DocsAgent Workflows

Providers

Supported coding agent CLIs and their capabilities

Overview#

Agentastic supports any terminal-based coding agent CLI. When you install an agent, Agentastic automatically detects it and makes it available for use. Launch agents from Agent Home, run them in isolated git worktrees or Docker containers, and review their changes with built-in diff and code review.

Structured Chat#

Beyond running a provider's CLI in a terminal, Agentastic can host a provider's session as a structured Chat transcript. Two transports do this:

Transport	Providers	How it works
Native	Claude Code, Codex	Agentastic owns the provider's own session format directly
ACP	Prime Agent, Poolside, Cursor, Grok Build, Qwen Code, Droid, OpenCode, GitHub Copilot, Auggie, Goose, Kimi, Kilocode, Devin, Kiro, Cline, Mistral Vibe, Hermes Agent, Junie, MiMo Code, Qoder CLI, Oh My Pi, CodeBuddy Code	Agentastic starts the provider's Agent Client Protocol server over stdio (for example `goose acp` or `copilot --acp`) and speaks JSON-RPC to it

A chat over either transport renders assistant text, thinking, nested tool calls, plans, and permission requests as structured events. Native Codex Chat supports mid-turn steering through App Server's turn/steer request. ACP's optional features — image attachments, session resume, mid-turn steering, slash commands, usage reporting, dynamic configuration — are negotiated with the provider during initialize, so each chat exposes only what that provider actually implements. Everything else in this page describes terminal launch behavior, which is unchanged.

Supported Agents#

Agentastic's public catalog includes 52 built-in agent definitions and auto-discovers them when installed. Any terminal agent not listed here can still be used via Settings > Connections.

Prime Agent (Prime Intellect)#

bash

curl -fsSL https://app.primeintellect.ai/prime-agent/install.sh | sh
prime-agent

Prime Intellect's open-source RLM coding and research agent combines a persistent Python environment, recursive subagents, and its Continual Harness for long-running work. On first launch, enter /login in Prime Agent and finish signing in.

Agentastic detects the prime-agent executable, passes the initial task as a positional message, and uses --continue to reopen the latest session for the current directory. An exact older session can still be opened manually with prime-agent --resume <path-or-id>. Agent Home exposes Prime Agent's native --thinking choices: off, minimal, low, medium, high, xhigh, and max. Structured Chat uses Prime Agent's ACP server through prime-agent --mode acp, and attached skills are delivered to .prime/agent/skills/<name>/. Run prime-agent update to update the CLI.

The Connections install action runs the official verified installer from the login shell selected in Settings > Terminal. It uses plain, non-interactive mode so terminal-only confirmations cannot block behind the settings sheet, defers the optional IPython runtime bootstrap until Prime Agent first needs it, and immediately refreshes prime-agent detection after npm finishes. Running the command above manually keeps Prime Agent's normal interactive installer prompts.

Prime Agent's --autonomous option is a bounded continuation mechanism, not a permission-bypass flag, so Agentastic does not present it as Auto-Approve. Prime Agent's Python and shell tools run with your user permissions; choose a worktree or container when you need an isolation boundary. See Prime Intellect's launch post and the open-source repository.

Claude Code (Anthropic)#

bash

npm install -g @anthropic-ai/claude-code

Anthropic's official CLI agent. Supports auto-approve, resume, and initial prompt.

Codex (OpenAI)#

bash

npm install -g @openai/codex

OpenAI's terminal coding agent. Supports auto-approve, resume, and initial prompt.

Warp Agent CLI#

bash

curl -fsSL https://app.warp.dev/download/agent-cli | bash

Warp's standalone agent CLI, independent of the Warp terminal app. Agentastic detects warp and uses --auto-approve for unattended runs. The CLI has no positional or flagged prompt argument and resumes only with an explicit server conversation token, so Agentastic launches the TUI and types your opening prompt once it is ready, and does not offer resume.

