ForgePilot MCP
OfficialClick 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., "@ForgePilot MCP research the Stripe API and generate a Claude MCP config snippet"
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.
Aitherium Agent Pack — ForgePilot
ForgePilot is a hackathon-ready agent pack for building and shipping agent integrations from wherever the work already happens.
The hackathon demo is deliberately self-referential: it shows the process used to build this pack. A plain-language request becomes research, a local aw* bootstrap plan, a scoped identity/policy boundary, desktop connectors, a phone PWA, and a shareable artifact.
It gives you one local runtime with four surfaces:
an Aitherium/awdk pack and fleet definition;
a small MCP server for Codex, Claude Desktop/Code, ChatGPT-compatible MCP clients, and any MCP host;
a WebMCP-enabled control room for browser agents;
generated OpenRouter and Exa adapter configs.
The demo story is simple: ask ForgePilot to research an API or integration, get grounded Exa sources, then produce a portable setup bundle for the coding client you are using.
The larger hackathon story is Agents Everywhere: bootstrap/AGENT_EVERYWHERE.md walks a new builder from zero to awdk, Bonsai/local inference, MCP, awsh, desktop, and secured remote connections.
Backend switching is a first-class path too: run bash bootstrap/switch-claude-backend.sh code (or the PowerShell equivalent) to use adk claude-model code for the DeepSeek Flash coding profile, then verify with adk claude-model check. The pack exposes the same plan through forgepilot_backend_switch_plan.
The pack is also designed to be opened alongside Aitherium Studio: use pack.yaml as the identity/tool contract, fleet.yaml as the orchestration view, and the control room as the browser-native demo surface. Use SHAREABLE_PACKS.md when you want to turn a new idea into a portable MCP + WebMCP + PWA tool that other builders can apply.
Quick start
Windows
Set-ExecutionPolicy -Scope Process Bypass
./setup.ps1macOS/Linux
chmod +x setup.sh
./setup.shThe setup script creates .venv, installs this project plus the aw* acceleration and authorization bricks, writes .env, and emits client config snippets under integrations/generated/. For an offline demo-only install, set SKIP_AWSDK=1.
Add EXA_API_KEY for live research and OPENROUTER_API_KEY for model-backed chat. The app still starts without either key in demo mode, so the control room and MCP handshake can be shown offline.
Start the control room:
python -m aitherium_pack serveOpen http://127.0.0.1:8787.
GitHub Pages
The repository ships an automated static deployment. Enable Settings → Pages → GitHub Actions, then push main. The pages.yml workflow builds the PWA and deploys it to the repository's Pages URL; verify.yml tests every pull request and checks the Pages bundle. The hosted page is a client only: configure its MCP endpoint to your own authenticated HTTPS ForgePilot/Aitherium node. See DEPLOY.md for CORS, token, and repository-variable setup.
Or create a durable run artifact directly:
python -m aitherium_pack run "Design a safe Codex and Claude MCP integration for this project"That writes artifacts/runs/<run-id>/ with research.json, response.json, report.md, stack.json, and a SHA-256 manifest.json.
If you want the native Aitherium fleet server after setup:
adk-serve --fleet fleet.yaml --port 8080Related MCP server: mcp-project-helper
What is included
pack.yaml awdk-style pack metadata
fleet.yaml awdk fleet definition
aitherium_pack/
cli.py one-command setup/serve/config generation
server.py stdlib HTTP + MCP JSON-RPC server
providers.py Exa and OpenRouter adapters
config.py generated client config helpers
web/index.html ForgePilot control room + WebMCP tools
skills/ portable pack skills
integrations/ checked-in examples and generated snippetsDesktop integrations
The generated files are intentionally plain JSON snippets so they can be copied into the client of choice:
Claude Desktop / Claude Code:
integrations/generated/claude.mcp.jsonCodex-style MCP clients:
integrations/generated/codex.mcp.jsonChatGPT or any remote MCP client:
integrations/generated/remote-mcp.json
For a local stdio client, the server command is python -m aitherium_pack mcp. For a remote client, expose the HTTP server through a trusted tunnel and use /mcp. The HTTP server binds to loopback by default; non-loopback mode requires FORGEPILOT_MCP_TOKEN plus an explicit trusted-LAN flag. CORS is origin-allowlisted and request bodies are bounded.
awdk connection
The pack is shaped for the current awdk workflow: pip install awdk, adk install pack:..., a named identity, and a fleet YAML. When awdk is installed, python -m aitherium_pack install-awdk --cloud runs the native cloud quickstart path and leaves this pack as the project-owned integration layer. The local MCP server remains dependency-light so the hackathon demo can run even before a full model backend is configured.
