local-forge
Provides an adapter for connecting to Ollama inference servers, sending prompts to locally hosted models and routing requests based on deterministic prompt classification, with runtime token and cost-comparison tracking. Defaults to a loopback endpoint but can be pointed at a caller-supplied HTTP(S) endpoint.
This server does not currently have a deployable release on Glama. Follow the repository's setup instructions to install and configure it for your MCP client, including any required credentials and local dependencies.
Local Agent Forge
Local Agent Forge is a pre-release, source-available toolkit for probing local inference servers, classifying prompts with a deterministic heuristic, exposing experimental MCP tools, and recording runtime cost comparisons.
This repository is not a published or production-approved release. As checked on 2026-08-30, neither@nymrel/local-forge on npm nor local-agent-forge on PyPI exists. Use the source checkout for evaluation. Do not treat the package names, version, examples, or workflow definitions as evidence of a registry release, configured provider account, deployed service, available local model, customer activation, privacy certification, cost reduction, or revenue.
Choose a starting point
Your goal | First step | What to expect |
Understand whether this fits your work, without coding | Read current boundaries and the security model. | This is a toolkit for an existing local inference setup, with no hosted app or guided model installation. You can describe your goal in a question before installing anything. |
Build or troubleshoot an integration | Follow source evaluation, then run | The CLI probes the five default local adapter endpoints. If a probe fails, check that its inference server is running at that default URL; a custom SDK endpoint needs an adapter-specific health check before sending a prompt. |
Connect an agent | Build the source checkout, then configure a compatible stdio host to launch | Start with tool discovery; use the returned input schemas. A routing decision is advisory, and a host may retain tool arguments. Do not infer execution or spending authority from a |
Related MCP server: LocalLama MCP Server
Implemented surfaces
TypeScript adapters for Ollama, vLLM, LM Studio, ComfyUI, and Whisper-compatible HTTP servers.
A Python adapter surface for the same five endpoint families.
Deterministic prompt classification and routing-decision objects.
A local CLI, an experimental JSON-RPC/MCP server, and a loopback-only HTTP proxy.
Runtime token and cost-comparison ledgers.
Node and Python unit contracts plus deterministic npm and Python distribution validators.
The adapters default to loopback endpoints. A caller can supply a remote HTTP(S) endpoint, so local execution and data residency are never inferred from the adapter name. Custom endpoints are validated as credential-free HTTP(S) base URLs before requests are constructed.
Current boundaries
The
0.85routing threshold is a hand-authored heuristic, not a benchmarked accuracy guarantee or a service-level objective.A
CLOUDrouting decision is advisory metadata. It does not configure credentials, authorize spend, or execute a cloud-provider request.Model identifiers in the current router are defaults and compatibility labels, not proof that a model is installed, current, licensed, or reachable.
Cost tables are illustrative comparison inputs. They may become stale and omit hardware, electricity, operations, provider discounts, and other real costs. Validate pricing and measured usage before relying on a comparison.
The MCP implementation exercises a bounded protocol surface and reports protocol version
2024-11-05; it is not an independent compatibility certification.The test suites prove the checked contracts, not full behavioral parity, exhaustive security, or a coverage percentage.
Supported development floor
Runtime | Policy |
Node.js |
|
npm |
|
Python |
|
TypeScript | Repository compiler pin: |
Evaluate from source
Node.js
npm ci --ignore-scripts --no-audit --no-fund
npm run check
node .\bin\local-forge.js helpUseful source-checkout commands:
node .\bin\local-forge.js health
node .\bin\local-forge.js route "Summarize this function"
node .\bin\local-forge.js start --port 4000
node .\bin\local-forge.js mcpThe proxy listens on 127.0.0.1 by default. Stop it when evaluation is complete.
Python
python -m unittest discover -s tests -p test_*.py
python -m build --outdir python-dist
python scripts\check_python_package.py python-dist
python -m pip install --force-reinstall .\python-dist\*.whl
local-forge-py --helpThese are local source-build instructions. They are not registry installation instructions.
Security model
The HTTP proxy binds to IPv4 loopback only.
Default adapter endpoints use
127.0.0.1.Custom adapter endpoints must be absolute
http:orhttps:URLs without embedded credentials, query strings, or fragments.URL validation is a syntax boundary, not a trust decision or a complete SSRF defense. Treat every configured endpoint as a data recipient and apply host, network, and egress policy appropriate to the deployment.
Local inference servers can have their own authentication, logging, persistence, plugin, model, and network behavior. Audit them separately.
MCP clients can invoke tools with sensitive prompts. Review client permissions and tool exposure before enabling the server.
See SECURITY.md for reporting and supported-boundary details.
Package and release policy
npm run check:package proves the npm archive allowlist. python scripts/check_python_package.py python-dist proves the wheel and source-archive contracts. Publication remains manual, existing-tag-only, main-ancestry-gated, checksum-producing, attested, and OIDC-only in the workflow definition.
Those controls do not prove that npm or PyPI trusted publishing, GitHub environments, repository Actions, ownership, or release approval are configured. RELEASE_READINESS.md is the release gate.
License
MIT. Copyright 2026 Nymrel / JalenBuilds LLC.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
This server cannot be deployed
Maintenance
Related MCP Connectors
Agent FinOps Spend Guard, Deterministic Budget Passport, LLM Token Cost Router & VIP Discounts.
Cuts coding agent token burn up to 72% continuously via contradiction-free memory & verification.
Pay-per-call ($0.01 USDC) model-routing for AI coding agents: which LLM to call, cost vs. quality.
Agent GPU compute scheduling: compare billing, approve budgets, run tasks and persist results.
Related MCP Servers
- AlicenseAqualityDmaintenanceOptimizes token costs by intelligently delegating low-complexity tasks to local LLMs via LiteLLM, enabling cost-effective development workflows.31MIT
- AlicenseNot gradedqualityBmaintenanceRoutes coding tasks to local and remote LLMs with intelligent cost-quality optimization, supports benchmarking and code search.43ISC
- FlicenseNot gradedqualityDmaintenanceEnables hybrid agentic coding by delegating token-heavy implementation, editing, and test-debug loops to local open-source models while frontier cloud models handle architecture and review, reducing premium API token usage and keeping code private.1-
- AlicenseAqualityBmaintenanceEnables Claude Code to delegate agent tasks to local LLMs via llama.cpp or vLLM servers on your own hardware, reducing API usage while maintaining full agent capabilities.1122MIT