OpenCode Advisor MCP
The OpenCode Advisor MCP server provides a single tool (ask_opencode_advisor) to request a read-only second-opinion code review of Git changes from a local OpenCode advisor agent.
What you can do:
Request a code review of current Git changes (diff and/or status)
Ask a specific question (
question) to focus the advisor on particular concerns (e.g., risks, missing tests, privacy issues)Provide a high-level goal (
goal) to give the advisor context about what the changes aim to achieveScope the review to a specific working directory (
cwd) or particular file paths (paths)Compare against a specific base ref (
base_ref) to control which Git diff is reviewedInclude or exclude the Git diff (
include_diff) or status (include_status) from the review contextLimit diff size (
max_diff_chars) to control how much diff content is sent to the advisorReceive structured JSON responses with stable error codes (e.g.,
invalid_cwd,git_failed,timeout) for predictable error handling
The advisor operates in read-only mode — the bundled template is designed to block writes and deny .env reads.
Allows Codex (OpenAI's coding assistant) to request a review of Git changes from an OpenCode advisor, focusing on risks, tests, privacy, and release readiness.
Click on "Install Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@OpenCode Advisor MCPReview the current git diff for potential bugs and security issues."
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
OpenCode Advisor MCP
OpenCode Advisor MCP gives Codex three local MCP tools:
ask_opencode_advisorfor a read-only reviewer passask_opencode_plannerfor a read-only planning passget_opencode_taskfor queued or running work
The MCP server has exactly two built-in roles: reviewer and planner. It does not reuse normal OpenCode, Codex, or Cockpit provider settings. The bundled agents can reason only over the request plus supplied Git status/diff context and, for the planner, the supplied plan; they have no file or shell tools.
Install And Configure
The current supported install path is the verified tarball attached to the GitHub v0.3.0 Release. This release process did not publish to npm, and no npm-registry install path is supported for v0.3.0.
Invoke-WebRequest https://github.com/henrydontbbai/opencode-advisor-mcp/releases/download/v0.3.0/opencode-advisor-mcp-0.3.0.tgz -OutFile opencode-advisor-mcp-0.3.0.tgz
Invoke-WebRequest https://github.com/henrydontbbai/opencode-advisor-mcp/releases/download/v0.3.0/SHA256SUMS.txt -OutFile SHA256SUMS.txt
$expected = "47a4697ad28e99fd85ba2951ac21289a566378948743526f2b1cde5cbd905fa1"
$manifest = (Get-Content .\SHA256SUMS.txt -Raw).Trim()
if ($manifest -ne "$expected opencode-advisor-mcp-0.3.0.tgz") { throw "Release manifest mismatch" }
$actual = (Get-FileHash .\opencode-advisor-mcp-0.3.0.tgz -Algorithm SHA256).Hash.ToLowerInvariant()
if ($actual -ne $expected) { throw "Release checksum mismatch" }
npm install -g .\opencode-advisor-mcp-0.3.0.tgz
opencode-advisor-setupFor another local packed tarball, install that exact file:
npm install -g <path-to-opencode-advisor-mcp.tgz>
opencode-advisor-setupFor source development or as a fallback, install dependencies and run setup from the checkout:
npm ci
npm run setupopencode-advisor-setup requires an interactive terminal. It asks for a third-party provider ID, display name, API base URL, transport (responses or chat_completions), model list, reviewer model and optional reasoning variant, planner model and optional reasoning variant, and API key. The key is hidden at entry and is never accepted from command arguments, MCP TOML, or a pipe. Setup does not inspect normal OpenCode, Codex, or Cockpit configuration, credentials, or account-login state; no OpenCode account login is required.
The setup command creates an independent profile under %USERPROFILE%\.codex\opencode-advisor on Windows or $HOME/.codex/opencode-advisor on POSIX. Windows credentials use CurrentUser DPAPI. POSIX uses a filesystem-permission fallback: private 0700 profile directories and a 0600 Base64-encoded credential envelope, rather than DPAPI-equivalent encryption. The profile contains the bundled codex-advisor and codex-planning-partner agent templates plus a non-secret provider manifest.
The credential is bound to the exact manifest fingerprint, and the generated OpenCode overlay must also match that manifest. An incomplete profile written during setup, stale overlay, or binding mismatch is fail-closed: MCP calls return setup guidance without queuing work. If setup ends before it writes profile artifacts, a previously valid profile remains usable. Rerun opencode-advisor-setup instead of editing profile artifacts by hand.
Related MCP server: codex-claude-bridge
MCP Configuration
For a source checkout, use Node and the absolute server path. The only required environment setting is OPENCODE_ADVISOR_ALLOWED_ROOTS:
[mcp_servers.opencode_advisor]
command = "node"
args = ["C:\\absolute\\path\\to\\opencode-advisor-mcp\\src\\server.mjs"]
startup_timeout_sec = 30
tool_timeout_sec = 420
[mcp_servers.opencode_advisor.env]
OPENCODE_ADVISOR_ALLOWED_ROOTS = "C:\\workspace\\allowed-repositories"For a GitHub Release or local tarball installed globally, use command = "opencode-advisor-mcp"; examples/codex-mcp.toml shows that installed-package form.