Poolside#

Poolside's official pool coding agent supports interactive terminal sessions, one-shot automation, and Agent Client Protocol (ACP) clients and servers.

On macOS, install pool from Terminal:

bash

curl -fsSL https://downloads.poolside.ai/pool/install.sh | sh
pool login
pool --version

The installer requires curl or wget and tar, asks you to accept the Poolside EULA, and installs to ~/.local/bin by default. If prompted, let it add that directory to your shell PATH. During pool login, choose Use Poolside Platform for free for the recommended flow, finish signing in in the browser, then paste the generated API key into the terminal. Organizations with their own deployment can instead use pool login --api-url <url>; the same login flow also supports OpenRouter and OpenAI-compatible providers. Run pool update when you want to update the CLI.

The Install action in Agentastic's Connections settings uses Poolside's documented headless-install variables to accept the EULA and add ~/.local/bin to your shell configuration. The command is shown for review before it runs. For CI, provide POOLSIDE_API_KEY; POOLSIDE_API_URL, POOLSIDE_TOKEN, POOLSIDE_STANDALONE_BASE_URL, and POOLSIDE_STANDALONE_MODEL cover deployment and OpenAI-compatible-provider setups.

Flags Agentastic uses#

Workflow	Poolside flag	Behavior
Initial prompt	`--prompt-queue <prompt>`	Queues the task until the interactive session is ready
Unattended	`--mode always-allow`	Auto-approves tool actions; explicit deny rules still apply
Plan	`--mode plan`	Plans without modifying the codebase
Resume	`-r` / `--resume`	Opens the current-directory session picker; a session ID resumes that exact session

Interactive flag reference#

Flag	Purpose
`-C`, `--directory <path>`	Set the session working directory
`-w`, `--worktree [branch]`	Create or reuse a Poolside-managed Git worktree; omit the branch to generate a name
`-r`, `--resume [session-id]`	Resume an exact session, or omit the ID to open the picker
`-m`, `--model <model>`	Override the saved model for this session
`--mode <mode>`	Select `default`, `accept-edits`, `always-allow`, or `plan`
`-s`, `--agent-server [server]`	Select an ACP server; omit the value to open the picker
`--sandbox <mode>`	Set sandbox usage to `required` or `disabled`
`-q`, `--prompt-queue <prompt>`	Queue a startup prompt; repeat the flag to queue more than one
`-h`, `--help`	Show CLI help
`-v`, `--version`	Show the installed version

The four interactive modes are: default (ask before first use of each tool type), accept-edits (auto-approve workspace reads and writes), always-allow (auto-approve tool actions), and plan (inspect and plan without changing the codebase). Agentastic already sets the working directory and manages its own worktree, so it does not add -C or -w to normal launches.

Models and thinking level#

Agentastic's model picker supplies the selected Poolside model with pool -m <model-id>:

Model	Model ID	Best suited for
Laguna S 2.1	`poolside/laguna-s-2.1`	Long-horizon reasoning and tool workflows
Laguna M.1	`poolside/laguna-m.1`	Complex, multi-step, quality-first work
Laguna XS 2.1	`poolside/laguna-xs-2.1`	Fast, efficient agentic iteration

Agent Home provides None and Max thinking choices for Poolside. Pool v1.0.14 does not expose a dedicated command-line flag for this setting, so Agentastic queues Pool's local slash command before the task:

bash

pool --prompt-queue '/thought-level max' --prompt-queue '<task>'

Pool processes the first queue entry as an ACP thought_level configuration change without sending it to the model, then sends the task. In an interactive Pool session, you can make the same change with /thought-level max or /thought-level none (/effort is an alias). ACP clients can instead set the option directly by session config ID:

yaml

agent_servers:
  Poolside:
    type: custom
    command: pool
    args:
      - acp
    default_config_options:
      model: poolside/laguna-xs-2.1
      thought_level: max

The Agent Home selection applies to that launch and is queued again when resuming the session. It does not change Pool's saved default.

