OMP features and highlights

Companion overview (who, install, stack, Pi/Paseo, release status): 2026-10-05 OMP agent TUI. This note is the feature surface as of the README on main and the docs at omp.sh, checked 2026-10-05. Latest GitHub tag is v18.6.1 (published 2026-10-05 00:12 CST). There is no separate “features” page on omp.sh; the homepage is the marketing cut of the README, and the real inventory is docs/tools.

Surfaces: TUI, one-shot, SDK, RPC, ACP

Same engine, five wrappers. The README groups the last four as “four entry points” and then documents the SDK beside them (README, SDK).

Surface	How you enter	Why it exists
Interactive TUI	`omp`	Default. Tool calls render as cards, edits preview before they land, ambiguity goes through `ask` (a structured option picker). The same cards also go out over ACP.
One-shot	`omp -p`	Answer one prompt and exit. Pair with `--no-session` when a pipeline must not write a transcript.
SDK	`bun add @oh-my-pi/pi-coding-agent`	In-process embed. `createAgentSession()` discovers credentials, settings, extensions, skills, rules, MCP, and LSP unless you override them. Hosts can add `customTools`, subscribe to stream events (`text_delta`, tool calls, compaction, retry), and `steer` / `followUp` / `abort`.
RPC	`omp --mode rpc`	NDJSON over stdio for a non-Bun host or process isolation. `--mode rpc-ui` adds tool cards, selectors, and dialogs as `extension_ui_request` frames the host must answer.
ACP	`omp acp`	Agent Client Protocol JSON-RPC, documented for Zed. `bash` maps to `terminal/create`, `read`/`write` to `fs/*`, and destructive tools pause on `session/request_permission`.

SDK correction. The README heading still says “embed in Node.” The SDK page is explicit: Bun ≥ 1.3.14, and “it is not a Node.js SDK.” The overview digest called it a Node SDK; that wording matches the README heading, not the current docs page. Do not put an RPC child behind the SDK just to call the same process twice.

Editing model: hashline, AST, conflicts

Hashline (edit). The model points at short content-hash anchors instead of retyping the lines it wants to change. Stale anchors are rejected before the file is written. The README’s own numbers, which are marketing until you re-run the bench: Grok Code Fast 1 edit success 6.7% → 68.3% once the format stopped eating the model; Gemini 3 Flash +5 pp over str_replace; Grok 4 Fast −61% output tokens; MiniMax 2.1× pass rate. Implementation lives in @oh-my-pi/hashline. v18.6.1 (#14254) accepts a follow-up edit after a line shift when the anchor still carries the current content, and still rejects a genuinely stale line.

ast_edit / ast_grep. Structural rewrites via ast-grep over 50+ tree-sitter grammars. ast_edit returns a proposed card; the disk write happens only after the agent writes a one-line reason to xd://resolve and the TUI shows Accept (atomic, all or nothing). Docs: ast_edit is on by default (astEdit.enabled); ast_grep search is off by default (astGrep.enabled). The README lists both in the tool catalog without that split.

write creates or overwrites a file, archive entry, or SQLite row. conflict://N turns each merge conflict into a URL; the agent writes @theirs, @ours, or @base (bulk: conflict://*).

Code intelligence: LSP and DAP

LSP is on by default (lsp.enabled; --no-lsp for one session). The README’s badge is 14 lsp ops. Documented uses: diagnostics, definitions, references, symbols, hover, code actions, raw requests, and rename through workspace/willRenameFiles so re-exports and aliased imports update before a file moves. Inspection is classed Read; fixes and renames are Write. The server still has to be installed or configured. People who left an IDE for the terminal cite this as the reason OMP beats a grep-only agent (see Field).

DAP (debug) is on by default (debug.enabled) but only appears when an adapter is actually available. README badge: 28 dap ops. The tool launches or attaches, sets breakpoints, steps, and reads threads, stack, scopes, and variables. README’s worked examples: lldb-dap on a native segfault, dlv on a hung Go service, debugpy on a wedged Python process. Docs also name gdb and js-debug-adapter. Inspection is Read; launch and process control are Exec.

Subagents, advisor, and the other multi-agent modes

task. Fans independent work to child agents, optionally in isolated worktrees (APFS clones, btrfs/zfs reflinks, overlayfs, projfs, rcopy — crate pi-iso). Workers can DM each other on an in-process IRC bus instead of routing every message through the parent. The parent reads a schema-validated object, not prose. Default recursion depth is 2 (task.maxRecursionDepth). v18.6.1 made subagent follow-ups background jobs you can wait or cancel, delivered once.

Agent Hub (Alt+A). Roster of live subagents: transcript, steering message, revive a parked worker, kill a stuck one without aborting the parent.

hub. Message peer agents, follow or cancel background jobs, supervise shared long-running processes (dev servers, debuggers). Available when collaboration is possible at the current task depth.

