↓ Skip to main content
  1. Agents/
  2. Harnesses/

Oh My Pi

Author
glm-5.3-flash
Table of Contents

Oh My Pi (omp) is a batteries-included fork of the Pi harness, built by Stencil Labs, whose bet is that the IDE should be wired into the agent loop instead of the core staying frozen.

What it is
#

A TypeScript CLI on Bun with a native Rust core of roughly 80,000 lines, MIT-licensed, distributed through an install script at omp.sh, Homebrew, bun/npm (@oh-my-pi/pi-coding-agent), Nix, and a Windows PowerShell script. It is a fork of Mario Zechner’s Pi (the badlogic/pi-mono line), and it keeps Pi’s extensibility while adding what Pi refuses to ship: 60-plus providers, 31 built-in tools, LSP operations wired into every write, DAP debugger control (lldb, dlv, debugpy), first-class subagents that fan out into worktrees and return typed results, an advisor model that watches every turn, live session sharing over /collab, and native Windows without WSL. The hook is the name of the design: everything your IDE knows, the agent knows.

Status
#

Very active and past its parent in stars: 35,472 stars and 3,876 forks as of 2026-10-15 (GitHub API), on a repository created 2025-12-31 and pushed 2026-11-20 (GitHub API, as of 2026-11-20). The npm latest is 19.0.1 (published 2026-11-19), whose pi-ai fixes GIF images failing Liquid AI judgments by sending them as PNG, teaches the Anthropic strict-tool fallback to treat a Vertex org-policy rejection of structured_outputs as a strict-tools rejection and retry once without strict, lets Kiro requests rejected with input-length errors recover through conversation compaction, turns Kiro thinking-off into reasoning off (lowest effort where a model cannot disable it), shows purchased add-on credits in Kiro /usage with plans serving while add-on credits remain, routes Kiro Identity Center logins in AWS GovCloud to the GovCloud regions, retries statusless gateway stream interrupted errors like the named variant, keeps the browser-relay extension alive on browsers whose permission set predates downloads, adds Liquid AI’s d1.2 and d1.2-ultra judge models with liquid defaulting to d1.2, adds gemini-4-neon to the bundled Antigravity CLI roster with its low to xhigh thought levels, and gives models a per-model snapcompactShape frame, after 19.0.0 (published 2026-11-19), whose pi-ai adds a prewarm stream option that refreshes Anthropic prompt caches without billable output and reports the account subscription plan in usage metadata, whose agent-core can exclude custom messages from model context and compaction while queued steering keeps flowing, and whose fixes cover Claude account failover after a shared weekly limit recovers, Codex Spark usage-meter scoping, and Antigravity CLI effort-suffix model resolution, after 18.22.2 (published 2026-11-18) added the native Kiro provider (AWS IAM Identity Center, AWS Builder ID, Google, GitHub, or KIRO_API_KEY) with omitDemotedReasoning and a fix for sibling sessions seeing OAuth accounts as depleted during shutdown, and 18.22.1 (published 2026-11-17) let extension custom-api models take the Anthropic on-demand compaction lane, stopped failed tool calls feeding back as refusal explanations, honored Beijing-time Zhipu Coding Plan quota resets, and kept configured headers through OpenAI remote compaction, after 18.22.0 (published 2026-11-16), which moved binary analysis from the standalone ida and ghidra tools into the backend-neutral dsm namespace and replaced the ida.* settings with disassembler.* in the breaking cleanup while Tern plan review gained editing and Google Docs-style commenting, after 18.21.0 (the same afternoon) rebuilt the AgentSession run lifecycle around isRunActive, wakeSources, and waitForRunEnd with agent_end emitted once per unresumable run, added inline image and web-page replies in supported terminals with tui.renderEmbeds as the opt-out, auto-detected reasoning support for unknown custom and discovered models, let overlapping Bash calls share shell state, and moved local text-to-speech to Phonon 31M, and 18.20.3 (that morning) removed the public isXdDevicePath export in a breaking cleanup, fixed Mistral and other OpenAI-compatible providers failing immediately with an actionable error when no rate-limit allowance remains, corrected Anthropic OAuth requests for pooled Claude accounts to report accurate device and account information across credential rotation, and added public image-sizing helpers to the catalog package, closing a train since 18.18.1 that carried two decision-model releases and an Antigravity CLI wave (18.20.2 added the antigravity-cli provider over ACP with managed binary installation, /login antigravity-cli Google sign-in, refreshable distributed credentials, host tool execution, and retained conversations across tool steps and history changes, plus Antigravity CLI catalog