Automated `pool exec` flags#

Flag	Purpose
`-p`, `--prompt <text>`	Supply prompt text; use `-` to read it from standard input
`-f`, `--prompt-file <file>`	Read the prompt from a file
`-d`, `--directory <path>`	Set the working directory
`-a`, `--agent-name <name>`	Choose an agent in tenant mode
`--api-url <url>`	Use an OpenAI-compatible API URL
`-o`, `--output markdown	json`
`--continue [run-id]`	Continue an exact run, or omit the ID to continue the latest run
`--unsafe-auto-allow`	Run without interactive tool approvals
`--sandbox required	disabled`
`--verbose`	Print verbose tool results

pool exec exits with 0 when the task succeeds and 4 when the agent explicitly reports that it could not complete the task; other codes indicate an unexpected error. See Poolside's installation guide and complete CLI reference.

Open Interpreter#

bash

curl -fsSL https://www.openinterpreter.com/install | sh

Provider-agnostic coding agent built on Codex. Agentastic detects the interpreter executable, passes an initial prompt directly to its TUI, resumes the latest session with resume --last, uses --yolo for unattended runs, selects the read-only sandbox for plan mode, and uses its non-interactive review command for Code Review. See the official docs.

Command Code (Langbase)#

bash

npm i -g command-code
command-code login

Langbase's coding agent with taste-aware workflows. Agentastic launches the canonical command-code executable (commandcode and cmdc are also detected), skips trust and onboarding prompts, supports --continue for resume, --yolo for auto-approve, --plan for plan mode, and can pass an initial prompt directly. See the official docs.

Gemini CLI (Google)#

bash

npm install -g @google/gemini-cli

Google's AI coding agent. Supports auto-approve, resume, and initial prompt.

Antigravity (Google)#

bash

curl -fsSL https://antigravity.google/cli/install.sh | bash

Google's Antigravity CLI for terminal-first agent sessions with shared Antigravity settings. Agentastic detects the agy executable, seeds the TUI with -i, and supports --dangerously-skip-permissions for unattended runs, --continue to resume, and --mode plan for plan mode. See the official docs.

Jules (Google)#

bash

npm install -g @google/jules

Google's CLI for Jules, the asynchronous coding agent. Launches the Jules TUI; Agentastic types your prompt in automatically once the TUI is ready. See the official docs.

Grok Build (xAI)#

bash

curl -fsSL https://x.ai/cli/install.sh | bash

xAI's official terminal coding agent. Agentastic detects the grok executable, launches its interactive TUI and delivers your prompt automatically, with --always-approve for unattended runs, --permission-mode plan for plan mode, and --continue to resume the most recent session. Grok Build picks up your AGENTS.md, skills, hooks, plugins, and MCP servers out of the box. See the official docs.

Muse Code (Meta)#

bash

curl -fsSL https://dev.meta.ai/install.sh | bash

Meta's terminal coding agent, powered by Muse Spark 1.2. Agentastic detects the muse executable, seeds the TUI with the prompt as a positional argument, and uses --yolo for unattended runs, resume --last to pick the session back up, and --disable-write --disable-shell for read-only plan mode (Muse ships planning as its bundled /plan skill rather than a launch flag). Every launch also passes --trust-workspace, so the run starts on its own instead of waiting on Muse's "Do you trust this workspace?" question — which would otherwise appear in each new agent worktree. That trust applies to the single run and is not saved. Muse Code reads your AGENTS.md, skills, hooks, plugins, and MCP servers, and can import skills and sessions from other agents. See the official docs.

Models and reasoning effort#

Agentastic's model picker supplies the selection with muse --model <model-id>:

Model	Model ID	Notes
Muse Spark 1.2	`muse-spark-1.2`	Meta's agentic coding model
Muse Spark 1.2 Contributor	`muse-spark-1.2-contributor`	Discounted tokens; your content may be used for product improvement

Effort maps to Muse's own --reasoning-effort flag, which accepts minimal, low, medium, high (Muse's default), xhigh, and ultra:

bash

muse --model muse-spark-1.2 --reasoning-effort xhigh

Muse's help text also lists none, but the meta provider that serves these models rejects it at launch, so Agent Home does not offer it.

