notes - agent: Register custom agents with add and inspect built-in agent availability and related state - clean: Remove disposable logs and caches while retaining custody evidence; reclaims a task's fallback cargo target only once the directory it was keyed from is gone, and reports how many it held back - show: Inspect task state, outcome, verification, events, context, output, result, transcript, summary, audit, or diff; prints Backup: (also backup_url in --json) once an artifact backup was uploaded; --diff --branch widens the diff from the task's own commits to the whole branch - export: Export a task in a supported format - usage: Report token and usage totals - cost: Report estimated costs by group, agent, or period - stats: Report outcome-based task success, declared difficulty versus outcomes, models, failures, and usage concentration - retry: Start a new attempt using prior task context and artifacts; supersedes a non-terminal task by stopping its live worker first - merge: Merge delivered code only when its outcome is successful by default; --force overrides a failed or inconclusive verification and records the reason - accept: Accept a completed task's delivered artifact as its principal - reject: Reject a completed task while preserving every artifact - gc: Delete accepted artifacts after recursive durability proof - respond: Codex: refused; use aid retry -f "" - reply: Codex: refused; use retry with feedback - stop: Stop one task or its retry tree while preserving artifacts; --force kills it immediately - steer: Codex: refused; use aid retry -f "" (or stop then retry) - unstick: Queue a recovery nudge or request escalation - ask: Run a focused research or exploration request with optional files - query: Query an LLM directly, optionally using automatic routing - classify: Ask typed questions (noul, choice, score) about a text or file via TypeSafe Jev; prints JSON - mcp: Start AID's stdio MCP server; task payloads expose additive outcome and verify_status fields - hook: Install or invoke supported AID hooks; task hook payloads expose additive outcome and verify_status fields - config: Inspect agents, pricing, installed skills, templates, and prompt budgets - group: Create and manage workgroups; summary , finding add/list/get/update , and broadcast handle shared results and publish messages to a file; member consumption is unconfirmed - container: Build, list, or stop development containers - build: Run cargo build/check and parse/deduplicate JSON compiler errors - test: Run cargo test with trusted guarantees (zero-match is an error) - worktree: Create or list AID-managed worktrees; it does not destroy them - store: Browse, install, inspect, and update community packages - team: Manage reusable team definitions - tool: Manage reusable tool definitions - doctor: Report repository/worktree hygiene, leaked operator symlinks, and NeedsHuman agent holds; --apply repairs only those symlinks without bypassing custody - byok: Manage BYOK providers (custom OpenAI-compatible endpoints) via opencode - credential: Manage credential-pool entries - project: Initialize, inspect, or synchronize project configuration - memory: Add, search, update, version, or forget project memory - kg: Knowledge graph — temporal entity relationships - experiment: Run and inspect metric-driven iterative experiments - upgrade: Upgrade AID after checking active-task safety - setup: Configure AID, optionally set/replace the TypeSafe (Jev) macOS keychain key with hidden terminal input and classify verification (OK or FAILED: ), and install bundled resources when needed - tree: Show task ancestry and retries - help: Print this message or the help of the given subcommand(s) ## Run flags - -q, --quiet: Suppress informational output (only errors/warnings shown). Also set via AID_QUIET=1 - --repo : Or --repo-root supplies the repository anchor - -d, --dir : Sets the task working directory - -o, --output : Selects a task output path - --result-file : Requires a durable result artifact - -m, --model : aid validates requested --model parameters against the target CLI's served model list - --difficulty : [possible values: trivial, simple, moderate, complex] - --budget : [possible values: free, cheap, standard, premium] - --urgency : [possible values: background, normal, urgent] - --rigor : Proof level; critical requires enabled --verify and post-task --audit (or project defaults) [possible values: draft, standard, critical] - --egress : Data egress: any (default), local (loopback only), or private-network (loopback or RFC1918/link-local) Possible values: - any: Any provider (default). No data-locality gate - local: Only a provider whose established endpoint is loopback - private-network: Only a provider whose established endpoint is loopback or private-network [default: any] - --kind : Task category; for a bug audit use debugging --read-only (--audit is a post-task check) [possible values: research, simple-edit, complex-impl, frontend, debugging, testing, refactoring, documentation] - -w, --worktree : Create/reuse a writable task branch; for a read-only checkout use --dir - -g, --group : Assign to workgroup - --verify []: Eligible plain cargo check|test|clippy without shell syntax uses aid's build runner outside remote-build/container runs, adding --message-format=json; everything else uses sh -c in the task directory; bare --verify auto-detects, an absent flag uses the project default, and a program missing under sh -c fails verification (missing cargo on the runner is infrastructure) - --judge []: AI judge evaluates output quality - --peer-review : Dispatch peer review after completion - --retry : Permits new attempts after failure - --context ...: for source material, not extra positional arguments - --checklist ...: or --checklist-file for explicit acceptance criteria - --scope ...: to state intended files - --remote-build []: Run Cargo builds on an rbox build box (bare flag selects automatically) - --context-from ...: Inject output from previous task(s) as context - --skill ...: Methodology skills to inject - --template