OpenAI Codex
- Surfaces represented
- Codex CLI
- Model context represented
- GPT-5.2
Practical agent comparison
Codex vs Claude Code
Both products document repository guidance, coding tools, file modification, shell execution, and approval boundaries. Codex prompt claims are scoped mainly to Codex CLI with GPT-5.2. Claude Code documentation covers CLI and IDE-oriented surfaces without fixing one model, and it documents repository auto memory. Comparable Codex memory is not established. Before moving automation, recheck context, memory, and permission assumptions.
- Reviewed
- Products
- OpenAI Codex · Claude Code
- Decision criteria
- 9 reviewed questions
Comparison orientation
Four reviewed dimensions at a glance
Each dimension links to its detailed evidence row. State labels describe the selected evidence.
Left productOpenAI CodexOpen product record
- Documented 01Repository / project instructionsOpen detailed evidence Documented
- Not established 02Memory / persistent contextOpen detailed evidence Documented
- Documented 03Planning / work decompositionOpen detailed evidence Documented
- Documented 04Tools / tool callingOpen detailed evidence Documented
Who this comparison is for
How do the documented instruction systems of Codex and Claude Code differ?
Developers and platform teams transferring repository-based coding workflows between Codex CLI and Claude Code.
Evidence-bounded scope
Product, surface, and model context
These summaries combine only the scope stated in the reviewed answers. Each criterion keeps its narrower limits.
Claude Code
- Surfaces represented
- Claude Code terminal CLI · Claude Code IDE integrations
- Model context represented
- Not established
Practical behavior comparison
What the selected evidence establishes
“Not established” records an evidence gap. It does not mean the product lacks the capability.
| Criterion | OpenAI Codex | Claude Code | Interpretation |
|---|---|---|---|
| Repository / project instructions Repository guidance must be placed in the target agent's documented instruction hierarchy. | Documented AGENTS.md repository guidance
| Documented Scoped CLAUDE.md hierarchy
| Read practical implication |
| Memory / persistent context Durable workflow context should not be assumed to persist across agent environments. | Not established Persistent memory not established
| Documented Repository-scoped auto memory
| Read practical implication |
| Planning / work decomposition Planning assumptions shape how teams structure, review, and resume multi-step work. | Documented Making and updating plans
| Documented Plan mode before source edits
| Read practical implication |
| Tools / tool calling A portable workflow must use only the tools established by the selected evidence. | Documented Terminal and patch function calls
| Documented Read, edit, execute, and external tools
| Read practical implication |
| Filesystem / file modification Repository reads and mutations require the correct workspace and review boundary. | Documented Workspace files and patch calls
| Documented Reads and edits codebase files
| Read practical implication |
| Shell / terminal execution Command-heavy automation depends on the target harness's execution controls. | Documented Terminal command function calls
| Documented Runs commands through codebase tools
| Read practical implication |
| Browsing / network access Current external information cannot be assumed available when browsing scope is unestablished. | Documented Cached or live CLI web search
| Documented WebSearch and WebFetch
| Read practical implication |
| Approvals / permissions Consent, confirmation, and isolation rules determine which delegated actions can proceed. | Documented Configuration-dependent approval escalation
| Documented Deny, ask, allow, and permission modes
| Read practical implication |
| Context scope / inheritance Instructions and tools remain dependable only when their surface, model, and configuration boundaries are preserved. | Documented Harness, workspace, and repository context
| Documented Managed, user, project, local, and memory scopes
| Read practical implication |
Criterion-by-criterion interpretation
Scope, limitations, and evidence
Behavior Intelligence supplies the factual answers. Editorial implications remain separate.
01
Open evidence across six products How do AI agents use repository or project instructions?
Why it matters. Repository guidance must be placed in the target agent's documented instruction hierarchy.
OpenAI Codex
DocumentedCodex supports repository-scoped guidance through AGENTS.md and AGENTS.override.md in the documented context assembly. Codex home guidance loads before project guidance; project discovery selects at most one file per directory from the project root to the working directory.
- Surface
- Codex CLI
- Model context
- Not established
- Configuration / version
- Instruction discovery follows the documented repository path and size limits.
Limits and uncertainty
- The selected evidence does not establish every runtime instruction input.
Claude Code
DocumentedClaude Code documents managed, user, project, and local CLAUDE.md instruction scopes for coding-agent work.
- Surface
- Claude Code terminal CLI · Claude Code IDE integrations
- Model context
- Not established
- Configuration / version
- Not established
Limits and uncertainty
- The complete default base system prompt is not established.
02
Open evidence across six products Does the product document memory or persistent context?
