Grok Agent Bridge
Runs OpenAI Codex as a native worker via its app-server protocol, letting the bridge route tasks to an exact provider/model, track durable routing receipts and session IDs, and surface verification results while keeping native approvals and account policy intact.
Click on "Deploy 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., "@Grok Agent BridgeStart a Codex task to fix the failing auth tests"
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.
Vesper Grok Bot Bridge
Use Grok Bot to supervise coding agents: start a task, inspect the agent's response, resolve delegated technical input, request improvements in the same session, and review real verification before accepting work.
Explore the system design diagrams, technology stack and architecture walkthrough.
Choose your setup
We maintain two focused Bot templates, with different tool contracts. Full route comparison.
What you use | Bot template | MCP server | Setup guide |
T3 Code | T3 already supplies its server through T3 Connect; no bridge install | ||
Coding CLIs without T3 | Install this project's CLI/MCP bridge once; enroll multiple ACP/native CLIs |
Recipients add a published Bot through the Grok Bot Marketplace or a shared template link. The profiles above are the publisher source files. Both templates are published and shareable: Vesper — CLI and Vesper — T3; inside the Grok Bot app the one-click add links are grokbot://app/v1/bot-template?id=BuUvKwDGbSoqwNvNDreQk and grokbot://app/v1/bot-template?id=H2br1687IFJlyapDHroic. Custom MCP servers are installed separately; importing a Bot does not install software or authenticate coding accounts. Template publishing details.
A combined Bot is optional for advanced users who connect both resources and select the route explicitly. It is not the default offering and must never silently move work between backends. There is one MCP server implementation in this repository, not one server per CLI. T3's server is a separate product.
Related MCP server: coding-agent-mcp
System architecture
The two routes are selected explicitly and use separate MCP servers and credentials.
flowchart TB
Owner["Owner / developer"]
Bot["Grok Bot / Vesper templates<br/>Supervision, technical answers, review"]
Owner -->|"Task and acceptance criteria"| Bot
subgraph T3Route["Route A - T3 product boundary"]
Relay["T3 Connect<br/>HTTPS /mcp + T3 OAuth"]
T3["T3 Code MCP server<br/>t3_* tools, projects and threads"]
T3Workers["T3-configured coding providers<br/>Native sessions and workspaces"]
Relay <-->|"T3-managed relay connection"| T3
T3 <-->|"T3-owned orchestration"| T3Workers
end
subgraph CLIRoute["Route B - owner coding computer"]
Gateway["Owner HTTPS ingress<br/>Reverse proxy / tunnel / optional Funnel"]
Bridge["Standalone MCP bridge<br/>Node.js 22+, TypeScript, Express 5<br/>MCP SDK + Zod"]
Router["TaskRouter + BridgeService<br/>Exact provider/model, session IDs<br/>Durable receipts and bounded monitoring"]
Native["Native workers<br/>Codex app-server / Claude stream-json<br/>ACP v1: OpenCode, Grok Build, enrolled adapters"]
Work["Detached Git worktrees<br/>Owner verifier + SHA-256 change evidence"]
State[("Owner-local files<br/>JSON snapshots, routing receipts<br/>Protected credentials and runtime records")]
Gateway <-->|"HTTP to 127.0.0.1<br/>Bridge OAuth bearer on /mcp"| Bridge
Bridge <-->|"Validated agent_* calls"| Router
Router <-->|"Child-process stdio protocols"| Native
Native <-->|"Repository edits and project checks"| Work
Router <-->|"Persist / recover"| State
Router -->|"Verify then finish"| Work
end
Bot <-->|"MCP over HTTPS<br/>T3 OAuth credential"| Relay
Bot <-->|"MCP over HTTPS<br/>Separate bridge OAuth credential"| Gateway
Native <-->|"Provider-managed network / authentication"| APIs["Selected provider services<br/>Native CLI owns account authentication"]
classDef core fill:#dbeafe,stroke:#2563eb,color:#10254a
classDef external fill:#f1f5f9,stroke:#64748b,color:#172033
class Bridge,Router core
class Owner,Bot,APIs,Relay,T3,T3Workers externalAll five diagrams are saved in docs/diagrams, with editable Mermaid sources and SVG exports. See the full system-design guide for networking, OAuth, task verification, optional integrations and the technology inventory.
