Skip to content
Prickle
Browse docs
On this page

Which agent for which job.

What each agent app is good at, organised by the job you need done. Every claim links to the vendor’s own documentation.

View as Markdown

Start from the job.

Each job lists agents that fit and why. Often more than one fits: pick the one your team already has, with the right tools connected.

A big refactor or a long task

It needs room to plan, split the work and keep going for a long time, then hand back a diff you can review.

Claude Code
Subagents and background agents split the work, cloud sessions keep running when your computer is off, and current Claude models have a 1M-token context window by default.[1][2][3]
Codex
Cloud tasks get their own workspace and keep working while your computer is asleep. Subagents start when you ask for them.[18][19]
Grok Build
/goal stops only when reviewer subagents agree the work is done, and parallel subagents each work in their own worktree.[36][37]
Cursor
Cloud agents run in isolated virtual machines, so your computer doesn’t need to stay online.[56]

A quick fix

Small, close to the code and cheap. You want to see the change before it lands.

Cursor
Its CLI has Agent, Plan and Ask modes, and it asks before using an MCP tool by default.[55][57]
Claude Code
Plan mode blocks edits until you approve its plan. Fast mode makes Opus up to 2.5 times faster, at a higher price.[4][5]
Codex
/model picks a lighter model and reasoning effort. On Plus, the efficient model allows many more messages per five hours.[16][22]
opencode
Works with more than 75 providers and with local models, so a small job can run on a small model.[47]

Design and assets

Images, video and checking a UI. Here the app matters more than the model: some make pictures and some can’t.

ChatGPT
Generates images with gpt-image-2 in the desktop and web apps, and in Codex too.[28]
Grok
The Grok app makes images and video with Imagine, with a watermark you can’t remove. Through xAI’s API, video clips run 1 to 15 seconds at up to 1080p.[43][45]
Claude
Draws HTML or SVG diagrams and charts, but doesn’t generate photos or illustrations.[15]

Research

Reading many sources and handing back something you can check.

ChatGPT
Work mode takes multi-step research through to a document, spreadsheet or slides, in a managed cloud that keeps going after you close the app.[24][25]
Grok
xAI’s X Search tool reads posts in real time, which suits questions about what people are saying right now. Through the API it’s billed per post read.[44]
Grok Build
/deep-research runs a research task from the terminal.[37]
Claude
Research can call your connectors’ tools, so it reads from the services you’ve connected. It does so without asking each time.[12]

Reviewing a PR

A second look that can read everything and change nothing.

Codex
/review checks uncommitted changes, a commit or a branch and reports findings without touching your working tree. codex exec runs read-only unless you allow edits.[16][17]
Grok Build
Its read-only sandbox profile is meant for reviewing and auditing code.[40]
Claude Code
Plan mode explores first and blocks edits until you approve a plan. Deny rules block in every mode.[5]
Cursor
Ask mode is read-only: it explores the code without making changes.[55]

Work that runs without you watching

Mentions only notify. Prickle’s runner is intended to wake an agent for an eligible task handoff; team staging currently stays in dry run until a launch profile is verified. Know what each client does when nobody is there to approve a step.

Claude Code
claude -p prints text or JSON. A remote MCP server that needs an OAuth sign-in can’t complete it in a -p run, so sign in first.[7][8]
Codex
codex exec is read-only by default and must run inside a git repository unless you tell it otherwise.[17]
Gemini CLI
With trusted folders turned on, a run in a folder you haven’t trusted stops with an error.[33]
Grok Build
Its sandbox is off by default. Choose a profile, such as workspace, before letting it run alone.[40]

Each agent in short.

What each app is, what it does well, what to watch for, and whether it connects to Prickle today.

  • Claude Code

    Anthropic. In the terminal, VS Code, JetBrains, the desktop app and the web.

    Good at
    Long coding work: subagents, background agents and cloud sessions that keep going when your computer is off. Current models have a 1M-token context window by default.[1][2][3]
    Tools and permissions
    MCP over stdio and HTTP, with OAuth. Modes run from approving every step to auto, where a classifier reviews actions. Its sandbox covers shell commands only, and native Windows runs them unsandboxed.[5][6][7]
    Cost
    A Pro, Max, Team or Enterprise plan, or API billing. Plan usage has a five-hour and a weekly window, shared with chat.[9][14]
    Watch for
    Claude models only. Fast mode costs more and, on a plan, bills to usage credits.[3][4]

    Prickle: Connects to a local Prickle server today.

    Connect Claude Code to Prickle →

  • Claude desktop app

    Anthropic. Chat, Cowork and Code tabs on macOS and Windows, with Linux in beta.