Devin (Cognition)#

bash

curl -fsSL https://cli.devin.ai/install.sh | bash

Cognition's Devin CLI for local, interactive coding sessions with Devin Cloud integration. Agentastic detects the devin executable, passes the prompt as a positional argument, and uses --permission-mode=dangerous for unattended runs and --continue to resume. See the official docs.

Hermes Agent (Nous Research)#

bash

curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash

Nous Research's open-source terminal agent. Agentastic launches hermes --tui, types the initial prompt once the interactive UI is ready, uses --continue for session recovery, and uses --yolo for unattended mode. The chat -q command is intentionally not used because it is a one-shot query. See the official docs and GitHub repository.

Cursor#

bash

curl https://cursor.com/install -fsSL | bash

Cursor's terminal coding agent. Agentastic detects and launches the cursor-agent executable (Cursor's shorter agent name is ambiguous — xAI's Grok installer claims it too), supports --continue for resume, --yolo for auto-approve, and can pass an initial prompt as a positional argument. Cursor also supports --sandbox enabled|disabled for controlling file and command permissions, --mode plan|ask for alternative interaction modes, and --cloud for offloading work to Cursor's cloud infrastructure. See the official docs.

GitHub Copilot#

bash

npm install -g @github/copilot

GitHub's AI coding assistant for the terminal. Agentastic seeds the session with -i <prompt>, uses --allow-all-tools for unattended runs, and --continue to resume the most recent session.

Junie (JetBrains)#

bash

curl -fsSL https://junie.jetbrains.com/install.sh | bash

JetBrains' LLM-agnostic coding agent CLI. Currently in beta. Agentastic seeds interactive mode with --prompt, uses --brave (Brave Mode) for unattended runs, and --resume to continue the last session.

OpenHands#

bash

uv tool install openhands

Open-source AI development agent. Supports auto-approve (--always-approve), resume, and initial prompt.

Letta Code#

bash

npm install -g @letta-ai/letta-code

Memory-first coding agent with persistent context across sessions. Agentastic uses --new for fresh launches, --yolo for auto-approve, and types the prompt once the TUI is ready. Letta's --resume selects an agent rather than an exact conversation, so Agentastic does not expose it as a generic session-resume shortcut.

Cortex Code (Snowflake)#

bash

curl -LsS https://ai.snowflake.com/static/cc-scripts/install.sh | sh

Snowflake's coding agent CLI for data engineering and development. Supports auto-approve, resume, and initial prompt.

OB-1 (OpenBlock Labs)#

bash

curl -fsSL https://dashboard.openblocklabs.com/install | bash

Autonomous coding agent. Supports auto-approve, resume, and initial prompt.

Amp#

bash

npm install -g @ampcode/cli@latest

Amp's agentic coding assistant. Agentastic supplies prompts through stdin, continues the latest thread with last, supports the accepted --dangerously-allow-all unattended flag, and exposes Amp's current reasoning modes.

Droid (Factory)#

bash

curl -fsSL https://app.factory.ai/cli | sh

Factory's AI agent for software development. Supports auto-approve (--auto high), resume, and a positional initial prompt.

Aider#

bash

curl -LsSf https://aider.chat/install.sh | sh

Open-source AI pair programmer. Supports auto-approve (--yes-always); Agentastic types your prompt into the chat once it launches (aider's positional arguments are files, not a prompt).

OpenCode#

bash

npm install -g opencode-ai

Open-source terminal coding agent. Supports resume, initial prompt, and unattended operation through the OPENCODE_PERMISSION environment policy.

Goose (Block)#

bash

curl -fsSL https://github.com/block/goose/releases/download/stable/download_cli.sh | bash

Block's autonomous coding agent. Supports initial prompt and resume. Goose exposes unattended operation through GOOSE_MODE=auto, so Agentastic sets that environment value only for Auto-Approve launches.

