omagents
Discovered★ 360Agent toolkit for OpenCode, Codex, and DeepSeek Harness: skills, MCP, background agents, and durable job tracking
OmAgents
OmAgents is a multi-host agent toolkit for OpenCode, Codex, and DeepSeek Harness (DSH). It bundles 18 OmAgents skills, 15 Superpowers skills, MCP servers, native background-agent integration, deep research, and Python tooling. A shared Core package keeps state and host-neutral behavior consistent.
Installation
Choose the package for your host. The installer takes no host positional argument.
# OpenCode
npx @omagents/omagents
# Codex
npx @omagents/omagents-codex
# DeepSeek Harness
dsh plugin --profile web add @omagents/omagents-dsh
Restart the host after installation. @omagents/core is installed automatically as a dependency; do not install it separately as a plugin. The current release candidate aligns all four packages at 0.14.1.
Python 3.11+ must be available on PATH for Python-backed skills. DSH requires Node.js 24.9 or newer and a DSH version in the package's tested compatibility range.
Optional API Keys
export EXA_API_KEY="your-exa-key"
export CONTEXT7_API_KEY="your-context7-key"
export GITHUB_TOKEN="your-github-token"
EXA_API_KEY: authenticates the hosted ExawebsearchMCP throughx-api-key; without it, the same endpoint works anonymously.CONTEXT7_API_KEY: optional rate-limit credentials for the local@upstash/context7-mcpserver.GITHUB_TOKEN: enables the GitHub Copilot MCP. Without it, OmAgents usesmcp.grep.appfor public code search.
For Codex, GitHub authentication is stored as the GITHUB_TOKEN environment-variable reference (bearer_token_env_var), not the token value. The Codex plugin .mcp.json contract has no equivalent indirection for Exa's x-api-key or Context7's stdio env, so configured Exa and Context7 values are written to that file. On platforms that honor POSIX modes, the installer creates private directories with mode 0700 and writes or rewrites .mcp.json with mode 0600.
Support Matrix
| Host or role | Package | Installation | Skills | Background execution and recovery | Management surface |
|---|---|---|---|---|---|
| OpenCode 1.x and OpenCode 2.x | @omagents/omagents | npx @omagents/omagents | 33: 18 OmAgents + 15 Superpowers | V1 uses task(background: true); V2 uses native subagent(..., background: true). Durable Core Job Board, completion delivery, and recovery. | /ps, parallel_status, cancel_task, parent-context results, and OpenCode TUI state |
| Codex | @omagents/omagents-codex | npx @omagents/omagents-codex | 33: 18 OmAgents + 15 Superpowers | Codex-native subagents. The OpenCode parallel engine and DSH Web Job Board are not installed. | Codex-native interface and skills |
| DeepSeek Harness (DSH) | @omagents/omagents-dsh | dsh plugin --profile web add @omagents/omagents-dsh | 33: 18 OmAgents + 15 Superpowers | Native asynchronous subagent lifecycle, durable job-board.json, cancellation, recovery, /ps, parallel_status, and cancel_task | Authorized Remote watchJobs stream plus a native React Web Job Board |
| Shared library, not a host plugin | @omagents/core | Automatic dependency; no separate user install | Shared skill definitions and Superpowers assets | Host-neutral execution state, persistence, MCP resolution, research assets, and Python environments | No standalone UI |
All host packages depend only on the aligned Core version, never on another host package.
Highlights
| Feature | What it provides |
|---|---|
| Loop Engineering | Durable task queues, retries, verification, and resumable state for iterative skills |
| Superpowers | Brainstorming, TDD, systematic debugging, plans, code review, worktrees, and verification |
| Deep Research | Multi-source items × fields research, gap detection, durable evidence, and Jinja2 reports |
| Parallel Execution | Host-native background agents with a shared Core execution model on OpenCode and DSH |
| DSH Web Job Board | Native React filters, details, notification state, confirmed cancellation, reconnect, and health displays |
| Built-in MCPs | agentmemory, codegraph, context7, websearch, and github/grep_app |
| Python Tooling | Persistent agent base and on-demand project overlays without modifying an existing project .venv |
| Document and Web Tools | MarkItDown conversion, OfficeCLI documents, and Playwright scraping |
| Code Intelligence | LSP guidance and AST-aware search/rewrite with grep fallback |
| Hyperplan and Refactor | Three-critic adversarial planning and verified iterative refactoring |
Python Environments
Only jinja2 is automatically provisioned in the persistent agent base. Optional Python tools such as MarkItDown and Playwright are installed on demand when their skills need them; they are not part of automatic setup. Project dependencies use an on-demand overlay and never require modifying an existing project .venv.
Use the agents-python-tools skill for environment selection, cross-platform paths, and CLI instructions. Project commands are implemented by packages/core/hooks/project-venv.js in this repository. Neither the base nor a project environment is automatically deleted.
Loop Engineering and Deep Research
Loop-based skills process one item at a time and persist state in .omagents/loops/<skill>/tasks.json. Failed tasks retry up to three times before becoming blocked. State survives context clearing so the workflow can resume instead of restarting.
Deep Research uses --workspace <dir> only for public deliverables. Plans, raw findings, checkpoints, derived artifacts, provenance, and run-scoped queues live under .omagents/deep-research/<run-id>/. New runs reject a workspace that would contain that internal state; legacy workspaces remain readable and resumable without automatic migration. Jinja2 templates live in packages/core/skills/deep-research/templates/; there is no root-level templates/ directory.