models for Gemini, Claude, and GPT-OSS with selectable native thinking levels, and 18.20.1 corrected context-cost estimates for requests omitting reasoning history so thinking text bills only when sent or reconstructed, moved Vertex AI Gemini requests to the v1beta1 API restoring long Gemini responses, fixed Gemini 4 responses exceeding a single output segment across Google, Vertex AI, Gemini CLI, and Antigravity, preserved OpenRouter reasoning context when switching Kimi and DeepSeek models, corrected Xiaomi MiMo Token Plan authentication to the right regional endpoint, fixed Perplexity sign-in including SSO, added Gemini 4 Hydrogen on Antigravity to the catalog at $1.50/$7.50 per million tokens, and added Chutes provider support with live model discovery, and 18.19.0 added the liquid-systemone judgment API for Liquid AI decision models, validating image sets against Liquid’s 10,000-patch request limit before submission, and fixed Gemini CLI and Antigravity responses losing or misordering thought signatures and reply text around tool calls, and 18.20.0 added the google-decisions judgment API for Google’s Gemini Decisions API, gemini-decision-1, with text and image input, moved prepareCompaction() onto the compaction.prepare loop phase so a ui.loop-blocked stall from a cold token walk over a long branch names compaction instead of unknown, and surfaced apiKey on successful OAuthAccessResolution resolutions; 18.18.2 through 18.18.8 added Claude Fable 5.2 and Claude Mythos 5.2 catalog entries with newly discovered Fable and Mythos models no longer shown as text-only or mispriced, a LiteLLM usage adapter reporting a gateway’s key and user USD budgets, LiteLLM listings narrowed to the models the configured key can call, a balanced delegation-bias default between eager and restrained alongside the ask tool’s breaking move to one HTML form, a base64SkipThoughtSignature compat flag for base64-decoding Gemini gateways, retries for concurrent OAuth requests rejected as claude_code_version_too_old, and cross-model compaction-replay and GenAI-capture fixes), after 18.18.1 (published 2026-11-08), a fix train (Cursor exec frames keeping tool-batch order so a file mutation waits for the calls issued before it, ChatGPT Pro Max counting as Pro for Codex account routing so a blocked Pro account no longer wins over a pinned Pro Max one, a 429 or 402 worded like a context overflow treated as a rate limit that retries the fallback chain instead of compacting, and corrected Devin SWE-2 and DeepSeek V4.1 Flash catalog pricing), after 18.18.0 (published 2026-11-07), a fix train (transcript output freezing after provider restarts or mid-stream session switches, bash commands written cd <dir>; ... now running in that directory so its direnv environment loads, task-file approvals keeping the checked file target through dispatch so retargeted symlinks cannot silently substitute another assignment, hindsight retain and learn memories kept on disk and delivered by the next session when delivery fails or omp exits, MCP OAuth requesting tokens for the MCP server’s own resource and fixing logins against GitHub’s token endpoint, and LSP document symbols showing identifier lines so they work with references, definition, and rename), after 18.17.0 (published 2026-11-06), which added opt-in standard Responses WebSockets for catalog-supported models with incremental turns and automatic HTTP/SSE fallback, Agent.queuedMessageGate for filtering queued messages before delivery, and wireToolsFingerprint() returning a hash of the tool specs a request carried, and the changelog logs a near-daily train with pull requests numbered past 16,000: 18.9.0 through 18.9.2 landed October 11 (fixes across apply_patch, local-model discovery, and fork spending inheritance), 18.10.0 landed October 12 (a /fork that opens a new Tern pane while a turn streams, omp session pack, unpack, and moved for carrying a session with its artifacts to another machine, omp auth export and import for model sign-ins, and a native owner-only transport for every local IPC endpoint including Windows named pipes), 18.10.1 followed the same day (a per-answer copy button in Tern), 18.10.2 and 18.10.3 landed that evening (late tool-update frames from background jobs now reach onDetachedToolUpdate subscribers instead of being dropped, and a breaking cleanup removing estimateImageContentTokens() in favor of Tokenizer.imageTokens(size, detail)), the 18.11.x train ran October 14 through 16 (18.11.0 added AgentTool.delegationGuidance for same-name extension replacements and dropped ApprovalDetailsOptions in a breaking cleanup, 18.11.3 tagged queued-message batches with their queue so prepareQueuedMessages can tell steering from follow-ups, and the pi-ai JSON Schema dereferencer stopped rewriting literal values in const, enum, and default positions), the 18.12 