Nexonoma Compare · Curated snapshot

Agentic Coding Tools

A comparison group for tools that plan, execute, and verify software tasks agentically in a repository.

Compare concrete tools against the same functional, non-functional, and commercial properties. The matrix presents evidenced differences. Weighting, recommendations, and selection will follow in Decide.

Catalog Explorer Decide with these candidates

Candidates

Select 2 to 4

Candidates	Claude Code Observed on 2026-08-24	OpenAI Codex Observed on 2026-08-24	Cursor Observed on 2026-08-24	OpenCode Observed on 2026-08-24
Functional
SurfacesProduct surfaces through which the tool can be used.	CLIIDEWeb The primary surface is the CLI, complemented by IDE and web access. 1	Desktop appCLIIDECloud Codex combines local surfaces with delegated cloud execution. 5	IDEWebMobileAPISlack The IDE is the core surface; web, mobile, API, and Slack extend remote workflows. 10 11	CLIDesktop appIDE extension Terminal, desktop app, and IDE extension cover local workflows. 15
Repository workCan read and modify files in a codebase and execute verification commands.	Supported Claude Code reads and changes files and runs repository-level commands through the CLI. 1	Supported Codex can implement repository changes and run the related checks. 6	Supported Local and remote agents can change code and execute terminal commands. 10	Supported OpenCode reads and changes files and runs shell tools in the project. 17
Plan modeProvides an explicit analytical or read-only mode before implementation.	Supported An explicit planning mode separates analysis from subsequent implementation. 2	Supported Planning mode enables structured analysis before implementation. 6	Supported A planning mode supports structured preparation of agentic changes. 9	Supported A dedicated plan agent operates with restricted tool permissions. 16
MCP integrationCan connect external tools and context through the Model Context Protocol.	Supported MCP servers can be connected as tools and data sources. 3	Supported MCP servers extend Codex with additional tools and context sources. 6	Supported MCP servers can expose local and external tools in Cursor. 12	Supported MCP servers can be configured as additional tools. 17
Delegated agentsCan use specialized subagents or parallel agents within a workflow.	Supported Custom subagents can be used with dedicated tasks and tool permissions. 4	Supported Multiple agents can work on specialized subtasks in parallel. 5	Partial Multiple remote agents are possible; the reviewed source does not describe a general configurable subagent model. 10	Supported Primary agents and subagents can be configured with dedicated roles and tools. 16
Remote runsCan execute tasks asynchronously in a managed remote environment.	Unknown The reviewed introductory documentation does not establish a clearly bounded asynchronous remote-run feature. 1	Supported Tasks can run asynchronously in a managed cloud environment. 5	Supported Background Agents continue tasks in remote environments. 10	Not supported Background subagents run within the local session; a managed remote execution environment is not documented. 16
Non-functional
Model choiceBreadth of supported model providers and models.	Anthropic models; operation through Amazon Bedrock and Google Vertex AI is also documented.1	OpenAI models optimized for Codex; no general multi-provider selection.5	Broad selection of current models from multiple providers plus automatic model selection.13	Provider and model selection through provider/model; local and external providers are configurable.16
Approval controlsGranularity for approving or blocking file, shell, network, and tool actions.	Supported Permission modes control which tool actions require approval or are allowed. 2	Supported Sandbox boundaries and approval rules control file, shell, and network access. 7	Partial Local MCP actions can require approval; remote agents execute terminal commands automatically. 10 12	Supported Rule-based permissions control access to tools and paths. 18
Execution isolationTechnical restriction of file-system, process, or network access during an agent run.	Unknown Permission modes are documented; the reviewed source does not describe a general local sandbox guarantee. 2	Supported Agent runs are technically constrained by configurable sandbox boundaries. 7	Supported Remote agents work in isolated environments; this does not apply universally to every local run. 10	Partial Granular access rules are available; the reviewed source does not promise a general operating-system sandbox. 18
Commercial
Source modelLicense and openness model of the local client or product.	Proprietary product with documented client integrations.1	CLI under Apache-2.0; app and cloud service proprietary.6	Proprietary product.9	Open source under the MIT license.19
Access modelBasic subscription, API, or usage-based access model at the assessment date.	Claude subscription or API/cloud-provider access.1	Included in supported ChatGPT plans; additional usage through credits or API-related metering.8	Free entry tier plus individual, team, and enterprise plans; agent usage is consumption-based.14	Client is freely usable; model costs depend on the selected provider or local operation.19

Claude Code

An AI-powered coding assistant for developers.