Identity and capability boundary
The device-flow and local-stack plans are designed around the Aitherium front door: awnboard receives the request, awiam identifies the caller/session, awbac makes a fail-closed capability decision, awdit records it, and awtunnel handles private reachability. Only after approval does awsh/awdk reach Bonsai-27B, awm, awrepl, awgraph, awgit, or ForgePilot MCP.
Safety boundary
ForgePilot’s first-party tools are read-only: health, web research, grounded chat, and config generation. Research results are marked untrusted before entering the model prompt. No email, purchase, file mutation, shell, or browser-navigation tool is exposed by default.
Sources
Available Tools
12 toolsforgepilot_backend_switch_planB
Show the safe adk/awsettings plan for switching Claude Code to a provider such as DeepSeek Flash. Read-only.
Inputs: profile.
forgepilot_chatC
Answer an integration question using optional Exa evidence through OpenRouter. Read-only.
Inputs: message (required), evidence.
forgepilot_community_bridge_planB
Show the safe ARC, awgym, awrelay, Relay, Forums, Spaces, and AwDesk connection plan. Read-only.
Inputs: no named parameters.
forgepilot_create_artifactB
Create a timestamped, read-only integration run bundle with evidence, response, report, and SHA-256 manifest.
Inputs: request (required).
forgepilot_create_community_bridge_packC
Create a no-secrets community bridge artifact for ARC, awgym, awrelay, Relay, Spaces, and AwDesk.
Inputs: name.
forgepilot_create_connector_artifactB
Generate runnable Claude Code and Codex MCP configs plus bootstrap scripts for the current local checkout.
Inputs: no named parameters.
forgepilot_generate_configB
Generate portable MCP client configuration snippets. Read-only.
Inputs: no named parameters.
forgepilot_healthB
Return the local ForgePilot runtime status. Read-only.
Inputs: no named parameters.
forgepilot_local_acceleration_planB
Inspect local aw* command/package presence and return the Bonsai model activation plan for Codex and Claude Code. Read-only.
Inputs: model.
forgepilot_searchB
Search the web with Exa and return compact, untrusted highlights with source URLs.
Inputs: query (required), num_results.
forgepilot_start_device_flowC
Create a consent-scoped Aitherium device-flow handoff for local inference, MCP, workspace, and LAN capabilities. No token is returned.
Inputs: scopes.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 12 tools
Several tools generate configs or artifacts (generate_config, create_connector_artifact, create_artifact, create_shareable_pack, create_community_bridge_pack) with overlapping 'create' verbs, and chat/search both use Exa, so misselection is possible despite descriptive text.
All tools use the forgepilot_ snake_case prefix, but four tools are noun-phrase plans (local_acceleration_plan, backend_switch_plan, community_bridge_plan) while the rest use verb or verb_noun patterns, a minor deviation.
12 tools is well within the ideal 3-15 range and each tool maps to a distinct ForgePilot integration task (health, search, config, artifacts, plans), so the count is appropriate.
The surface covers many one-way generation/planning tasks but lacks lifecycle operations: no list, get, update, or delete for artifacts/packs, and device flow has only a start tool with no completion/poll tool.
Maintenance
Related MCP Connectors
Hosted MCP server for live public-data APIs and Skills for AI agents.
Agent-native MCP server over the public saagarpatel.dev corpus. Read-only, stateless.
Remote MCP server for supportsheep: run AI interviews and manage support content for your blog.
Remote MCP server for supportsheep: run AI interviews and manage support content for your blog.
Related MCP Servers
- AlicenseNot gradedqualityCmaintenanceLocal-first MCP server that provides project context, verification gates, and structured tools for coding agents to discover knowledge, run diagnostics, and execute allowlisted commands within a repository.40 npmMIT
- FlicenseBqualityDmaintenanceProvides a minimal MCP server with safe tools for AI assistants to search project files, read files, search local documentation, and run pre-approved checks within a single project directory.4-
- AlicenseAqualityBmaintenanceLets ChatGPT, Claude Desktop, Claude Code, and other MCP clients run an allowlisted set of local commands and read files from a chosen directory, so an assistant can analyse a codebase or trigger local generators (images, sprites, assets) without metered API tokens. The server is read-only and deny-by-default by default, with no shell, path containment, opt-in writes, and required auth when exposed over HTTP.34MIT
- AlicenseNot gradedqualityBmaintenanceExposes MCP tools that give desktop agents durable two-tier memory with recall and graph traversal, project code intelligence (symbol lookup, impact walking, notes search), and web search with page-to-Markdown conversion. It also runs a local model gateway that routes Anthropic- and OpenAI-shaped traffic across providers with usage accounting.MIT