Cline#

bash

npm install -g cline

Terminal-based coding agent. Supports auto-approve (--auto-approve true), plan mode (--plan), and a positional initial prompt.

Autohand Code#

bash

curl -fsSL https://autohand.ai/install.sh | bash

Autonomous coding agent using the ReAct pattern with Agent Skills. Supports auto-approve (--unrestricted); Agentastic types the prompt into the TUI post-launch (the CLI's -p runs one-shot command mode).

mini-SWE-agent#

bash

uv tool install mini-swe-agent

Lightweight ~100-line coding agent from the SWE-bench team. Supports auto-approve and initial prompt.

Qwen Code#

bash

npm install -g @qwen-code/qwen-code

Alibaba's terminal coding agent. Supports auto-approve, resume, and initial prompt.

Kilo Code#

bash

npm install -g @kilocode/cli

Terminal-based coding agent. Agentastic detects and launches the canonical kilo executable, supports resume and initial prompt via --prompt, and does not advertise auto-approve because the current CLI exposes no such flag.

Kiro (AWS)#

bash

curl -fsSL https://cli.kiro.dev/install | bash

AWS's spec-driven development agent. Supports auto-approve (--trust-all-tools), resume, and a positional initial prompt via its chat subcommand.

Charm#

bash

npm install -g @charmland/crush

Charmbracelet's TUI-based coding agent. Supports auto-approve (--yolo); Agentastic types the prompt into the TUI post-launch.

Auggie (Augment Code)#

bash

npm install -g @augmentcode/auggie

Enterprise coding agent with Context Engine. Supports resume and a positional initial prompt.

Rovo Dev (Atlassian)#

bash

brew install atlassian/acli/acli && acli rovodev auth login

Atlassian's development agent, run via the Atlassian CLI (acli). Supports auto-approve; Agentastic types the prompt into the TUI post-launch.

Continue#

bash

npm install -g @continuedev/cli

Open-source coding assistant. Supports auto-approve (--auto), plan mode (--readonly), resume, and initial prompt via --prompt.

Codebuff#

bash

npm install -g codebuff

AI coding agent for the terminal. Supports resume (--continue) and a positional initial prompt.

Freebuff#

bash

npm install -g freebuff

Standalone Codebuff-compatible package. It supports resume with --continue; because its only positional command is login, Agentastic launches the TUI and types the initial prompt after it is ready.

Mistral Vibe#

bash

curl -LsSf https://mistral.ai/vibe/install.sh | bash

Mistral's terminal coding agent. Supports auto-approve (--agent auto-approve), plan mode (--agent plan), resume (-c), a positional initial prompt, and model selection through VIBE_ACTIVE_MODEL.

Kimi (Moonshot AI)#

bash

uv tool install kimi-cli

Moonshot AI's coding agent. Supports auto-approve (--yolo), plan mode (--plan), resume (--continue), and initial prompt.

Pi#

bash

npm install -g @mariozechner/pi-coding-agent

Open-source coding agent. Supports resume and a positional initial prompt.

MiMo Code (Xiaomi)#

bash

npm install --global --prefix "$HOME/.local" @mimo-ai/cli@latest

Agentastic detects mimo, seeds the interactive TUI with --prompt, uses --dangerously-skip-permissions for unattended runs, and uses --continue to resume the latest session. MiMo also accepts --session <session-id> for native exact resume; Agentastic does not yet persist that ID. The current CLI does not expose a native plan-mode flag.

Qoder CLI#

bash

npm install --global --prefix "$HOME/.local" @qoder-ai/qodercli@latest

Agentastic detects qodercli, seeds the TUI with -i, uses --dangerously-skip-permissions for unattended runs, --permission-mode plan for native plan mode, and --continue for the latest session. Qoder also accepts -r <session-id> for native exact resume; Agentastic does not yet persist that ID.