Functional

Surfaces: CLIIDEWeb
The primary surface is the CLI, complemented by IDE and web access.
1
Repository work: Supported
Claude Code reads and changes files and runs repository-level commands through the CLI.
1
Plan mode: Supported
An explicit planning mode separates analysis from subsequent implementation.
2
MCP integration: Supported
MCP servers can be connected as tools and data sources.
3
Delegated agents: Supported
Custom subagents can be used with dedicated tasks and tool permissions.
4
Remote runs: Unknown
The reviewed introductory documentation does not establish a clearly bounded asynchronous remote-run feature.
1

Non-functional

Model choice: Anthropic models; operation through Amazon Bedrock and Google Vertex AI is also documented.1
Approval controls: Supported
Permission modes control which tool actions require approval or are allowed.
2
Execution isolation: Unknown
Permission modes are documented; the reviewed source does not describe a general local sandbox guarantee.
2

Commercial

Source model: Proprietary product with documented client integrations.1
Access model: Claude subscription or API/cloud-provider access.1

OpenAI Codex

An agentic coding tool for planning, implementing, reviewing, and delegating software tasks.

Functional

Surfaces: Desktop appCLIIDECloud
Codex combines local surfaces with delegated cloud execution.
5
Repository work: Supported
Codex can implement repository changes and run the related checks.
6
Plan mode: Supported
Planning mode enables structured analysis before implementation.
6
MCP integration: Supported
MCP servers extend Codex with additional tools and context sources.
6
Delegated agents: Supported
Multiple agents can work on specialized subtasks in parallel.
5
Remote runs: Supported
Tasks can run asynchronously in a managed cloud environment.
5

Non-functional

Model choice: OpenAI models optimized for Codex; no general multi-provider selection.5
Approval controls: Supported
Sandbox boundaries and approval rules control file, shell, and network access.
7
Execution isolation: Supported
Agent runs are technically constrained by configurable sandbox boundaries.
7

Commercial

Source model: CLI under Apache-2.0; app and cloud service proprietary.6
Access model: Included in supported ChatGPT plans; additional usage through credits or API-related metering.8

Cursor

An AI-centered development environment for conversational code changes and agentic tasks.

Functional

Surfaces: IDEWebMobileAPISlack
The IDE is the core surface; web, mobile, API, and Slack extend remote workflows.
10 11
Repository work: Supported
Local and remote agents can change code and execute terminal commands.
10
Plan mode: Supported
A planning mode supports structured preparation of agentic changes.
9
MCP integration: Supported
MCP servers can expose local and external tools in Cursor.
12
Delegated agents: Partial
Multiple remote agents are possible; the reviewed source does not describe a general configurable subagent model.
10
Remote runs: Supported
Background Agents continue tasks in remote environments.
10

Non-functional

Model choice: Broad selection of current models from multiple providers plus automatic model selection.13
Approval controls: Partial
Local MCP actions can require approval; remote agents execute terminal commands automatically.
10 12
Execution isolation: Supported
Remote agents work in isolated environments; this does not apply universally to every local run.
10

Commercial

Source model: Proprietary product.9
Access model: Free entry tier plus individual, team, and enterprise plans; agent usage is consumption-based.14

OpenCode

An open-source coding agent for the terminal that helps plan, implement, and refactor code.

Functional

Surfaces: CLIDesktop appIDE extension
Terminal, desktop app, and IDE extension cover local workflows.
15
Repository work: Supported
OpenCode reads and changes files and runs shell tools in the project.
17
Plan mode: Supported
A dedicated plan agent operates with restricted tool permissions.
16
MCP integration: Supported
MCP servers can be configured as additional tools.
17
Delegated agents: Supported
Primary agents and subagents can be configured with dedicated roles and tools.
16
Remote runs: Not supported
Background subagents run within the local session; a managed remote execution environment is not documented.
16

Non-functional

Model choice: Provider and model selection through provider/model; local and external providers are configurable.16
Approval controls: Supported
Rule-based permissions control access to tools and paths.
18
Execution isolation: Partial
Granular access rules are available; the reviewed source does not promise a general operating-system sandbox.
18

Commercial

Source model: Open source under the MIT license.19
Access model: Client is freely usable; model costs depend on the selected provider or local operation.19

Sources

Numbers link each property to its evidence.