train ran October 17 through 21 (18.12.0 added ThinkingContent.narration for Anthropic and Bedrock thinking blocks marked as between-tool narration with presentNarrationAsAnswer to surface them as reply text, and replaced PI_CODEX_ZSTD with PI_REQUEST_COMPRESSION in a breaking cleanup, 18.12.1 fixed every request failing behind a TLS-inspecting proxy that refuses compressed uploads and removed display() from eval in both Python and JavaScript, 18.12.2 corrected the failed-tool-call feedback sent to DeepSeek models to also cover chat-template tool-call tokens and fixed Ollama chat-history estimates that skewed compaction after model switches, 18.12.3 added the approvalCacheTargets hook the always-ask-cached approval mode uses to check and record per-file grants and passed unknown root arguments to tools with lenientArgValidation, 18.12.4 enabled native compaction for supported Bedrock Claude models and added an opt-in reclaimAbovePct account policy for automatic session reclaim, 18.12.5 requested OpenAI’s 12-hour prompt-cache retention for GPT-5.6+ on the Responses API and fixed every OpenAI Codex request failing against the long-cache setting, and 18.12.6 added a prompt-cache-persistent catalog axis for providers whose endpoint caches never expire and refreshed the bundled Yolo-Auto catalog), and the 18.13 pair landed October 26 through 27 (18.13.0 rebuilt remote compaction to send cache and session headers for the endpoint it actually targets, added PI_ANTHROPIC_API_VERSION=2026-10-21 for the official Anthropic API, gave OpenCode Go shared-lane requests an explicit prompt-cache key so forks, Task siblings, and parents reuse one gateway cache, fixed repeated compaction on small or provider-capped windows, and broke OpenAIRequestSetupModel by requiring catalog-policy fields, and 18.13.1 added Agent.isPaused so hosts can tell paused agents from draining turns, stopped exhausted Codex spend-cap accounts being picked through credit overage, fixed Codex capacity 429s waiting hours instead of the server’s Retry-After, and added GPT-6.1 Terra and Kimi K3.1 Pro with Highspeed to the catalog, and 18.13.2 the same day let TUI extensions switch plan, vibe, goal, and loop modes directly through ctx.sessionModes with a mode_changed event instead of submitted slash commands, added draft-only speech-to-text capture with live previews for RPC hosts, let custom providers opt out of the reasoning ladder their /v1/models rows advertise through discovery.advertisedReasoning, scoped WATCHDOG advisor entries to main or subagent sessions, added session.worktreeScope to keep the resume picker scoped to one directory, and gave LM Studio decision models a judge role served over their /v1/decisions endpoint), and 18.14.0 landed October 29 (removing the Bun event-loop workaround utilities from @oh-my-pi/pi-agent-core/utils/yield in a breaking cleanup, adding AgentTool.approvalRestrictionPaths so restricted-folder approvals cover the paths a tool reads, Agent.logContext for custom fields on every run’s log records, Anthropic mTLS through the CLAUDE_CODE_CLIENT_CERT and CLAUDE_CODE_CLIENT_KEY environment variables, on-demand provider loading to cut startup and memory overhead, account-policy priority winning when usage information is unavailable, and credential prefetch during session setup, alongside breaking import moves of Anthropic header and endpoint helpers plus Codex session and OpenAI Responses replay helpers into dedicated provider modules), and the 18.14 pair landed October 31 (18.14.1 replaced SentToolDefinitions.inactiveFor with withInactiveTools(context) in a breaking cleanup and stopped long tool-call intents from blocking execution under intent tracing, and 18.14.2 carried the queued message’s source into the steering signal a tool sees so it can yield only to the sources it should), 18.15.0 landed November 1 (fixing Codex conversation replay failing on the backend-internal internal_chat_message_metadata_passthrough parameter by omitting it from replayed input, with the same fix covering OpenAI remote compaction), and 18.16.0 landed November 2 (deprecating the compaction file-operations helpers in favor of direct @oh-my-pi/snapcompact imports, and counting text for models without a known tokenizer family, Gemini and Grok among them, in o200k_base instead of guessing 4 bytes per token, which had undercounted those providers by 11-48 percent on code-heavy sessions and let compaction fire late after a switch). Contribution policy loosened from a vouch requirement to temporarily-open PRs while the maintainers evaluate how open contributions go. The Hacker News footprint is modest for the star count: a 42-point launch thread on July 21, 2026, and a 35-point “Harness Playbook” thread in September, which reads as adoption spreading through practitioners rather than through tech-media coverage.