    Good at
    Everyday work next to code. Cowork runs tasks on Anthropic’s servers while your computer sleeps; the Code tab runs Claude Code locally or in the cloud.[1][10][11]
    Tools and permissions
    Custom connectors (remote MCP) connect from Anthropic’s cloud, so the server must be public. Local servers come as desktop extensions. The Free plan allows one custom connector.[12][13]
    Cost
    The same plans and shared limits as Claude Code. The Code tab needs a paid plan.[1][14]
    Watch for
    It can’t generate photos or illustrations. Cowork tasks that use local files or the browser need the app open.[11][15]

    Prickle: Connects to a local Prickle server today.

    Connect Claude Desktop to Prickle →

  • Codex

    OpenAI. The Codex CLI, the IDE extension, the ChatGPT apps and Codex cloud.

    Good at
    Cloud tasks that keep working while your computer is asleep, local code review, and subagents when you ask for them.[16][18][19]
    Tools and permissions
    MCP over stdio and streamable HTTP, with a bearer token or OAuth. Read-only, workspace-write and full-access sandbox modes, enforced by the operating system.[20][21]
    Cost
    A ChatGPT plan: Plus and above include the CLI, IDE and web. An API key covers the CLI and IDE but not cloud tasks. Limits vary by model.[22]
    Watch for
    ChatGPT sign-in offers OpenAI models. The CLI also supports custom providers through a compatible Responses API endpoint. Cloud tasks and subagents use more of your allowance. On Linux, install bubblewrap for the sandbox.[18][21][22][23]

    Prickle: Connects to a local Prickle server today.

    Connect Codex to Prickle →

  • ChatGPT

    OpenAI. Web, desktop and mobile apps.

    Good at
    Work mode for multi-step tasks that end in a deliverable: research, documents, spreadsheets and slides. Image generation with gpt-image-2.[24][28]
    Tools and permissions
    A custom MCP server needs a public HTTPS address or a Secure MCP Tunnel. In Business and Enterprise workspaces, publishing MCP apps is off until an admin allows it.[26][27]
    Cost
    ChatGPT plans from Free and Go to Plus, Pro, Business and Enterprise.[22]
    Watch for
    It connects from OpenAI’s cloud, so a Prickle server on your own computer needs a public address or a tunnel first.[26]

    Prickle: Works on team staging over OAuth. Elsewhere it needs Prickle at a public address.

  • Gemini CLI

    Google. An open-source terminal agent.

    Good at
    Open source and scriptable, with a sandbox you can turn on, backed by Seatbelt, Docker, Podman or gVisor.[29][32]
    Tools and permissions
    MCP over stdio, SSE and HTTP, with OAuth. OAuth needs a browser on the same machine, so it fails over SSH.[31]
    Cost
    Since 18 June 2026 it no longer serves Google AI Pro, Ultra or free personal use. It works with Code Assist Standard or Enterprise and paid Gemini API keys.[30]
    Watch for
    Google points unpaid users to the Antigravity CLI and says Gemini CLI isn’t optimised for Gemini 4 Argon. Prickle doesn’t support the Antigravity CLI: Google’s terms for its consumer plans don’t allow another app to run it.[29][34]

    Prickle: Connects to a local Prickle server today.

    Connect Gemini CLI to Prickle →

  • Grok Build

    xAI. A terminal agent, also headless and over ACP.

    Good at
    Long tasks with /goal and reviewer subagents, /deep-research, and parallel subagents in their own worktrees. The grok-build-0.1 model has a 256,000-token context window.[36][37][41]
    Tools and permissions
    MCP over stdio and HTTP, and it reads Claude Code and Cursor MCP configs too. It asks by default, and rm and git push always prompt. The sandbox is off by default.[38][39][40]
    Cost
    The X Subscription (Free, Plus, Premium, Super, Ultra), with one usage pool shared across Grok and Cursor. An xAI API key also works.[35][42]
    Watch for
    Blocking network access for child processes is enforced only on Linux.[40]

    Prickle: Connects to a local Prickle server today.

    Connect Grok Build to Prickle →

  • opencode

    Anomaly (SST). An open-source terminal app, desktop app and IDE extension.

    Good at
    Choice: more than 75 providers, local models, and sign-in with some subscriptions (ChatGPT Plus or Pro, GitHub Copilot, GitLab Duo, SuperGrok). Plan and Build modes keep looking separate from changing.[46][47]
    Tools and permissions
    Local and remote MCP, with OAuth started for you. Allow, ask and deny rules, with most actions allowed by default. No sandbox is documented.[48][49]
    Cost
    Your provider’s pricing, or opencode’s own pay-as-you-go Zen or Go subscription.[50][51]
    Watch for
    MCP servers fill the context window. Its docs say Anthropic prohibits using Claude Pro or Max through opencode plugins.[47][48]

    Prickle: Connects to a local Prickle server today.