OpenClaude#

bash

npm install --global --prefix "$HOME/.local" @gitlawb/openclaude@latest

Agentastic passes the prompt positionally, uses --dangerously-skip-permissions for unattended runs, --permission-mode plan for plan mode, and --continue for the latest session. OpenClaude also accepts -r <session-id> for native exact resume; Agentastic does not yet persist that ID.

Ante#

bash

curl -fsSL https://ante.run/install.sh | ANTE_INSTALL_DIR="$HOME/.local/bin" bash

Agentastic launches Ante's interactive TUI, types the initial prompt after it is ready, and uses --yolo for unattended runs. Ante resumes only by explicit ID with -r <session-id>, so Agentastic does not expose a generic latest-session Resume shortcut.

The MiMo, Qoder, OpenClaude, and Ante install commands place their executables in ~/.local/bin, which Agentastic checks directly instead of relying on a version-manager-specific global npm prefix.

Oh My Pi#

bash

curl -fsSL https://omp.sh/install | sh -s -- --binary

Agentastic forces OMP's native release binary so an installed but outdated Bun cannot block setup. It detects omp, passes the initial prompt positionally, uses --auto-approve for unattended runs, and uses --continue for the latest session. OMP also supports native exact resume with --resume <session-id>.

OpenClaw#

bash

curl -fsSL --proto '=https' --tlsv1.2 https://openclaw.ai/install-cli.sh | bash -s -- --prefix "$HOME/.local" --no-onboard

Agentastic uses OpenClaw's CLI-only installer, including its supported Node runtime, and places the wrapper in ~/.local/bin for reliable detection. It launches openclaw tui --local so the session is self-contained and does not require a running gateway, then seeds it with --message. OpenClaw accepts a stable session key through --session <key>, but it exposes no generic latest-session, auto-approve, or plan-mode flag.

CodeBuddy Code (Tencent)#

bash

npm install -g @tencent-ai/codebuddy-code

Tencent's terminal coding agent. Agentastic detects either codebuddy or the shorter cbc wrapper, passes the initial prompt positionally, uses --dangerously-skip-permissions for unattended runs, and continues the latest session with --resume.

Zero#

bash

npm install -g @gitlawb/zero

A minimal terminal coding agent. Agentastic detects zero and uses --skip-permissions-unsafe for unattended runs. Its interactive TUI takes no launch-time prompt argument, so Agentastic types the opening prompt once the interface settles.

Feature Support#

Not all agents support the same capabilities. The table below shows which features are available per provider.

Agent	Auto-Approve	Plan	Resume Latest	Initial Prompt
Prime Agent	-	-	Yes	Yes
Claude Code	Yes	Yes	Yes	Yes
Codex	Yes	Yes	Yes	Yes
Warp Agent CLI	Yes	-	-	typed
Poolside	Yes	Yes	Yes	Yes
Open Interpreter	Yes	Yes	Yes	Yes
Command Code	Yes	Yes	Yes	Yes
Gemini CLI	Yes	Yes	Yes	Yes
Antigravity	Yes	Yes	Yes	Yes
Jules	-	-	-	typed
Grok Build	Yes	Yes	Yes	Yes
Muse Code	Yes	Yes	Yes	Yes
Devin	Yes	-	Yes	Yes
Hermes Agent	Yes	-	Yes	typed
Cursor	Yes	Yes	Yes	Yes
GitHub Copilot	Yes	-	Yes	Yes
Junie	Yes	-	Yes	Yes
OpenHands	Yes	-	Yes	Yes
Letta Code	Yes	-	-	typed
Cortex Code	Yes	-	Yes	Yes
OB-1	Yes	Yes	Yes	Yes
Amp	Yes	-	Yes	stdin
Droid	Yes	-	Yes	Yes
Aider	Yes	-	-	typed
OpenCode	Yes	-	Yes	Yes
Goose	Yes	-	Yes	Yes
Cline	Yes	Yes	-	Yes
Autohand Code	Yes	-	-	typed
mini-SWE-agent	Yes	-	-	Yes
Qwen Code	Yes	Yes	Yes	Yes
Kilo Code	-	-	Yes	Yes
Kiro	Yes	-	Yes	Yes
Charm	Yes	-	-	typed
Auggie	-	-	Yes	Yes
Rovo Dev	Yes	-	-	typed
Continue	Yes	Yes	Yes	Yes
Codebuff	-	-	Yes	Yes
Freebuff	-	-	Yes	typed
Mistral Vibe	Yes	Yes	Yes	Yes
Kimi	Yes	Yes	Yes	Yes
Pi	-	-	Yes	Yes
MiMo Code	Yes	-	Yes	Yes
Qoder CLI	Yes	Yes	Yes	Yes
OpenClaude	Yes	Yes	Yes	Yes
Ante	Yes	-	-	typed
Oh My Pi	Yes	-	Yes	Yes
OpenClaw	-	-	-	Yes
CodeBuddy Code	Yes	-	Yes	Yes
Zero	Yes	-	-	typed