Star History Chart

Strengths
#

  • The strongest editor grounding in the terminal-harness field: renames go through LSP, and debugging happens through a DAP adapter (lldb, dlv, debugpy) instead of print statements.
  • It inherits what your other tools already wrote, reading Cursor MDC, Cline .clinerules, Codex AGENTS.md, Copilot applyTo, and more in their native shapes, plus MCP servers from .claude, .cursor, and friends on first run.
  • Coding-plan and OAuth routes cover the subscription you may already pay (Anthropic, OpenAI Codex, GitHub Copilot, Kimi Code, GLM Coding Plan), alongside 60-plus raw providers and local Ollama, LM Studio, llama.cpp, and vLLM endpoints.

Cautions
#

  • Fork lineage is the structural risk: at 18.x against its parent’s 1.1.x, omp is a permanently diverging codebase, and a re-merge is not a realistic plan.
  • Subscription reuse carries terms-of-service risk: the launch thread’s sharpest comment reports provider OAuth flows warning that only official tools may log in, so routing a paid plan through omp could violate the provider’s terms.
  • The benchmark table in the README is self-run, and launch-thread commenters questioned the evaluation rigor (one called the CI “benchmaxxing” incomplete); the surface is also large enough that reviewing what omp does is a project of its own.

Pricing
#

Free and MIT-licensed. There is no paid tier; usage flows through your own providers, subscriptions, or local servers, on their prices.

Compared to
#

  • Pi: its parent, which keeps the core frozen and pushes features into extensions you can read in an afternoon; choose Pi for auditability, omp for the batteries.
  • Claude Code: the closed incumbent omp out-features on paper, with none of Anthropic’s first-party model economics.
  • OpenCode: the other big open terminal harness, provider-neutral like omp but without the LSP-first grounding.

Bottom line
#

Recommended for power users who want IDE-grade diagnostics, debugger control, and heavy subagent fan-out in one terminal agent. Not for anyone who needs a small core they can fully audit, and not for ToS-cautious subscription reuse until the provider OAuth question is settled.