Standalone MCP: install once
Requires Node.js 22+, npm, Git and your chosen authenticated coding CLI/ACP adapter. Install the published package from npm:
npm install -g grok-bot-mcp-agent
grok-agent-bridge setup --repository C:/Code/my-project
grok-agent-bridge doctorPrefer a checksum-verified offline archive? Download the runtime archive from a trusted GitHub release, verify its SHA-256 sidecar, then install it by local path (npm install -g ./grok-bot-mcp-agent-<version>.tgz).
For repositories without a suitable detected test/check command, enroll the real verifier explicitly. Then enroll your exact ACP executable/model, configure a stable owner-controlled HTTPS gateway and approve OAuth locally. Continue with the complete standalone guide; the three commands above alone do not connect Grok Bot's cloud connector to your computer.
Setup persists per-user paths and credentials and configures supported sign-in startup. Workers run in background isolated Git worktrees; no visible terminal or T3 installation is required. A single installed bridge can operate multiple owner-enrolled CLIs. Local MCP hosts can use grok-agent-bridge connect over stdio. Grok Bot uses the authenticated HTTPS/OAuth route. No xAI API key is needed when Grok Bot supplies reasoning.
T3 MCP: connect once
Install and configure T3 Code and its coding provider, enable T3 Connect, then add your environment's HTTPS /mcp URL as a remote Grok Bot connector and approve native OAuth. Do not install this bridge merely to use T3. Follow T3 setup and acceptance and the T3 Bot role.
Permissions, accounts and availability
Use your own accounts and approved authentication. Keep native approvals, administrator controls and budgets in force. Credentials, MFA, CAPTCHA, payment and human-only approvals remain human handoffs. Stop on quota/policy failures; no account rotation, token extraction, silent model substitution or uncertain task replay. The bridge is not an OS sandbox or a multi-tenant credential proxy.
Read the current provider-policy review before use or publication. It covers OpenAI, Anthropic, xAI and other workers, including subscription-versus-API and hosted/commercial boundaries. ACP support and passing tests are not vendor approval, and nobody can guarantee immunity from account restrictions.
CLIProxyAPI-style multi-account pooling, automatic failover on quota/authentication errors, and client fingerprint cloaking are not integrated; the assessment explains the account and contract risk.
A VPN or Tailscale tunnel protects the network path, not the provider account. See remote access and network safety for T3 Connect, Tailscale Serve/Funnel and tunnel boundaries.
The coding computer must stay awake and online with its selected backend and tunnel running. Persisted enrollment is not a promise of indefinite OAuth validity. Native Grok Bot approval cards can still interrupt unattended work; physical reboot and app-closed operation remain separate acceptance limits. Live results and limitations, verification history.
Develop and distribute
npm ci
npm run lint
npm test -- --no-file-parallelism --maxWorkers 1
npm run build
npm run pack:communityThe final command builds the installable CLI/MCP runtime and a publisher source kit containing both Bot definitions, with SHA-256 sidecars and a release manifest under build/releases/. Ordinary marketplace users do not need the source kit. Nothing is published, enrolled or installed globally by that build. Distribution guide.
See CONTRIBUTING.md for the directory map and contribution workflow, and the documentation index for all guides. The optional legacy API-key supervisor is documented separately.
Unofficial community integration under MIT; no vendor endorsement, account-policy approval, public npm publication or universal provider certification is implied.
This server cannot be deployed
Maintenance
Related MCP Connectors
Remote MCP learning coach for coding agents.
Coordinate coding agents through MCP using existing AI plans, saved work, and independent checks.
A paid remote MCP for OpenAI Codex agent coordination MCP, built to return verdicts, receipts, usage
Hosted MCP server and Agent Skills for specs, engineering threads, and project memory.
Related MCP Servers
- AlicenseNot gradedqualityDmaintenanceEnables Codex to delegate code reviews, coding tasks, image/video generation, and background investigations to Grok via MCP tools.28Apache 2.0
- AlicenseAqualityBmaintenanceEnables supervising and orchestrating local coding agents like Muse and AGY through a client-agnostic MCP interface, with isolated Git worktrees, safe execution, and deterministic verification.107 npmMIT
- AlicenseNot gradedqualityBmaintenanceEnables supervising and coordinating local coding agents across allowed repositories, including inspecting context, reviewing sessions, launching or continuing tasks, and retrieving redacted logs under owner-controlled access.7 npm1MIT
- AlicenseNot gradedqualityCmaintenanceEnables Claude and Codex agents to delegate implementation, testing, research, or review tasks to each other through local MCP tools, while supporting persistent sessions, isolated editing workspaces, and reviewed integration. It also allows progress tracking, cancellation, model selection, and management handoff between providers.2MIT