omp-claude-bridge

omp extension that integrates Claude Code via the Agent SDK. It is a port of pi-claude-bridge by Eli Dickinson (from its 0.8.0 release) to omp's provider contract; all the streaming, MCP tool bridging, session sync and resume machinery is upstream's.

1. Provider — Use Claude Opus, Sonnet and Haiku as models in omp, with all tool calls flowing through omp's TUI
2. AskClaude tool — Delegate tasks or questions to Claude Code when using another provider

Using pi rather than omp? Install upstream pi-claude-bridge. This package registers only an omp entry and follows omp's system-prompt, message and event shapes.

Subscription terms. The bridge drives an unmodified Claude Code (claude -p through the Agent SDK) logged in with your own Claude subscription, which Anthropic support confirmed is permitted (pi-claude-bridge#118). It does not use omp's own Anthropic OAuth login. Anthropic has announced and then unannounced a billing change for Agent SDK tools; it currently uses your regular subscription quota just like Claude Code.

 

Install

omp plugin install @thealtoclef/omp-claude-bridge

Verified against omp 18.8.9. Claude Code must be logged in (claude → /login) with your subscription.

Provider

Use /model to select claude-bridge/opus, claude-bridge/sonnet or claude-bridge/haiku. These are Claude Code's model aliases, passed to it as-is: Claude Code resolves each to the newest model of its family (today Opus 5.5, Sonnet 5.5 and Haiku 5.5), so a new model arrives when Claude Code moves the alias, with no bridge or omp update. To pin a version, set Claude Code's ANTHROPIC_DEFAULT_OPUS_MODEL, ANTHROPIC_DEFAULT_SONNET_MODEL or ANTHROPIC_DEFAULT_HAIKU_MODEL for the omp process.

Behind the scenes, omp's tools are bridged to Claude Code over MCP, but it should all work like normal in omp. Bash commands get a 120-second default timeout (matching Claude Code's default). Steering works mid-turn: a message sent while Claude is running a tool reaches it at that tool boundary, not after the whole turn finishes.

Context and thinking: the aliased models run with a 1M context window by default (no [1m] suffix needed) and 128K output, and take every Claude Code effort level: low, medium, high, xhigh and max. The thinking level you pick in omp is sent as Claude Code's effort; omp's minimal is sent as low.

What Claude Code sees of omp's system prompt

Claude Code keeps its own claude_code system prompt; omp's prompt is appended to it, block by block:

• omp's harness block (the "You are omp's trusted coding assistant" template) is not forwarded — it describes omp's own tool loop. Only its <skills>, <generic-rules> and <domain-rules> sections are carried over. Skills and domain rules are listed with an instruction to read skill://<name> / rule://<name> through the bridged read tool, and are left out when read is not available.
• Everything else is forwarded verbatim: the <project-context> block (your AGENTS.md/CLAUDE.md context files, workstation, workspace roots), blocks other extensions add in before_agent_start, and a custom system prompt if you set one.

Summaries, /clear, and session switches

• omp's one-off summary requests — compaction, branch summaries, handoff — run in a separate, isolated Claude Code process with omp's summarization prompt, so they never enter the shared Claude Code session.
• After /clear, the next turn rebuilds the Claude Code session from what omp still has, rather than resuming the cleared conversation.
• Switching, resuming, forking or branching a session drops the shared Claude Code session, and the next turn rebuilds or resumes it from omp's history.

AskClaude Tool

Opt-in: set askClaude.enabled to true (see Configuration). Available when using any non-claude-bridge provider. omp's model can delegate tasks to Claude Code and wait for it to answer a question or perform a task. Examples of how to use:

• "Ask Claude to plan a fix"
• "If you get stuck, ask claude for help"
• "Ask claude to review the plan in @foo.md, implement it, then ask an isolated=true claude to review the implementation"
• "Find all the places in the codebase that handle auth"

Delegated calls are isolated from your ~/.claude estate: children don't read global CLAUDE.md files or Claude Code's skill listing, get no omp skills, and always get Claude Code's own system prompt preset.