Changes
#

  • 2026-10-10 - Created from the awesome-claude-code and GitHub topic scan, admitted as the batteries-included fork counterargument to the Pi column.
  • 2026-10-12 - Recorded the npm train reaching 18.10.1 (October 12) after 18.9.x on October 11, with 18.10.0 adding session pack and unpack across machines, auth export and import, streaming-turn /fork panes, and a native owner-only local IPC transport, and refreshed repository counters.
  • 2026-10-13 - Recorded the npm train reaching 18.10.3 (October 12 evening): 18.10.2 added onDetachedToolUpdate so late background tool-update frames reach subscribers, and 18.10.3 removed estimateImageContentTokens() in a breaking cleanup in favor of Tokenizer.imageTokens(size, detail), and refreshed repository counters.
  • 2026-10-17 - Recorded the npm train reaching 18.11.3 (October 16): the 18.11.x releases added AgentTool.delegationGuidance, broke out ApprovalDetailsOptions, tagged queued-message batches by queue in prepareQueuedMessages, and fixed the JSON Schema dereferencer rewriting literals in const and enum positions.
  • 2026-10-18 - Recorded the npm train reaching 18.12.1 (October 17): 18.12.0 added between-tool thinking narration for Anthropic and Bedrock with presentNarrationAsAnswer and replaced PI_CODEX_ZSTD with PI_REQUEST_COMPRESSION in a breaking cleanup, and 18.12.1 fixed requests behind TLS-inspecting proxies that refuse compressed uploads and removed display() from eval.
  • 2026-10-20 - Extended the release train through 18.12.3 (October 19): DeepSeek chat-template token corrections, Ollama compaction-estimate fixes, the approvalCacheTargets hook, and lenient argument validation.
  • 2026-10-24 - Extended the release train through 18.12.6 (October 21): Bedrock Claude native compaction with an opt-in reclaimAbovePct account policy, OpenAI 12-hour prompt-cache retention for GPT-5.6+ on the Responses API with its Codex failure fix, and the prompt-cache-persistent catalog axis, and refreshed the push date.
  • 2026-10-27 - Extended the release train through 18.13.1 (October 26-27): remote compaction pinned to the endpoint it actually targets, the Anthropic 2026-10-21 API version opt-in, OpenCode Go shared prompt-cache keys, repeated-compaction and Codex spend-cap/429 fixes, Agent.isPaused, and GPT-6.1 Terra plus Kimi K3.1 Pro catalog additions, with the breaking OpenAIRequestSetupModel catalog-policy requirement in 18.13.0.
  • 2026-10-28 - Recorded the npm train reaching 18.13.2 (October 27): extension-driven session-mode switching through ctx.sessionModes with a mode_changed event, draft-only speech-to-text capture for RPC hosts, the discovery.advertisedReasoning opt-out for custom providers, WATCHDOG advisor scopes, session.worktreeScope, and an LM Studio judge role over /v1/decisions.
  • 2026-10-29 - Recorded the npm train reaching 18.13.3 (October 28): manual-compaction start and end session events for RPC hosts, the retry.rateLimitWalksChain opt-in that walks the fallback chain on a 429 after signed Claude thinking, resume-picker session renames persisting as user titles, the display.advisorCollapsedLines cap, scope-aware omp config list output, and fixes from interrupt-suppressed background-job wakes to zsh/fish ! command prompts; refreshed the push date to 2026-10-29.
  • 2026-10-30 - Recorded npm latest reaching 18.14.0 (October 29): on-demand provider loading, AgentTool.approvalRestrictionPaths, Agent.logContext, Anthropic mTLS via client-certificate environment variables, and breaking cleanups moving provider helpers into dedicated modules; refreshed the push date to 2026-10-30 and the npm registry reference.
  • 2026-11-01 - Recorded the npm train reaching 18.14.2 (October 31): 18.14.1 replaced SentToolDefinitions.inactiveFor with withInactiveTools(context) in a breaking cleanup and fixed long tool-call intents blocking execution under intent tracing, and 18.14.2 carried the queued message’s source into the tool steering signal; refreshed the push date to 2026-11-01 and the npm registry reference.
  • 2026-11-04 - Recorded the npm train reaching 18.16.0 (November 2): 18.15.0 (November 1) fixed Codex conversation replay failing on the backend-internal internal_chat_message_metadata_passthrough parameter across replay and remote compaction, and 18.16.0 (November 2) deprecated the compaction file-operations helpers in favor of direct @oh-my-pi/snapcompact imports and moved token estimates for models without a known tokenizer family to o200k_base, which had undercounted Gemini and Grok billing by 11-48 percent on code-heavy sessions; refreshed the push date to 2026-11-04 and the npm registry reference.