Why it matters. Durable workflow context should not be assumed to persist across agent environments.
OpenAI Codex
Not establishedThe limited Codex excerpt and repository-guidance source do not establish a separate persistent memory mechanism.
- Surface
- Not established
- Model context
- Not established
- Configuration / version
- Not established
Limits and uncertainty
- Workspace context and AGENTS.md instructions are not treated as memory.
No source link is attached because the reviewed evidence does not establish this answer.
Claude Code
DocumentedClaude Code documents repository-scoped auto memory that stores selected learnings separately from human-authored CLAUDE.md instructions.
- Surface
- Claude Code terminal CLI · Claude Code IDE integrations
- Model context
- Not established
- Configuration / version
- Not established
Limits and uncertainty
- Auto memory is context, not hard policy enforcement.
03
Open evidence across six products How is planning or work decomposition documented?
Why it matters. Planning assumptions shape how teams structure, review, and resume multi-step work.
OpenAI Codex
DocumentedThe selected Codex prompt excerpt explicitly names making and updating plans as part of the assistant's working interface.
- Surface
- Codex CLI
- Model context
- GPT-5.2
- Configuration / version
- Not established
Limits and uncertainty
- The excerpt does not prescribe one universal decomposition procedure.
Claude Code
DocumentedPlan mode researches and proposes changes before source edits. It can run exploratory shell commands and write a plan. Outside sessions with bypass permissions available, source edits remain blocked until plan approval. Command review depends on mode, settings, and version; approval exits planning into the selected execution mode. CLI evidence is not a universal read-only sandbox claim.
- Surface
- Claude Code terminal CLI
- Model context
- Not established
- Configuration / version
- Current product documentation consulted 2026-09-14; enabled mode, settings, and deployment still determine availability.
Limits and uncertainty
- This claim does not update the dated prompt snapshot or establish every runtime configuration.
04
Open evidence across six products What tools or tool-calling capabilities are documented?
Why it matters. A portable workflow must use only the tools established by the selected evidence.
OpenAI Codex
DocumentedCodex's selected prompt excerpt names function calls for terminal commands and patch application.
- Surface
- Codex CLI
- Model context
- GPT-5.2
- Configuration / version
- Actual tool availability and escalation depend on the specific run configuration.
Limits and uncertainty
- The excerpt is not a complete tool inventory.
Claude Code
DocumentedClaude Code documents tools for reading code, editing files, running commands, and connecting to external tools.
- Surface
- Claude Code terminal CLI · Claude Code IDE integrations
- Model context
- Not established
- Configuration / version
- Permission rules and modes can constrain tool execution.
Limits and uncertainty
- The selected evidence does not establish that every deployment exposes every external tool.
05
Open evidence across six products Can the agent read or modify files?
Why it matters. Repository reads and mutations require the correct workspace and review boundary.
OpenAI Codex
DocumentedCodex's selected excerpt says the harness can supply workspace files and that the assistant can emit patch function calls.
- Surface
- Codex CLI
- Model context
- GPT-5.2
- Configuration / version
- Workspace access and patch approval depend on the run configuration.
Limits and uncertainty
- The selected excerpt does not establish unrestricted filesystem access.
Claude Code
DocumentedClaude Code's official overview says the product reads code, edits files, and runs commands.
- Surface
- Claude Code terminal CLI · Claude Code IDE integrations
- Model context
- Not established
- Configuration / version
- Permission rules distinguish read-only and mutating actions.
Limits and uncertainty
- Access remains bounded by the client environment and configured permissions.
06
Open evidence across six products Can the agent execute shell or terminal commands?
Why it matters. Command-heavy automation depends on the target harness's execution controls.
OpenAI Codex
DocumentedCodex's selected excerpt explicitly names function calls that run terminal commands.
- Surface
- Codex CLI
- Model context
- GPT-5.2
- Configuration / version
- Commands may require user approval depending on the run configuration.
Limits and uncertainty
- The selected excerpt does not establish unrestricted command execution.
Claude Code
DocumentedClaude Code's official overview documents command execution as part of codebase operations.
- Surface
- Claude Code terminal CLI
- Model context
- Not established
- Configuration / version
- Permission rules and modes determine when command execution is allowed or asks for approval.
Limits and uncertainty
- The exact command environment is deployment-specific.
07
Open evidence across six products Can the product browse or access web and network resources?
Why it matters. Current external information cannot be assumed available when browsing scope is unestablished.
OpenAI Codex
DocumentedInteractive Codex CLI defaults to cached web search; --search selects live web search. This tool capability does not establish unrestricted shell network access or every session configuration.