Parameters

• prompt — the question or task for Claude Code
• mode — read (default, read files and search/fetch on web), none, or full (read+write+bash, disable this mode with allowFullMode: false in config)
• model — opus (default), sonnet, haiku, or a full model ID; passed to Claude Code as-is
• thinking — effort level: off, minimal, low, medium, high, xhigh
• isolated — when true, Claude gets a clean session with no conversation history (default: false)

Configuration

Config: ~/.omp/agent/claude-bridge.json (global) or .omp/claude-bridge.json in the project (merged over global).

{
  "askClaude": {
    "enabled": true,
    "allowFullMode": true,
    "defaultIsolated": false,
    "description": "Custom tool description override"
  },
  "provider": {
    "strictMcpConfig": true,
    "pathToClaudeCodeExecutable": "/home/you/.nix-profile/bin/claude"
  }
}

askClaude:

• enabled — register the AskClaude tool (default false). If it's unset, the startup notice below points this out once.
• name — override the tool's omp-side name (default "AskClaude")
• label — override the TUI label (default "Ask Claude Code")
• description — override the tool description
• defaultMode — "read" (default), "none", or "full"
• defaultIsolated — start each call in a fresh session (default false)
• allowFullMode — allow mode: "full"; set false to lock it out

provider:

• strictMcpConfig — block MCP servers from ~/.claude.json / .mcp.json (default true). Cloud MCP (Gmail/Drive via claude.ai OAuth) is always blocked.
• autoMemoryEnabled — enable Claude Code's auto-memory system (default false)
• pathToClaudeCodeExecutable — path to the claude binary. Useful if your OS/filesystem has the SDK's bundled musl/glibc binaries in a place where they can't run (e.g. Nix: "/home/you/.nix-profile/bin/claude").

Startup notice: the first interactive session to reach Claude Code points out that askClaude.enabled is unset, then records startupNoticeShown (the date, YYYY-MM-DD) in the global config so it doesn't nag again.

Tests

• npm run test:unit — offline tests (tests/unit-*.mjs).
• node tests/omp-smoke.mjs — loads the extension in your installed omp and checks the registered models. Makes no model call.
• npm test — upstream's integration suite (tests/int-*.{sh,mjs}). It drives pi, not omp, and makes real Claude Code calls, so it does not validate this port; it is kept for upstream merges.

Debugging

Set CLAUDE_BRIDGE_DEBUG=1 to enable debug output:

• Bridge log at ~/.omp/agent/claude-bridge.log — every provider call (with the system-prompt blocks received and the size of what was appended), session sync decision, tool result delivery, and CC's stderr. Override location with CLAUDE_BRIDGE_DEBUG_PATH.
• Per-query Claude Code CLI logs at ~/.omp/agent/cc-cli-logs/<timestamp>-<tag>-<seq>.log — the CC subprocess's own debug stream, one file per query() call.

Known issues

Sessions get rebuilt more often than they need to be, and a rebuild is expensive. The bridge rewrites Claude Code's session from omp's history whenever the history moves underneath it — after an abort, compaction, tree navigation, /clear, or an API error. A rebuild usually loses the prompt cache.

Files Claude Code edits are not carried across a rebuild. The edit itself survives in the history as a tool call and its result; Claude loses only the post-edit file snapshot. @file expansions are carried.

Exported Anthropic environment variables override the Claude Code child. An ANTHROPIC_BASE_URL, ANTHROPIC_API_KEY or ANTHROPIC_AUTH_TOKEN exported for another gateway redirects Claude Code as well, and every turn fails with that gateway's auth error. Unset them for the omp process.

Release

Published from the omp branch by the Publish omp-claude-bridge GitHub workflow (manual dispatch). Versions are CalVer: the release date YYYY.M.D, then YYYY.M.D-post.N for further releases the same day. The repo's package.json stays upstream's; publish/stage.mjs writes the published manifest from publish/package.template.json, runs the typecheck and unit suite, packs the tarball, installs it, and loads it in omp. Run it locally with:

node publish/stage.mjs --version 2026.10.17

The workflow publishes with npm provenance (trusted publishing), or with an NPM_TOKEN secret if one is set.

Credits

Eli Dickinson's pi-claude-bridge, itself based on claude-agent-sdk-pi by Prateek Sunal. MIT licensed.