The Observatory · 8 min mission

The Grand Comparison: Every Agent, One Matrix

Compare all eight hub agents across surface, cost, model lock, and trust — and find the one row that fits how you work.

comparisonobservatorytool-selectionmcpsandboxingpricingFact-checked 2026-06-15
On this page

Reference matrix for the eight coding agents in this hub — Claude Code, Codex, Gemini CLI, Cursor, GitHub Copilot, Windsurf (now Devin Desktop), Cline, and Roo Code — across eight dimensions: surface, context window, free tier, pricing model, models, MCP, sandboxing, and extensibility. All eight run the same agentic loop (read repo, edit files, run commands, read errors, iterate); the matrix compares the dimensions that differ. Baseline cells are the 2026-06-15 snapshot; transition and Codex/Gemini model facts were refreshed 2026-08-01.

The grand comparison matrix — eight tools, eight dimensions

The grand comparison

Eight coding agents, eight dimensions that actually decide a choice. They run the same agentic loop — the differences are where it lives, what it costs, whose model drives it, and how much it locks down. Highlight the axis that matters to you, filter the field, and click any cell for the exact spec.

Highlight an axis

Filter the field

All eight agents in the hub.

Tip: tap a column header to focus it, or an axis chip above to highlight a group.

Agent
Claude CodeAnthropic
Codex (CLI)OpenAI
Gemini CLIGoogleIndividual access moved · 2026-06-18
CursorAnysphere
GitHub CopilotGitHub / Microsoft
Windsurf → Devin DesktopCognitionRebranded · 2026-06-02
ClineCline Bot Inc.
Roo CodeRooCodeIncDiscontinued · 2026-05-15

Three recent product changes — this matrix is dated

  1. 2026-05-15

    Roo Code

    VS Code extension shut down; repo archived/read-only (final v3.54.0). Do not install — successors are Zoo Code (fork) and Cline (upstream).

  2. 2026-06-02

    Windsurf → Devin Desktop

    The Windsurf desktop app became Devin Desktop; Windsurf remains the JetBrains plugin name and in compatibility paths. Devin Local is primary, while Cascade was still documented on 2026-08-01 with no confirmed shutdown date.

  3. 2026-06-18

    Gemini CLI → Antigravity

    Gemini CLI stopped serving Google AI Pro, Ultra, and free-tier individual accounts; those users moved to Antigravity CLI. API-key, Google Cloud, and Code Assist Standard/Enterprise routes continue.

Prefer Devin Local for new work, but do not invent a Cascade shutdown date: Cognition still documented Cascade on 2026-08-01. Do not recommend installing Roo Code.

Baseline cells were checked 2026-06-15; Codex, Gemini, Roo/Zoo, and Devin/Cascade facts were refreshed 2026-08-01. Prices, model rosters, and versions move weekly — re-check live vendor pages before relying on a number.

Sort by any column, filter to one axis (surface, cost, model, trust), expand a cell for the exact string and its source, and read the transition badges on the three tools mid-migration. Dim every column except the one you care about to read one decision at a time.

The matrix, frozen for reference

Static copy of the dataset the island renders. "Per model" means the cell inherits the chosen model's value, so a single number would be wrong. Roo Code's row is historical.

ToolSurfaceContextFree tierModelsSandboxing
Claude CodeTerminal CLI1M (Haiku 200k)NoClaude onlyBuilt-in /sandbox
CodexTerminal CLIPer modelNoOpenAI GPT-5 onlyOn by default
Gemini CLITerminal CLI1MAPI-key free tier: Flash onlyGemini onlyOpt-in (--sandbox)
CursorEditor (VS Code fork)Per model (Max Mode)Yes — HobbyMulti-vendor + ComposerPer-action approval
GitHub CopilotIn-editor + cloud agentPer modelYes — Free $0Multi-vendor + AutoApproval / cloud-isolated
Windsurf / DevinEditor (VS Code fork)Per modelYes — Free $0Multi-vendor + SWE-1.xPer-action approval
ClineVS Code / JetBrains / CLIPer model (BYO)Yes — free OSSAny provider (BYO)Per-action approval
Roo CodeDiscontinued 2026-05-15Per model (BYO)Was free OSSAny provider (BYO)Per-action approval
Every agent across the eight deciding dimensions. Baseline checked 2026-06-15; Codex, Gemini, and Devin/Cascade rows refreshed 2026-08-01.