Advisor. A second model on the advisor role, with its own context, reads each turn and injects a quiet aside, a concern, or a hard blocker. The main agent course-corrects or explains why it won’t. This is the feature people name unprompted (see Field).

Other director modes (tools index):

/review — reviewer subagents over a branch, commit, uncommitted work, or PR; findings ranked P0–P3 with a ship/no-ship verdict.
/vibe — the main session becomes a read-only director over persistent fast / good workers.
/plan — planning turn on the plan role, sandboxed read-only, then execute-and-purge, keep the transcript, or compact.
/goal and /guided-goal — one durable, budgeted objective across turns.
/autoresearch — repeated benchmark-backed experiments on an autoresearch/* branch. Some reset controls are destructive.
Magic words in prose only (not code spans): ultrathink, orchestrate, workflowz.

MCP inheritance

OMP is an MCP client, not a separate MCP product. On first run it reads rules, skills, and MCP servers already on disk. The README names .claude, .cursor, .windsurf, .gemini, .codex, .cline, .github/copilot, and .vscode. The MCP docs are slightly tighter for servers: Claude Code, Codex, Gemini CLI, OpenCode, Cursor, Windsurf, VS Code, installed plugins, and extension packages. /mcp list shows which file supplied each server. First definition wins; duplicates are not merged. /mcp disable hides an imported server without editing the other tool’s config. OAuth writes only omp-managed files.

Own config: .omp/mcp.json (project), ~/.omp/agent/mcp.json (user), or ~/.omp/profiles/<name>/agent/mcp.json. Transports: stdio (default), http (Streamable HTTP), sse (legacy; a URL ending in /sse does not select it by itself). Tool names surface as mcp__<server>_<tool>. Prompts become /server:prompt. Resource-update injection is off unless mcp.notifications is on.

A stdio server runs as your user. The docs tell you to review project MCP files before opening an unfamiliar repo, and to use a separate profile for untrusted checkouts because OAuth is bound to profile + URL.

Models and providers

Nine roles: default, smol, slow, plan, commit, vision, task, advisor, tiny. Swap mid-session with /model; cycle the active role’s models with Ctrl+P. Launch overrides: --smol, --slow, --plan.

README claim: 60+ providers, a thousand models. Auth is API key, OAuth (/login), or a coding-plan subscription. Custom OpenAI-compatible (and several other wire APIs) go in ~/.omp/agent/models.yml. Routing knobs that are actually specified, not slogans:

Fallback chains under retry.fallbackChains — next entry takes the rest of the turn on 429/quota, restored on cooldown.
Path-scoped enabledModels / disabledProviders so one repo can pin a different set.
Round-robin credentials with session affinity and per-key backoff.

Local: Ollama, LM Studio, llama.cpp, vLLM, LiteLLM. Full list: omp.sh/docs/providers.

web_search is one tool over a 23-provider chain (auto, or pin one). Site-aware extractors cover GitHub/GitLab, package registries, arXiv, Stack Overflow, Reddit, HN, MDN, docs.rs. Exa accepts /login exa or a public MCP fallback. Vuln lookups hit NVD, OSV, and CISA KEV.

Sessions

Sessions are append-only JSONL trees under ~/.omp/agent/sessions/, one history per working directory, ids are UUID v7 (sessions).

omp -c continues the most recent session in this cwd, preferring a per-terminal breadcrumb so tmux panes don’t steal each other.
omp -r / /resume picks; a short id prefix is enough.
/branch starts a new leaf in the same file. /fork (or omp --fork) clones into a new file with a parentSession marker.
/tree moves the leaf without forking. Shift+L labels survive compaction.
/compact summarizes the older half of the active branch; the file on disk still holds the pre-compact history.
/fresh resets provider stream state (stale prompt cache, wedged stream) without wiping the local transcript.
/handoff writes a structured wrap-up. /share renders HTML and uploads a secret gist via gh, or calls ~/.omp/agent/share.ts. /export is the HTML without sharing.
--no-session is ephemeral and disables /fork, /export, and /share.

/collab. Live session on a relay, link plus QR. omp join or a browser guest. /collab view is read-only. Frames are sealed client-side. v18.6.1 lets guests answer setting-change and tool-issue prompts.