- Surface
- Codex CLI
- Model context
- Not established
- Configuration / version
- Current product documentation consulted 2026-09-14; enabled mode, settings, and deployment still determine availability.
Limits and uncertainty
- This claim does not update the dated prompt snapshot or establish every runtime configuration.
Claude Code
DocumentedWebSearch returns search titles and URLs; WebFetch retrieves and processes a page, with lossy summarization and possible truncation. Tool/domain rules and provider support apply. Amazon Bedrock does not expose the server-side search tool; an Azure-hosted Foundry deployment does not support it. Web tools do not establish arbitrary network access.
- Surface
- Claude Code terminal CLI
- Model context
- Not established
- Configuration / version
- Current product documentation consulted 2026-09-14; enabled mode, settings, and deployment still determine availability.
Limits and uncertainty
- This claim does not update the dated prompt snapshot or establish every runtime configuration.
08
Open evidence across six products What approval or permission controls are documented?
Why it matters. Consent, confirmation, and isolation rules determine which delegated actions can proceed.
OpenAI Codex
DocumentedCodex's selected excerpt says terminal and patch calls may be escalated to the user for approval depending on configuration.
- Surface
- Codex CLI
- Model context
- GPT-5.2
- Configuration / version
- The specific run determines which calls require escalation.
Limits and uncertainty
- The excerpt does not define every sandbox or approval mode.
Claude Code
DocumentedClaude Code documents deny, ask, and allow rules with precedence, plus permission modes that constrain tool execution.
- Surface
- Claude Code terminal CLI · Claude Code IDE integrations
- Model context
- Not established
- Configuration / version
- Rules can be managed at different settings scopes and by the client environment.
Limits and uncertainty
- The selected evidence does not establish every deployment policy.
09
Open evidence across six products How is context scoped or inherited across products and agents?
Why it matters. Instructions and tools remain dependable only when their surface, model, and configuration boundaries are preserved.
OpenAI Codex
DocumentedCodex's selected sources establish harness-supplied workspace context and authored guidance from Codex home followed by project instruction files. Current discovery documentation is separate from the pinned GPT-5.2 prompt excerpt.
- Surface
- Codex CLI
- Model context
- Not established
- Configuration / version
- The current instruction-discovery claim covers Codex CLI; the historical harness excerpt remains scoped to GPT-5.2.
Limits and uncertainty
- The complete runtime context assembly is not established.
Claude Code
DocumentedClaude Code documents managed, user, project, and local CLAUDE.md scopes plus separate repository-scoped auto memory.
- Surface
- Claude Code terminal CLI · Claude Code IDE integrations
- Model context
- Not established
- Configuration / version
- The reviewed evidence covers terminal and IDE-oriented product surfaces with a selectable Claude model.
Limits and uncertainty
- The complete default system prompt is not established.
Workflow guidance
Migration considerations
These implications follow from the reviewed criteria. They do not express a product preference.
Example: move a payment-service review from Codex to Claude Code
Keep the same review contract while separating team instructions, personal preferences, and the permission to edit.
- Start with three explicit requirements: the repository requires its payment tests, your personal preference is concise findings, and this task permits review only. These are example requirements, not product defaults.
- Keep the test requirement in repository AGENTS.md for Codex and project CLAUDE.md for Claude Code. Move the personal preference from Codex home AGENTS.md to Claude Code user-scope CLAUDE.md; do not commit personal guidance into the project merely to make it visible.
- Before starting Codex from a nested payment directory, check AGENTS.override.md along its root-to-working-directory chain. It can replace AGENTS.md at that level. Inspect Claude Code managed, user, project, and local context separately; renaming files does not transfer precedence.
- Request the same artifact: affected call paths, findings with file evidence, and proposed tests. Claude Code Plan mode can support analysis before source edits, but shell behavior and bypass availability still matter. The Codex pinned planning excerpt does not establish an identical permission mode.
- Handoff condition: the destination can identify the applicable team rule and produce the review artifact with the intended tool permissions. Carry unresolved findings in the task; do not assume Claude Code auto memory becomes Codex memory, or that an instruction file enforces a tool prohibition.
Keep plans and durable context explicit
A documented Codex planning interface and Claude Code auto memory do not create equivalent behavior across both products; state required artifacts and persistence needs.
Revalidate tool execution
Preserve filesystem, shell, and approval boundaries when moving automation, because tool descriptions do not imply identical execution policy.
Check external-information requirements
For a documentation lookup, require a source URL and the relevant page evidence. Check whether Codex search is live or cached; on Claude Code verify search-provider support and permission to fetch the destination domain.
