Add Local Claude / Local Codex subscription options to /agent (use my own CLI, not Warp credits)

[sydneyfunnelAIO]

Pre-submit Checks

• I have searched Warp feature requests and there are no duplicates
• I have searched Warp docs and my feature is not there

Describe the solution you'd like?

Add "Local Claude (subscription)" and "Local Codex (subscription)" options to the /agent model picker. When selected, the agent runs the user's prompt through the locally-installed claude (Claude Code) or codex (OpenAI Codex CLI) binary on disk, using whatever auth those CLIs already have set up — i.e. the user's existing Claude / OpenAI subscription. No Warp credits should be consumed for these turns.

The conversation surface in Warp should look identical to the cloud-agent path:

• Tool calls (Bash, Read, Edit, Grep, Glob, Write) render as the same Warp ToolCall cards used by cloud agents
• Multi-turn context is preserved (prior turns visible to the CLI)
• The top-bar AI status pill shows the active local model + token usage (e.g. Local Claude · 23.4k / 200k)
• A bare clear in the terminal also resets the local-agent conversation so the CLI's context window matches what the user sees

This is opt-in per-conversation via the model picker; the existing cloud-agent flow is untouched when those options aren't selected.

Is your feature request related to a problem? Please describe.

I already pay for a Claude (Claude Code) subscription and an OpenAI (Codex) subscription. When I use Warp's /agent mode, every turn also burns Warp request credits — even though my own subscriptions could already cover the call. Today there's no way to tell /agent "for this conversation, use my local CLI instead of Warp's backend."

The closest existing options don't solve this:

• BYO API key (Enable BYO Provider (OpenAI API key) on my Warp account #7422) — uses raw API tokens metered by token, not the flat-rate subscription quota I'm already paying for. Doesn't help if I'm subscribed to Claude Pro / ChatGPT Plus.
• Cloud agent with model selector — still routes through Warp and counts against Warp credits.
• Just running claude / codex directly in the terminal — works, but loses Warp's UI affordances: tool-call cards, scrollable conversation history, terminal-block context auto-injection, attachments.

The user need is "give me Warp's /agent UX, but billed against my own existing subscription, not Warp's."

Additional context

I have a working prototype on a feature branch (ahmet/local-subscription-agent) gated behind a local_subscription_agent cargo feature. Implementation sketch:

• New synthetic model ids local-claude / local-codex registered in app/src/ai/llms.rs
• Dispatch in app/src/ai/agent/api/impl.rs short-circuits local-* ids to a synthetic ResponseStream backed by a tokio task
• Task spawns the CLI as a subprocess (claude -p - / codex exec --json -), writes the prompt envelope to stdin, parses stream-json output line-by-line
• Each tool_use / tool_result content block (Claude) and command_execution item (Codex) is translated into the matching RunShellCommand / ReadFiles / Grep / FileGlobV2 / ApplyFileDiffs proto so the UI is identical to cloud
• Stderr is drained in a background task to avoid the 64 KB pipe deadlock with --verbose
• Token usage per turn is recorded in a global OnceLock<RwLock<Option<LocalUsageSnapshot>>> slot read by the workspace top-bar

Happy to file a spec PR (specs/GH<this-issue>/product.md + tech.md) once the issue is triaged with ready-to-spec if maintainers want the full spec flow before reviewing the code.

Operating system (OS)

macOS

How important is this feature to you?

4

[warp-for-oss]

@sydneyfunnelAIO

Here's what I found while triaging this issue:

Thanks for the detailed proposal. Existing BYOK and third-party Claude/Codex CLI-agent support cover adjacent workflows, but they do not appear to provide /agent model-picker entries backed by local subscription-authenticated CLIs, so this is being triaged as a distinct Agent enhancement rather than a duplicate of #4339, #7326, #9288, #9581, or #9588.

Maintainer details

I concluded that feature request to add /agent model-picker options that run local Claude Code or Codex CLIs using the reporter's existing subscriptions instead of Warp credits..

Triage

This is an enhancement request for /agent to add local Claude Code and Codex model choices that run the user-installed CLIs with the user's existing subscription auth, while preserving Warp's Agent Mode conversation and tool-call UI and avoiding Warp credit consumption for those turns.

Evidence from the current code/docs:

• Agent model choices are currently represented as server-provided ModelsByFeature / AvailableLLMs data and rendered through the existing model selector.
• /agent requests currently build a warp_multi_agent_api::Request with the selected model IDs and call ServerApi::generate_multi_agent_output; BYOK settings can be attached, but there is no local subprocess provider branch in this path.
• Warp already supports Claude Code and Codex as third-party CLI agents with detection/plugin/harness code, but those are terminal/agent-SDK integrations rather than /agent model-picker entries that stream local CLI output into Agent Mode tool-call cards.
• Public docs cover BYOK as an API-key flow and Claude/Codex as third-party terminal CLI agents; this request is distinct because it asks for subscription-backed local CLIs inside the /agent UX.

Closest related issues checked: #4339 (local LLMs/Ollama), #7326 (ACP), #9288/#9581 (BYOK availability), #9588 (credit inefficiency), and #9550 (editor state for external CLI agents). These are related but not close enough to mark this as a duplicate because they do not cover local Claude/Codex subscription CLIs as /agent model choices with Warp's native tool-call rendering.

Likely implementation areas include the model registry/selector, Agent Mode request dispatch and response conversion, CLI-agent harness reuse boundaries, conversation reset semantics, and usage/credit display. Relevant SMEs from STAKEHOLDERS are likely @zachbai for Agent Mode/blocklist UX, @abhishekp106 for model selector and execution profiles, @harryalbert for usage display, and @bnavetta/@ianhodge for agent SDK/harness overlap.

This is my automated analysis and may be incorrect. A maintainer will verify the details.

Powered by Oz

[yashs33244]

Can i pick this up ..?

[sydneyfunnelAIO]

@yashs33244 already code it.

[peicodes]

Using third-party agents in Warp should be solved by implementing ACP