Standout UX, beyond the tool list

In-process shell. bash is brush plus a builtins crate (README: 58 utilities in one place, 46 “coreutils” in the tool blurb, 67 in the crate table — the counts don’t match). No WSL required; the same binary is the Windows port. Persistent shell sessions, optional PTY.
eval. Persistent Python and JavaScript by default; Ruby and Julia are opt-in. Kernels can call back into agent tools.
Stream rules. A regex match aborts the stream mid-token, injects a dormant rule, and retries. Injections survive compaction, so you don’t pay the rule on every turn.
Internal URLs through the same read/grep you use on files: pr://, issue://, agent://, skill://, ssh://, conflict://, xd://, plus artifact:// and memory:// in the docs conventions. agent://…/findings.0.path pulls a field out of a subagent result.
ask. Option picker mid-turn, also over ACP.
omp commit. Splits the working tree into dependency-ordered atomic commits; cycles rejected before anything is written; lockfiles excluded from analysis. Push is opt-in.
Browser. Headless Chromium, CDP attach, or a relay extension that adopts Chrome tabs you already have open. README also claims the same API can drive Electron apps (Slack called out). Stealth on by default.
computer. Real desktop: windows, screenshots, native input, accessibility tree, clipboard. Off by default (computer.enabled). Not the same as browser.
Git diff pane. Space stages or unstages the focused hunk (v18.6.1).
/btw in Tern. Side questions scroll instead of clipping (v18.6.1, #14331).
Memory. memory.backend is off by default. Tools page: recall / retain / reflect need hindsight or mnemopi; memory_edit is Mnemopi only; learn also accepts local but needs autolearn.enabled. README’s “local, Hindsight, or Mnemopi” is the backend enum, not “every memory tool is on.”
Gated off by default (README): github, security_scan, generate_image, tts, checkpoint, rewind, and the memory tools. inspect_image appears automatically when the active model can’t see.
tools.xdev (on by default) keeps rare tools behind xd:// so their descriptions aren’t in the prompt up front. read xd:// lists them.
Approvals. Persistent mode defaults to yolo (Read, Write, and Exec all auto-allowed). always-ask, write, or per-capability allow/prompt/deny exist. An approval prompt is not a sandbox.
Extensions. TypeScript modules with the same tool, slash, and hotkey APIs. /reload-plugins hot-reloads. Plugins can ship to a marketplace or npm.
Voice crate (pi-voice: Opus, WebRTC) and v18.5’s GPT live-voice RPC events are in the overview’s release notes; this pass did not re-read the voice docs, so treat “live voice” as release-note, not a verified user workflow.

What people highlight

X search 2026-10-05, English, not from @grok, roughly July–September 2026. The repeated nouns are advisor, LSP, hashline, and subagents. MCP inheritance is a docs feature; it did not show up as something people brag about. Posts that mention MCP are usually “OMP is one of N clients,” not “it imported my .cursor config.”

brandon galang (2026-07-03): favorite addition is the advisor, “another model riding shotgun.”
Jared Smith (2026-07-03): Anthropic and ChatGPT subscriptions, configurable task subagents, advisor “like my eyes.”
Gokay Yilmaz (2026-08-17): “built in LSP is the hack,” annoyed it was missing from a harness tier list.
NOVA (2026-07-10): LSP ops so “everything your IDE knows” is in the session.
Chris Short and EveryDev.ai (2026-07-27): the short list they repeat is hash-anchored edits, LSP, DAP, subagents, browser.
brrrock (2026-07-14): didn’t import Claude Code config; set a custom PowerShell LSP and fallbacks per model role.
aaronjmars (2026-09-17): ranks omp first for “deploy today,” citing omp -p JSONL (tokens, USD, tool events) and a profile-isolated credential vault.
Kevin Gray (2026-09-17): recommends OMP over Pi specifically for subagents.

Caveats / what it is not

Not Pi. Pi’s pitch is four tools and a small prompt. OMP’s pitch is the opposite. The context-weight complaint (a large starting prompt, most tools unused) is argued in the overview digest’s July 2026 HN thread; this note did not re-fetch that thread.
Not a sandbox. Default approval mode is yolo. MCP stdio servers, project extensions, and computer run as you. Isolation worktrees are for sibling agents not stepping on each other’s files, not a security boundary.
Not “31 tools always in the prompt.” Several are off until you flip a setting (github, memory, computer, ast_grep, image, tts, security scan). tools.xdev hides more behind xd://. /tools is the live list.
Not a Node library. Bun SDK, or RPC/ACP if the host isn’t Bun.
Not a quiet Windows port, and not “needs WSL.” Native Windows is a stated goal; 18.5–18.6 still shipped a pile of Windows fixes (see the overview).
Not done with compaction or hashline edge cases. v18.6.1 is still patching screenshot token estimates and shift-consistent anchors.
Counts drift inside the README (coreutils 46 vs 58 vs 67). Cite the docs page for whether a tool is on, and the release notes for what last week’s patch actually changed.
omp² is a homepage tease (“soon”), not a feature in this tree.
security_scan “drives Codex Security cloud scans” is a README phrase for an off-by-default tool. This note did not verify that cloud path.

Sources

README, main, fetched 2026-10-05: can1357/oh-my-pi
Docs: overview · tools · sessions · sdk · mcp · providers
Homepage (no separate features page): omp.sh
Release: v18.6.1 (2026-10-04 16:12 UTC / 2026-10-05 00:12 CST)
X posts linked inline. Searches: feature keywords and MCP-inherit keywords, lang:en, excluding @grok, 2026-10-05.
Prior overview, not rewritten: 2026-10-05 OMP agent TUI