Free tier — check this cell first

The most-mistaken column. Three tools have no free tier; five have a free-to-start route. Gemini CLI's route is a Flash-only API key, while individual Google AI accounts use Antigravity CLI.

ToolFree tier?Detail
Claude CodeNoAnthropic's $0 plan excludes Claude Code; cheapest entry is Pro $20/mo
Codex (CLI)NoCheapest entry is ChatGPT Plus $20/mo or pay-as-you-go API key. Free-across-all-plans applies to *Codex Mobile*, not the CLI
Gemini CLIYes — API key, Flash only250 requests/day on the unpaid API-key tier. Google AI Pro, Ultra, and free-tier individual accounts moved to Antigravity CLI
CursorYesHobby plan, no card: limited Agent requests + Tab completions
GitHub CopilotYes — $0Limited completions + chat/agent; no cloud agent on Free
Windsurf / DevinYes — $0Light agent quota, limited models, unlimited inline edits + Tab
ClineYes — free OSSExtension is free/open-source; you pay only for inference (BYO key)
Free tier per tool, with the exact entry point. Baseline checked 2026-06-15; Gemini route refreshed 2026-08-01.

Pricing — four kinds, not just four prices

The billing model matters more than the dollar amount. Each tool falls into exactly one kind.

ToolBilling kindTiers (verified)
Claude CodeFlat subscriptionPro $20/mo ($17 annual) · Max from $100/mo · or Anthropic API per-token
CodexBundled into existing subIncluded in ChatGPT Plus/Pro/Business/Edu/Enterprise (usage-metered) · or API key
Gemini CLIFree + paid upgradesFree/paid Gemini API key · Google Cloud · Code Assist Standard/Enterprise
CursorFree + paid upgradesHobby free · Pro $20 · Teams $40/user (Pro+/Ultra named but not itemized)
GitHub CopilotUsage-metered (AI Credits)Free $0 · Pro $10 · Pro+ $39 · Max $100 · Business $19/seat · Enterprise $39/seat
Windsurf / DevinUsage-based quotasFree $0 · Pro $20 · Max $200 · Teams $40/seat · Enterprise (ACUs)
ClineBring-your-own-keyFree OSS, pay provider directly · optional at-cost managed inference · Enterprise custom
Pricing model and verified tier amounts. Baseline checked 2026-06-15; Gemini access route refreshed 2026-08-01.

Sandboxing — three trust tiers

"Sandboxed" means OS-level isolation. Editor tools instead use permission modes, optional auto-approval, and checkpoints, which are not OS-parity with the CLIs; Copilot's Autopilot explicitly overrides fine-grained approvals and auto-approves every tool call after its one-time warning.

ToolPostureControl / mechanism
CodexOS sandbox, on by defaultsandbox_mode = read-only / workspace-write (default) / danger-full-access; macOS Seatbelt, Linux/WSL2 bubblewrap
Claude CodeOS sandbox, one commandRun /sandbox; macOS Seatbelt, Linux/WSL2 bubblewrap; no native Windows (use WSL2)
Gemini CLIOS sandbox, opt-inPass --sandbox / -s or set GEMINI_SANDBOX; Docker / Podman / sandbox-exec / gVisor / LXC
GitHub CopilotCloud-isolated (agent)Cloud agent runs in an ephemeral GitHub Actions container with a 59-min cap and one branch; assigning an issue always creates one PR, while prompt-started sessions are branch-only by default and may request at most one PR later
CursorApproval gatePer-tool approve/deny + Auto-review mode (3.6) + checkpoints
ClineApproval gatePer-action approval on every edit/command; auto-approve toggle; checkpoints
Windsurf / DevinApproval gateDevin Local adds sandboxing and capability permissions; Cascade remains documented for Memories and Workflows, with no verified shutdown date
Sandbox posture, the exact control to set, and the tier. As of 2026-06-15.

