OrgX para Claude Code
useorgx/orgx-claude-code-plugin
Integra Claude Code com OrgX para orquestração de agentes e sync de trabalho
Ver no GitHubMCP server que conecta Claude Code ao OrgX através de OAuth nativo, permitindo sincronização de skills, orquestração de agents e gerenciamento de outputs revisáveis. Oferece acesso a ferramentas de busca, inspeção de status, progresso de iniciativas e atualização de trabalho sem expor credenciais ou dados sensíveis.
Instalação
npx convez-ai useorgx/orgx-claude-code-plugin1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Sincronizar status e progresso de trabalho entre Claude Code e OrgX
- Orquestrar agentes e executar operações através de MCP
- Anexar e revisar outputs gerados pelo Claude Code no fluxo OrgX
- Consultar próximas ações e crônicas operacionais sem deixar o editor
Veredito

Plugin maduro de integração que resolve acoplamento real entre Claude Code e OrgX via MCP com segurança apropriada (OAuth, sem credenciais em manifesto). Ideal para fluxos onde sincronização bidirecional de trabalho é crítica; requer familiaridade com MCP e OrgX.
README
A tradução para o português deste README ainda está na fila.
OrgX for Claude Code
OrgX connects Claude Code to current work and reviewable outputs through Claude Code's native OAuth flow. Status reads and explicit user-requested updates use the authenticated account's grants.
The plugin surface is intentionally small:
- one remote HTTPS MCP connection
- four user-invoked commands:
/orgx-claude-code-plugin:orgx-login,/orgx-claude-code-plugin:orgx-status,/orgx-claude-code-plugin:orgx-operator-chronicle, and/orgx-claude-code-plugin:orgx-attach-output - one static setup skill for connection troubleshooting
- one
Stophook and a dependency-free attachment outbox helper
Version 0.1.17 can retain an attachment request that the user explicitly queues.
At the end of a response, the Stop hook can ask the current Claude session to
retry that request through the existing OrgX MCP connection. The helper reads
only the hook's session identifier and loop flag; it does not read project files,
session text, or transcripts. OrgX receives only the requested operation's
inputs, subject to the authenticated account's permissions. The hook makes no
HTTP requests and reads no OAuth credentials.
Install from the public marketplace
claude plugin marketplace add https://github.com/useorgx/orgx-claude-code-plugin.git --scope user
claude plugin install orgx-claude-code-plugin@orgx --scope user
Anthropic directory publication is a separate distribution state. Anthropic's
documentation describes both a general, community-driven plugin directory and
the claude-plugins-official Claude Code marketplace. Treat the provider
submission receipt and resulting catalog entry as authoritative for the actual
publication destination, then verify a fresh install from the marketplace named
there. Until Anthropic confirms publication, the HTTPS marketplace above is the
public self-serve installation path.
The OrgX wizard is another installation path and requires its own release
receipt. A compatible build must copy root .mcp.json, the Stop hook
configuration and scripts/attachment-outbox.mjs, and remove older runtime
files. Verify those installed files and confirm that /mcp lists orgx;
an older wizard's successful exit does not establish compatibility.
Connect OrgX
- Open
/mcpin Claude Code. - Select
orgx. - Choose Authenticate and complete the OrgX consent flow in the browser.
- Return to
/mcpand confirm thatorgxis connected.
No API key or bearer token belongs in the plugin manifest, chat, or a checked-in file. Installation alone is not proof that OAuth succeeded.
The plugin connects to:
https://mcp.useorgx.com/mcp?profile=claude-code
That closed-world profile exposes exactly eight OrgX tools:
orgx_searchorgx_inspectorgx_get_next_actionsorgx_get_agent_statusorgx_get_initiative_progressorgx_get_operation_statusorgx_get_operator_brieforgx_update_work
The update operation accepts either a bounded content patch or one output attachment. It requires a stable retry key and the account's write grant. Attachment enters review; it does not approve or complete the work. The profile does not expose business-data deletion or workflow state-transition tools. Standard OrgX MCP usage may be recorded by the hosted service for operation and metering. The installed plugin adds no local telemetry or background reporting.
The claude-code selector uses the seven named informational operations plus
the existing orgx_update_work operation. The release manifest binds this exact
profile into its fingerprint. Existing read-only clients retain their seven
informational operations. Broader directory and runtime profiles are separate.
After updating an older installation, refresh its MCP connection and verify the eight authenticated tools above. A read-only OAuth grant can hide the write operation; the native consent flow must grant write access before attachment. If the profile is unavailable, stop the OrgX command and report the connection mismatch. Do not select a broader profile or guess a replacement tool. Profile changes require server deployment before a client release; local source updates are not evidence of publication.
