Terminal protocol bridge

Your terminal works as before,
the request chain knows your project.

Keep your coding tools. Connect project context and request records in one place.

Start freeSee the setup

Illustrative product interaction, not a live task or operational data.

live protocol handshake
CONNECTED
DEEPCODEX GATEWAY
  1. project resolved
  2. memory · 6 hits
  3. model routed
  4. metering attached
tools stay local
MCPlocal
Skillslocal
Hookslocal
Sandboxlocal
LATENCY42 ms
CONTEXTinjected
TOOLSpreserved

Pick a terminal to follow its protocol flow

How it works

Codex CLI and Claude Code keep their tools, permissions and local execution; DeepCodex adds project context, routing and metering only where the model request passes through.

Compatibility contract

Who owns what, stated plainly

DeepCodex does not clone either terminal or take over local tools. It fits the orchestration you already use and enhances the part the model request passes through.

Terminal native

Tools and execution

Agent runs, Skills, MCP, Hooks, approvals, the sandbox and worktrees stay with Codex CLI or Claude Code.

DeepCodex adds

Context and model

Project profile, task memory and hard-task enhancement injected into the request, routed between the fast and deep model by stage.

One view

Usage and results

Every parent task and child request traces back to model, equivalent tokens, cost, TTFT, duration and failure reason.

Feature map

Familiar capabilities, same model chain

Names follow each terminal's own concepts; DeepCodex describes only what it adds on the request side.

Project instructions
Codex CLIAGENTS.md
Claude CodeCLAUDE.md

A per-project profile with the stable stack, conventions and pitfalls

Dedicated agent runs
Codex CLISubagents
Claude CodeSubagents

Model mix, quota and per-call tracing for concurrent requests

Reusable flows
Codex CLISkills
Claude CodeSkills

Keeps the terminal flow; picks the fast or deep model by stage

External tools
Codex CLIMCP
Claude CodeMCP

Tools stay connected by the terminal; model requests land in one ledger

Execution boundary
Codex CLISandbox + approvals
Claude CodePermissions + hooks

Never crosses terminal permissions; adds key, project and budget bounds

Large-scale parallelism
Codex CLISubagent batches
Claude CodeAgent teams + workflows

Groups concurrency, models, equivalent tokens and results by parent task

Experimental capabilities such as Agent Teams and dynamic workflows depend on the terminal's current version and configuration; DeepCodex does not change the terminal's permission decisions.

Terminal coverage

From two flagship terminals to the whole coding toolchain

Each terminal is labelled by its real protocol and configuration; an integration limited by the terminal is never dressed up as full support.

  • ResponsesSupportedAGENTS.md · Subagents · Skills · MCP · Sandbox + approvalsPro Plus
  • Anthropic MessagesBetaCLAUDE.md · Subagents · Skills · MCP · Hooks · PermissionsPro Plus
  • Chat CompletionsSupportedAGENTS.md · MCP · Custom providerNo published plan offers this client
  • Chat CompletionsSupportedDiff edits · Pair modeNo published plan offers this client
  • Chat CompletionsSupportedCustom provider · Chat · AutocompleteNo published plan offers this client
  • Chat CompletionsSupportedOpenAI-compatible · MCP · Browser toolsNo published plan offers this client
  • Chat CompletionsLimitedChat专业版

Goal-driven execution

You state the goal, the system organises the path

A goal is not a longer prompt; it is a checkable delivery contract with target, boundary, verification and budget settled first.

  • Goal: what must be delivered
  • Boundary: what may be read, changed or committed
  • Verification: test commands, checks and success criteria
  • Budget: concurrency, equivalent tokens and stop conditions
Goal: harden the auth module
Done when: tests pass · no high-risk findings · PR opened
  1. 01
    Understand
    Read the project profile and change scope
  2. 02
    Split
    Dispatch review, implementation and test runs
  3. 03
    Execute
    Combine the fast and deep model by role
  4. 04
    Verify
    Run checks and shape a reviewable delivery

Orchestration patterns

From a single agent run to repeatable large workflows

Different task sizes suit different structures. DeepCodex covers all of them with the same project context and metering.

Dedicated agent runs

Hand exploration, implementation, tests or review to an isolated context; the main thread only collects results.

For 2–6 clear roles

Isolated parallelism

Several tasks advance in separate workspaces so the same files are never edited at once.

For splittable changes

Dynamic workflows

Let a script own loops, branches and intermediate results; the model handles only the current step.

For audits and large migrations

Cross-verification

Independent agent runs produce separately; a summariser compares the evidence and resolves conflicts.

For high-risk engineering work

Keep your terminal, make the model chain project-aware

Keep the agent runs, Skills and tooling you know in Codex CLI or Claude Code, and gain the model mix, project memory and transparent metering.

Start freeSee the setup