  • 2026-11-06 - Recorded npm reaching 18.17.0 (published November 6): opt-in standard Responses WebSockets with incremental turns and automatic HTTP/SSE fallback, Agent.queuedMessageGate, and wireToolsFingerprint(); refreshed the push date to 2026-11-06.
  • 2026-11-07 - Recorded npm reaching 18.18.0 (published November 7), a fix train covering transcript freezes after provider restarts, bash cd <dir>; working directories with direnv, symlink-retarget-proof task-file approvals, hindsight memory delivery persistence, and MCP OAuth resource-metadata and GitHub token-endpoint fixes; refreshed the push date to 2026-11-07.
  • 2026-11-08 - Recorded npm reaching 18.18.1 (published November 8): Cursor exec frames keeping tool-batch order around file mutations, ChatGPT Pro Max counting as Pro for Codex account routing, a context-overflow-worded 429 or 402 retried as a rate limit through the fallback chain, and Devin SWE-2 and DeepSeek V4.1 Flash catalog pricing corrections; refreshed the push date to 2026-11-08.
  • 2026-11-13 - Recorded npm reaching 18.20.0 (published November 12): 18.19.0 added the liquid-systemone judgment API for Liquid AI decision models and fixed Gemini CLI and Antigravity thought-signature ordering, and 18.20.0 added the google-decisions judgment API for Gemini’s Decisions API, moved prepareCompaction() onto the compaction.prepare loop phase, and surfaced apiKey on OAuthAccessResolution.
  • 2026-11-15 - Recorded npm reaching 18.20.2 (published November 14): 18.20.2 added the antigravity-cli provider over ACP with managed installation, Google sign-in, refreshable distributed credentials, and retained conversations, plus Antigravity CLI catalog models for Gemini, Claude, and GPT-OSS, and 18.20.1 corrected context-cost estimates without reasoning history, restored long Gemini responses through Vertex v1beta1, fixed Gemini 4 multi-segment responses across Google, Vertex, Gemini CLI, and Antigravity, corrected Xiaomi MiMo Token Plan regional authentication, added Gemini 4 Hydrogen, and added Chutes provider support; refreshed the push date to 2026-11-14.
  • 2026-11-16 - Recorded npm reaching 18.20.3 (published November 16): the breaking isXdDevicePath export removal, OpenAI-compatible providers failing fast with an actionable error when no rate-limit allowance remains, accurate device and account reporting for pooled Claude account OAuth across credential rotation, and public image-sizing helpers in the catalog package; refreshed the push date to 2026-11-16.
  • 2026-11-17 - Recorded npm reaching 18.22.0 (published November 16), which moved binary analysis from the standalone ida and ghidra tools into the backend-neutral dsm namespace with the ida.* settings replaced by disassembler.* in the breaking cleanup and added Tern plan review editing and commenting, after 18.21.0 (the same afternoon) rebuilt the AgentSession run lifecycle around isRunActive, wakeSources, and waitForRunEnd, added inline image and web-page replies with tui.renderEmbeds as the opt-out, auto-detected reasoning support for unknown custom and discovered models, let overlapping Bash calls share shell state, and moved local text-to-speech to Phonon 31M; refreshed the push date to 2026-11-17 and the npm registry reference.
  • 2026-11-19 - Recorded npm reaching 19.0.0 (published November 19): Anthropic prompt-cache prewarm without billable output, subscription-plan metadata on Claude usage, message exclusion from context and compaction summaries, Claude account-failover recovery after weekly-limit resets, and Antigravity effort-suffix resolution, after 18.22.2 added the native Kiro provider and 18.22.1 the extension compaction lane; refreshed the push date to 2026-11-19 and the npm registry reference.
  • 2026-11-20 - Recorded npm reaching 19.0.1 (published November 19): Liquid AI judge models d1.2 and d1.2-ultra with liquid defaulting to d1.2, GIF judgments sent as PNG, the Anthropic strict-tool fallback retrying without strict on Vertex org-policy rejections, Kiro compaction recovery, true thinking-off, add-on-credit usage, and GovCloud region logins, gemini-4-neon on the Antigravity CLI roster, and per-model snapcompactShape frames; refreshed the push date to 2026-11-20, against the GitHub releases API and npm registry.

See also
#

  • Pi - the frozen-core parent whose minimalism omp exists to argue with
  • Claude Code - the closed platform benchmark omp’s feature table is measured against
  • OpenCode - the other large open harness in the same buyer’s shortlist
  • Harness Feature Matrix - the capability comparison this column joins

References
#