Do not put provider URLs, model IDs, API keys, tokens, or OPENCODE_CONFIG_CONTENT in the MCP configuration. See docs/CONFIGURATION.md for non-secret runtime settings.
Provider Compatibility
The setup transport is explicit:
responsesgenerates an@ai-sdk/openaiprovider. OpenCode owns the Responses streaming SSE exchange, including output-text events such asresponse.output_text.delta/response.output_text.done, error events (error,response.failed), and function-call events.chat_completionsgenerates an@ai-sdk/openai-compatibleprovider and uses standard/v1/chat/completionsstream chunks followed by[DONE].
Each role can optionally select an OpenCode model variant. The choices are independent even when reviewer and planner use the same model; a common setup is reviewer high and planner max. Leave a variant empty to use the model default. For a responses provider/model that supports the selected variant, the generated OpenCode model variant reaches the Responses request as reasoning.effort. high and max are examples, not values guaranteed by every provider or model.
The built-in agents deny all tool execution. A streamed function call is covered only as a fail-closed condition: it is not executed, does not complete a tool round-trip, and the built-in-agent fixture expects a timeout rather than a reviewer or planner result. Raw provider tool events are not MCP output; only structured assistant text can become a reviewer or planner result. The provider key belongs only to the independent Advisor profile.
Verify
From a source checkout:
npm run smoke
npm test
npm run test:doctor
$env:OPENCODE_ADVISOR_ALLOWED_ROOTS = "<allowed-root>"
npm run doctoropencode-advisor-doctor checks the independent profile, both roles, structured JSON output, and sanitized MCP response shape. A missing or unreadable profile remains an opencode_failed MCP result with setup guidance; it never creates a queued task.
Use opencode-advisor-doctor --json (or npm run --silent doctor -- --json from source) for a machine-readable report. JSON mode preserves the same checks and exit status as the default text report and does not add provider settings or credentials.
Security Boundaries
Reviewer and planner child processes run with
--pureand an explicitprovider/model.OpenCode receives isolated XDG and OpenCode configuration paths plus a generated
OPENCODE_CONFIG_CONTENToverlay.OPENCODE_DISABLE_PROJECT_CONFIG=1prevents the reviewed repository's OpenCode configuration from being merged.Inherited OpenCode, XDG, provider-key, token, secret, and password environment variables are removed before OpenCode starts.
Queue files, runner logs, doctor reports, and MCP responses do not include provider URLs, model selections, role variants, or API keys.
A custom
OPENCODE_ADVISOR_OPENCODE_CMDon Windows must name an existing, operator-trusted absolute.exe; command strings with arguments and.cmd/.batwrappers are rejected.Keep
OPENCODE_ADVISOR_ALLOWED_ROOTSnarrow and use this only for repositories you are authorized to disclose to the configured provider.
Further details:
Maintenance
Resources
Unclaimed servers have limited discoverability.
Looking for Admin?
If you are the server author, to access and configure the admin panel.
Related MCP Servers
- AlicenseAqualityFmaintenanceAn MCP server that provides code review functionality using OpenAI, Google, and Anthropic models, serving as a "second opinion" tool that works with any MCP client.Last updated11533MIT
- Alicense-qualityAmaintenanceMCP server that enables Claude Code to send code and plans for review by OpenAI Codex, returning structured feedback inline.Last updated44820MIT
- Alicense-qualityCmaintenanceAn MCP server that provides agentic code review powered by OpenAI-compatible models, designed for use with Claude Code.Last updated1MIT
- Alicense-qualityBmaintenanceMCP server enabling Claude to consult Codex (GPT-5.x) mid-task for second opinions, plan/diff review, brainstorming, and codebase exploration via structured debates and permission-controlled interactions.Last updated2MIT
Related MCP Connectors
An MCP server that gives your AI access to the source code and docs of all public github repos
A MCP server built for developers enabling Git based project management with project and personal…
Official MCP server for Lovable, the AI-powered full-stack app builder.
Latest Blog Posts
- Who's Calling? MCP Hosts Are an Identity Blind Spot (And the Spec Knows It)By Om-Shree-0709 on .mcpAgent IdentityOAuth 2.1
- Your AI Chatbot Just Exposed Your CEO's Salary to an InternBy Om-Shree-0709 on .Agent IdentityMCP SecurityOAuth Delegation
- Why MCP Servers Need Execution Sandboxing (And Why Your Current Stack Isn't Enough)By Om-Shree-0709 on .Agentic AiPrompt InjectionWebAssembly
MCP directory API
We provide all the information about MCP servers via our MCP API.
curl -X GET 'https://glama.ai/api/mcp/v1/servers/henrydontbbai/opencode-advisor-mcp'
If you have feedback or need assistance with the MCP directory API, please join our Discord server