Comparison-specific evidence
Differences in available evidence
Available first-party evidence covers different parts of the Codex and Claude Code instruction systems. Those evidence differences do not establish a product preference. The analysis below adds prompt and product context to the shared behavior criteria.
- 01
Available prompt evidence differs
Codex has a limited first-party prompt excerpt. The selected Claude Code sources establish an instruction architecture but not a complete default prompt.
- 02
Both document repository guidance
Codex documents repository/project guidance through AGENTS.md, while Claude Code documents scoped CLAUDE.md instructions and a separate repository auto-memory mechanism.
- 03
Both separate tools from approval controls
Both products document tool use and harness-level control boundaries, but the selected sources expose different levels of permission detail.
Source-specific detail
Product and prompt evidence
These criteria cover prompt visibility, product identity, and other evidence specific to this comparison.
| Criterion | OpenAI Codex | Claude Code |
|---|---|---|
| Product and model roleEditorial interpretation. Both products are coding agents whose active model remains separate from the product itself. | documented Codex CLI is the product surface; GPT-5.2 is the model named in the selected pinned prompt, not every Codex runtime. | documented Claude Code is a multi-surface coding product with separately selectable Claude model context. |
| Base-prompt visibilityEditorial interpretation. Evidence availability differs; absence from the selected source set is not proof that a product lacks internal base instructions. | documented Promptcred maps a limited prompt excerpt published in OpenAI's first-party repository, with completeness and rights limits attached. | not established The selected official Claude Code documentation does not establish a complete default system prompt; its public reader is architecture analysis. |
| Project instructionsEditorial interpretation. Both products document repository guidance, but Codex assembles AGENTS.md files along the project path while Claude Code documents managed, user, project, and local CLAUDE.md scopes. | documented Codex documents repository and project guidance through AGENTS.md and AGENTS.override.md files aggregated from the Git/project root to the working directory. | documented Managed, user, project, and local CLAUDE.md scopes are documented as instruction inputs. |
| ToolsEditorial interpretation. Both products expose agentic operations, but the sources describe them at different levels of detail. | documented The Codex excerpt names terminal-command and patch calls and says availability depends on configuration. | documented Claude Code documentation describes reading code, editing files, running commands, and connecting to external tools. |
| Permission controlsEditorial interpretation. Both products enforce some tool permissions through the product harness, outside the prompt text. | documented The prompt excerpt makes approval escalation configuration-dependent but does not provide a complete permission-rule taxonomy. | documented Claude Code documents deny, ask, and allow rule evaluation plus permission modes. |
| Persistent memoryEditorial interpretation. Only Claude Code's selected source set supports a repository-scoped memory claim. | not established The limited Codex excerpt does not establish a persistent memory mechanism. | documented Claude Code documents auto memory as a repository-scoped store separate from human-authored CLAUDE.md files. |
Using the evidence
How to read this evidence
Match prompt claims to the available evidence
Codex supports limited excerpt-level statements. Claude Code supports architecture claims, not a full-prompt transcription.
Inspect the context layer you depend on
Both products document repository guidance: Codex through AGENTS.md and Claude Code through scoped CLAUDE.md files. Claude Code's selected sources also establish a separate auto-memory mechanism.
Treat permissions as runtime controls
Tool descriptions do not fully define what an agent may execute. Both products retain harness and configuration boundaries.
Primary-source trail
Sources used in this comparison
Each source links to its product evidence, with provenance and publication limits preserved.
OpenAI Codex
- Tier A · vendor-publishedCodex custom instructions with AGENTS.md
- Tier A · vendor-publishedCodex CLI command reference — web search
- Tier B · first-party open sourceOpenAI Codex prompt file
- Tier A · vendor-publishedUnrolling the Codex agent loop
Claude Code
- Tier A · vendor-publishedClaude Code permission modes — Plan mode
- Tier A · vendor-publishedClaude Code tools reference — WebSearch and WebFetch
- Tier A · vendor-publishedClaude Code overview
- Tier A · vendor-publishedHow Claude remembers your project
- Tier A · vendor-publishedClaude Code permissions
Deeper cross-agent context
Behavior questions in this comparison
Open the six-product answer for broader evidence and scope.
Repository / project instructionsHow do AI agents use repository or project instructions?Memory / persistent contextDoes the product document memory or persistent context?Planning / work decompositionHow is planning or work decomposition documented?Tools / tool callingWhat tools or tool-calling capabilities are documented?Filesystem / file modificationCan the agent read or modify files?Shell / terminal executionCan the agent execute shell or terminal commands?Browsing / network accessCan the product browse or access web and network resources?Approvals / permissionsWhat approval or permission controls are documented?Context scope / inheritanceHow is context scoped or inherited across products and agents?
Continue comparing