User-invoked workflows
/orgx-claude-code-plugin:orgx-loginexplains the native OAuth connection flow./orgx-claude-code-plugin:orgx-statuscallsorgx_get_initiative_progressfor a requested initiative./orgx-claude-code-plugin:orgx-operator-chroniclecallsorgx_get_operator_brieffor the last 30 days./orgx-claude-code-plugin:orgx-attach-outputinspects the requested target, attaches one durable output throughorgx_update_work, then inspects the stored artifact. It presents the canonical answer, linked decisions, source notes, context, files and receipt when OrgX has recorded them. The user can explicitly queue an attachment with its selected HTTPS output URL, list queued requests, retry a selected request, clear a local request, or runrecoverafter an interrupted local queue write.
The source files keep their short names under commands/; Claude Code adds the
plugin namespace to installed slash-command invocations.
The commands are user-invoked. The Stop hook handles only explicitly queued
attachment requests from its own session.
Session security boundary
Every attachment attempt uses current orgx_inspect results to check the target
and workspace, then calls the existing orgx_update_work through Claude Code's
native OAuth connection. Stored request data and an earlier MCP response grant
no continuing authority. Expired or missing authorization requires native OAuth
reauthentication; a denied or invalid request stays paused.
The plugin does not read or transmit the current working directory, repository
path, or path aliases. Queued requests live in the user's home directory at
~/.config/useorgx/claude-code/attachment-outbox, with directory mode 0700 and
file mode 0600. It retains only the selected attachment request and its retry
state, with at most 20 entries of 64 KiB each. It does not discover files, upload
local paths, or retain tokens, session text or transcripts. Native OAuth and the
hosted MCP service own authorization expiry.
An ambiguous attempt retains its exact payload and stable retry key. The hook emits bounded instructions for the current Claude session; it cannot attach an output itself or publish a newly discovered output. Automatic retries stop after three attempts, and requests older than seven days pause. Authentication, permission and validation failures also pause. Only an explicit retry command resumes a paused request. Another session cannot automatically retry it. Clear removes the local request only; it does not delete an OrgX artifact.
Use the attachment command's explicit recover action if a crash interrupted a
local queue write. It refuses recovery while the recorded owner process is
alive. For a dead owner, it can release the stale lock and restore a valid
uncommitted request as paused with reason recovered, preserving its exact
payload and retry key. An explicit retry is still required. Recovery discards
incomplete staging data and rejects unknown files; it makes no MCP call. Do not
force a live lock or delete unrelated files.
Claude Code does not run Stop when the user interrupts a response or when an
API error ends the turn. Those paths leave an attachment pending; the hook does
not establish delivery.
Attachment succeeds only when the current MCP response identifies the stored artifact. Review, client acceptance and work completion remain separate. The outbox does not replace Wizard's runtime capture.
plugin.manifest.json records this boundary as a machine-checked contract.
Version 0.1.17 requires a new directory review. Local protocol tests establish
the hook and queue behavior; they do not prove native OAuth, a live attachment,
directory publication or a signed package. Deploy the matching app attachment
route and MCP claude-code profile before distributing this version.
Data and support
- Privacy: https://useorgx.com/privacy
- Terms: https://useorgx.com/terms
- Support: https://useorgx.com/support
- Security contact: reviewers@useorgx.com
- License: MIT
Development and validation
Requirements: Node.js 18+ and a current Claude Code CLI supporting exec-form command hooks. Plugin validation was exercised with Claude Code 2.1.309.
npm ci
npm run check
claude plugin validate . --strict
npm pack --dry-run --json
Record the actual CLI validation result for the release. Claude Code 2.1.309's strict validator currently reports pre-existing plugin-name and missing-icon warnings for this package; strict validation has not passed. Normal validation and a live OAuth attachment need their own recorded results.
Run the local plugin in an isolated Claude session:
claude --plugin-dir . -p "Reply with exactly: plugin-smoke-ok"
Release and directory-review evidence is tracked in:
docs/release-checklist.mddocs/anthropic-plugin-directory-submission.md
Source readiness, submission, review, approval, directory publication, and an Anthropic Verified badge are distinct states.
About OrgX
OrgX is proof for AI-delivered work — the operating record your agents and tools share. Agents in ChatGPT, Claude, Cursor, and Codex work from one record of decisions, approvals, and work receipts.
- OrgX FAQ: what it is, who it's for, and what it costs
- Agent work receipts: how finished agent work is checked and signed off
- Pricing
- Hosted MCP server: useorgx/orgx-mcp