Skills and MCP Servers
The 18 OmAgents skills are:
deep-research, parallel-execution, agents-python-tools, markitdown-converter, officecli, playwright-web-scraping, init-deep, doctor, remove-ai-slops, remove-deadcode, github-triage, tech-debt-audit, lsp-guide, ast-grep, work-with-pr, pre-publish-review, hyperplan, and refactor.
Superpowers contributes 15 additional process skills. Core renders shared skills for opencode-v1, opencode-v2, codex, and dsh, using complete host-specific overrides where control flow differs. DSH uses native subagent semantics without an OpenCode background argument.
| MCP | Type | Behavior |
|---|---|---|
agentmemory | Local | Session memory and audit |
codegraph | Local | Codebase symbol graph and exploration |
context7 | Local | Documentation search through @upstash/context7-mcp |
websearch | Remote | Hosted Exa MCP, anonymous or authenticated with EXA_API_KEY |
github / grep_app | Remote | GitHub Copilot MCP with a token; otherwise public code search through mcp.grep.app |
User-defined MCP configuration takes precedence over built-in registrations.
DSH Job Board Behavior
The DSH package installs a complete Host and Web Client, not a skills-only plugin. The Web Job Board consumes revisioned whole-set Remote frames, starts with a fresh baseline after reconnect, and does not poll running jobs. It provides all/running/attention/finished filters, result and error details, notification state, cancellation confirmation, stale/permission/degraded views, and DSH theme integration.
Cancellation acceptance is not confirmed termination. Local interruption stops the child's current turn, and its descendants continue unless cancelled separately. External one-shot runs stop permanently and cannot resume. Viewing details in the UI never acknowledges a model notification.
On restart, DSH reconciles persisted jobs against authoritative host state. Missing or unreadable outcomes become interrupted only when no active run is authoritatively observable, never fabricated successes. An observed active external run remains active on a cold session-not-found instead of being misclassified as interrupted. Remote access is limited by the active DSH profile and authorized parent-session scope. See the DSH package guide and Job Board guide for UI, cancellation, recovery, authorization, security, and troubleshooting details.
Architecture
┌─────────────────────────────────────────────────┐
│ User choice layer (not bundled) │
│ OpenSpec · gstack · custom workflows · none │
├─────────────────────────────────────────────────┤
│ Process skills layer (bundled Superpowers) │
├─────────────────────────────────────────────────┤
│ OmAgents infrastructure and shared Core │
│ Skills · MCP · Job Board · research · Python │
├─────────────────────────────────────────────────┤
│ OpenCode · Codex · DeepSeek Harness │
└─────────────────────────────────────────────────┘
OmAgents supplies infrastructure, not a mandatory development methodology. Superpowers supplies reusable process skills. Methodologies such as OpenSpec and gstack remain separate user choices.
Development
The repository is a private npm-workspaces root containing packages/core, packages/opencode, packages/codex, and packages/dsh. The repository .opencode/plugins/index.js is only a development re-export; product code lives in the packages. Deep-research templates remain under packages/core/skills/deep-research/templates/.
# All root and workspace tests
npm test
# Build and test DSH
npm run build --workspace @omagents/omagents-dsh && npm test --workspace @omagents/omagents-dsh
# Packed-artifact and browser verification
npm run pack:all
npm run test:packed
npm run test:browser
Every release must use one aligned version and build, test, pack, and publish all four packages together, in this deterministic order:
@omagents/core → @omagents/omagents → @omagents/omagents-codex → @omagents/omagents-dsh
Core is published first because every host depends on it. The hosts are sequential only for deterministic logs and failure handling; there are no host-to-host dependencies. See CONTRIBUTING.md for verification requirements and the current platform/browser confirmation scope.
Uninstallation
Use the selected host's plugin management to remove or disable its OmAgents package. For DSH, disabling the bare @omagents/omagents-dsh Loader row disables both the Host and its Web Client. Core remains only as a dependency of an installed host. Python environments are not automatically deleted.
License
MIT — see LICENSE.
Comments
Loading…
Similar plugins
by useorgx
OrgX Work Ledger, MCP tools, skills, proof, and governed execution for DeepSeek Harness
★ 0
MIT
JavaScript
Oct 29, 2026
dsh plugin --profile web add @useorgx/deepseek-harness-pluginby allentnetus
Job Hunting DSH plugin and runtime skill scaffold for local job intelligence workflows.
★ 0
MIT
JavaScript
Aug 22, 2026
dsh plugin --profile web add dsh-job-huntingby leechen298
Generate Function, MCP, Agent Skill, and offline test packages from existing code; installable as a DeepSeek Harness bundle.
★ 11
MIT
Python
Oct 7, 2026
dsh plugin --profile web add @leechen298/code2skillby limuyang2
Multi-agent team collaboration for DeepSeek Harness, with independent models, skills, MCP tools, contexts, and a shared workspace.
★ 52
↓ 590/wk
MIT
TypeScript
Aug 26, 2026
dsh plugin --profile web add @limuyang2/dsh-agent-teamby pure-craft
DeepSeek Harness plugin: see what your agent can actually reach — skills, MCP servers, system tools with true in-context state, and per-session / per-preset switches
★ 13
↓ 588/wk
MIT
TypeScript
Oct 19, 2026
dsh plugin --profile web add dsh-capability-panelby chushixixin
Expose DeepSeek Harness agent capabilities as an MCP server (brain=Hermes, arms=Harness)
★ 16
↓ 99/wk
MIT
TypeScript
Aug 15, 2026
dsh plugin --profile web add @chushixixin/dsh-harness-mcp-server