Feature Descriptions#

Auto-Approve - Allows skipping permission prompts for file read/write operations. Useful for hands-off workflows where the agent works autonomously.
Plan - Uses the provider's native read-only or planning mode. Agentastic hides Plan when every selected agent cannot enforce it.
Resume Latest - Re-runs the provider's current-directory or latest-session continuation command. It does not yet guarantee a specific historical session ID.
Initial Prompt - Passes instructions when the agent starts. typed means the TUI cannot accept an interactive prompt through launch arguments, so Agentastic types it once the TUI is ready; stdin means the prompt is piped in.

All flags in this table were verified against each CLI's actual --help output or parser (through 2026-08). Rovo Dev requires Atlassian authentication before its CLI responds, so its flags follow Atlassian's documentation.

Adding Agents#

Agentastic scans your system PATH for known agent CLIs. If you install a new agent:

Install the agent using its official instructions
Make sure the binary is in your PATH
Open Settings > Connections and click Refresh to re-scan

Custom Agents#

Any terminal-based tool can be used as an agent:

Go to Settings > Connections
Click Add Connection
Enter the CLI command for your agent
The agent will start in the worktree's directory

Requesting New Agents#

Missing a tool? Open an issue on GitHub or let us know on Discord.

Providers

Overview#

Structured Chat#

Supported Agents#

Prime Agent (Prime Intellect)#

Claude Code (Anthropic)#

Codex (OpenAI)#

Warp Agent CLI#

Poolside#

Install and sign in#

Flags Agentastic uses#

Interactive flag reference#

Models and thinking level#

Automated pool exec flags#

Open Interpreter#

Command Code (Langbase)#

Gemini CLI (Google)#

Antigravity (Google)#

Jules (Google)#

Grok Build (xAI)#

Muse Code (Meta)#

Models and reasoning effort#

Devin (Cognition)#

Hermes Agent (Nous Research)#

Cursor#

GitHub Copilot#

Junie (JetBrains)#

OpenHands#

Letta Code#

Cortex Code (Snowflake)#

OB-1 (OpenBlock Labs)#

Amp#

Droid (Factory)#

Aider#

OpenCode#

Goose (Block)#

Cline#

Autohand Code#

mini-SWE-agent#

Qwen Code#

Kilo Code#

Kiro (AWS)#

Charm#

Auggie (Augment Code)#

Rovo Dev (Atlassian)#

Continue#

Codebuff#

Freebuff#

Mistral Vibe#

Kimi (Moonshot AI)#

Pi#

MiMo Code (Xiaomi)#

Qoder CLI#

OpenClaude#

Ante#

Oh My Pi#

OpenClaw#

CodeBuddy Code (Tencent)#

Zero#

Feature Support#

Feature Descriptions#

Adding Agents#

Custom Agents#

Requesting New Agents#

Automated `pool exec` flags#