    Connect opencode to Prickle →

  • Cursor

    Anysphere. An editor, a CLI and cloud agents.

    Good at
    Working in the editor with Agent, Plan and Ask modes, and cloud agents you can start from the web, Slack, GitHub or Linear. Models from many vendors plus its own.[54][55][56]
    Tools and permissions
    MCP over stdio, SSE and HTTP, with OAuth, asking before each tool by default. Its own docs say auto-review is not a security boundary.[57][58]
    Cost
    Part of The X Subscription, with a weekly pool shared with Grok. Plus has no third-party models; Premium and above pay API list price for them.[52][53]
    Watch for
    Cloud agents need a paid plan and bill at API pricing.[56]

    Prickle: Not verified with Prickle yet.

Using this in Prickle.

What an agent can do depends as much on the tools its owner connected (a design tool, a payments account, an analytics board) as on its model. Two teammates on the same app can be good at very different jobs.

In a project room, hand the job to the teammate whose agent fits. Prickle carries the handoff through that owner’s permission rules and records it. The work runs in their app, on their plan.

When you want a second opinion, ask a teammate whose agent runs a different model from the one that did the work.

Sources

Vendor documentation, each opened on 17 October 2026. Follow the link before relying on a plan, a limit or a model name.

  1. 1.Claude Code overview Anthropic, checked
  2. 2.Claude Code on the web Anthropic, checked
  3. 3.Claude Code model configuration Anthropic, checked
  4. 4.Claude Code fast mode Anthropic, checked
  5. 5.Claude Code permission modes Anthropic, checked
  6. 6.Claude Code sandboxing Anthropic, checked
  7. 7.Connect Claude Code to tools via MCP Anthropic, checked
  8. 8.Run Claude Code programmatically Anthropic, checked
  9. 9.Manage costs in Claude Code Anthropic, checked
  10. 10.Claude Code in the desktop app Anthropic, checked
  11. 11.Get started with Cowork Anthropic, checked
  12. 12.Custom connectors using remote MCP Anthropic, checked
  13. 13.Local MCP servers on Claude Desktop Anthropic, checked
  14. 14.Understanding usage and length limits Anthropic, checked
  15. 15.Can Claude generate images? Anthropic, checked
  16. 16.Codex CLI OpenAI, checked
  17. 17.Codex non-interactive mode OpenAI, checked
  18. 18.Codex subagents OpenAI, checked
  19. 19.Codex cloud OpenAI, checked
  20. 20.Codex and MCP OpenAI, checked
  21. 21.Codex sandboxing OpenAI, checked
  22. 22.Codex pricing and usage limits OpenAI, checked
  23. 23.Codex models OpenAI, checked
  24. 24.ChatGPT on the web OpenAI, checked
  25. 25.Using ChatGPT OpenAI, checked
  26. 26.Connect a plugin to ChatGPT OpenAI, checked
  27. 27.MCP in ChatGPT OpenAI, checked
  28. 28.Image generation OpenAI, checked
  29. 29.Gemini CLI documentation Google, checked
  30. 30.Transitioning Gemini CLI to Antigravity CLI Google, checked
  31. 31.Gemini CLI MCP servers Google, checked
  32. 32.Gemini CLI sandboxing Google, checked
  33. 33.Gemini CLI trusted folders Google, checked
  34. 34.Antigravity FAQ Google, checked
  35. 35.Grok Build overview xAI, checked
  36. 36.Grok Build CLI announcement xAI, checked
  37. 37.Grok Build modes and commands xAI, checked
  38. 38.Grok Build MCP servers xAI, checked
  39. 39.Grok Build permissions xAI, checked
  40. 40.Grok Build sandbox xAI, checked
  41. 41.grok-build-0.1 model xAI, checked
  42. 42.The X Subscription pricing xAI, checked
  43. 43.Grok FAQ xAI, checked
  44. 45.Video generation xAI, checked
  45. 46.opencode documentation opencode, checked
  46. 47.opencode providers opencode, checked
  47. 48.opencode MCP servers opencode, checked
  48. 49.opencode permissions opencode, checked
  49. 50.opencode Zen opencode, checked
  50. 51.opencode Go opencode, checked
  51. 52.Cursor pricing Cursor, checked
  52. 53.Cursor usage limits Cursor, checked
  53. 54.Cursor models and pricing Cursor, checked
  54. 55.Cursor CLI overview Cursor, checked
  55. 56.Cursor cloud agents Cursor, checked
  56. 57.Cursor MCP Cursor, checked
  57. 58.Cursor run modes Cursor, checked