MCP — universal, so compare scope not presence

All eight tools support the Model Context Protocol; it is table stakes. The differences are config path, capability scope, and limits.

ToolConfig pathNuance
Claude Code/mcpTool definitions deferred by default — only server names enter context until a tool is used
Codex~/.codex/config.tomlNative
Gemini CLI~/.gemini/settings.jsonstdio + remote
Cursor.cursor MCP configMCP + skills + hooks on paid plans
GitHub Copilot.vscode/mcp.jsonCloud agent ships GitHub + Playwright MCP tools-only (no resources/prompts, no OAuth remote)
Windsurf / Devin~/.codeium/windsurf/mcp_config.jsonFull tools/resources/prompts; 100-tool cap; MCP Marketplace
Cline~/.cline/mcp.json (CLI)MCP Marketplace one-click install; can scaffold/clone servers as an agent task
MCP config location and the meaningful per-tool nuance. As of 2026-06-15.

Pick a tool: weight the axis that matches how you work

  1. Need OS-level isolation in a terminal

    Choose a sandbox-first CLI: Codex (on by default) or Claude Code (/sandbox). Gemini CLI with --sandbox also qualifies on its supported API-key, Google Cloud, and organization-managed routes.

  2. Need zero cost to start

    Choose Cursor (Hobby), GitHub Copilot (Free), Windsurf/Devin (Free), Cline (free OSS, pay only for inference), or Gemini CLI with an unpaid API key (250 requests/day, Flash only). Avoid Claude Code and the Codex CLI — neither has a free tier. Individual Google AI accounts use Antigravity CLI.

  3. Have a model loyalty

    Single-vendor lock decides it: Claude Code (Claude), Codex (GPT-5), Gemini CLI (Gemini). Refuse all lock-in? Use Cline (any provider via BYO key).

  4. Prefer an editor over a terminal

    Choose Cursor or Windsurf/Devin (VS Code forks), or GitHub Copilot (in-editor + async cloud agent). Or run a CLI agent inside the editor over ACP.

  5. Before committing

    Re-pull the live vendor page from Sources — prices, model rosters, and version numbers move fastest, and the recent Roo, Devin, and Gemini product changes remain easy to misstate.

Knowledge check

A beginner asks which of these tools they can start using today at no cost. Which answer is correct as of 2026-08-01?

Filter the matrix by one axis at a time
… scroll to run this session
Each axis returns a different set of tools. Filter the island to one column to read a single decision.

Three recent product changes

Three of the eight tools changed identity or access inside an eight-week window. Re-pull the live vendor page before relying on any cell for these three.

Roo Code is discontinued. Repo archived read-only 2026-05-15 (final v3.54.0); cloud services offline. Do not recommend installing it. Successors: Zoo Code (community fork, inherits the role-based modes 1:1) and Cline (the upstream project Roo forked from). Roo's modes — Code, Ask, Architect, Debug, Orchestrator — remain the canonical teaching example.

The Windsurf desktop app is now Devin Desktop. Rebranded 2026-06-02; Windsurf remains the JetBrains plugin name and in compatibility paths. Devin Local is the primary local agent. Cascade remained documented on 2026-08-01 and still owns Memories and Workflows; no primary source confirmed a final shutdown date. Note: Claude Fable 5 was pulled from Devin Desktop 2026-06-12 by an export directive (the models doc still lists it — treat as unavailable there).

Gemini CLI moved individual accounts on 2026-06-18. It no longer serves Google AI Pro, Ultra, or free-tier individual accounts; those users install Antigravity CLI. Gemini CLI continues for supported API-key, Google Cloud, and Code Assist Standard/Enterprise routes, and the repo remains Apache-2.0.

Reach the end and this star joins your charted sky.