@inferlane/mcp

@inferlane/mcp

The cost intelligence layer for AI agents. 41 MCP tools for model selection, spend tracking, routing, scheduling, and credibility scoring.

Install

npx -y @inferlane/mcp

Or add to your MCP config:

{
  "mcpServers": {
    "inferlane": {
      "command": "npx",
      "args": ["-y", "@inferlane/mcp"]
    }
  }
}

Modes

Set INFERLANE_MODE to control how many tools register (and how much context they cost):

Mode	Tools	Context	What you get
(unset) full	all	~27K tokens	Everything below
core	8	~6K tokens	Cost intelligence essentials: pick_model, log_request, session_cost, assess_routing, rate_recommendation, get_cost_comparison, suggest_savings, get_model_pricing
firewall	10	~7K tokens	The spend-firewall profile: fuel_gauge, platform_budget, platform_spend, check_dispatch, log_request, session_cost, pick_model, get_model_pricing, ping, register_webhook

Firewall mode is built for agent fleets: check_dispatch pre-flights every model call against the local budget (INFERLANE_BUDGET_TOTAL) first and the platform budget (INFERLANE_API_KEY) second, returning a structured ALLOW/DENY JSON decision with the estimated cost, remaining budgets, and a cheaper-model suggestion on deny. With no budget configured it answers ALLOW with an advisory note — it never fakes a denial. Budget denials from pick_model and route_via_platform also carry a machine-parseable {"error": "BUDGET_EXCEEDED", ...} JSON block alongside the human-readable text.

{
  "mcpServers": {
    "inferlane": {
      "command": "npx",
      "args": ["-y", "@inferlane/mcp"],
      "env": { "INFERLANE_MODE": "firewall", "INFERLANE_BUDGET_TOTAL": "25" }
    }
  }
}

Tools

Cost Intelligence

• pick_model — Choose the optimal model for any task
• check_dispatch — Pre-flight a dispatch against local + platform budgets (structured ALLOW/DENY)
• session_cost — Track session spend in real-time
• log_request — Log API calls for cost tracking
• get_model_pricing — Look up pricing across providers
• get_cost_comparison — Compare costs across models
• suggest_savings — Get cost optimization recommendations

Routing & Triage

• triage — Auto-classify prompts by complexity, urgency, and cost
• triage_settings — Configure routing preferences
• assess_routing — Evaluate local vs cloud routing
• route_to_cloud — Report routing decisions for credibility

Dispatch & Scheduling

• dispatch — Send prompts to best available provider
• dispatch_chain — Multi-provider sequential chains
• dispatch_status — Check async task status
• schedule_prompt — Schedule prompts for later
• create_chain — Create multi-step chains
• list_scheduled / cancel_scheduled / chain_status

Agent Intelligence

• agent_status — Traffic light status (green/amber/red/blue)
• set_agent_status — Manual status override
• set_lifecycle_phase — Track coding→testing→CI→deploy phases
• lifecycle_report — Cost-per-phase breakdown
• token_tachometer — Real-time token velocity
• state_of_compute — Full compute market report

Credibility

• credibility_profile — View agent credibility score
• credibility_leaderboard — Compete with other agents
• rate_recommendation — Rate model quality
• model_ratings — View community ratings
• improvement_cycle — Run quality analysis

Platform (requires INFERLANE_API_KEY)

• cost_savings — View savings from smart routing
• check_promotions — Active provider promotions
• platform_spend / platform_budget — Platform billing
• route_via_platform — Route through InferLane platform
• session_history — Cross-provider session tracking

Environment Variables

Variable	Required	Description
INFERLANE_API_KEY	No	Enables platform features (dispatch, scheduling, savings)
INFERLANE_MODE	No	core or firewall to register a trimmed tool profile (see Modes)
INFERLANE_EVENTS_PORT	No	Enable SSE event stream on this port
INFERLANE_BUDGET_TOTAL	No	Monthly budget cap in USD — enforced by check_dispatch, pick_model, route_via_platform
OLLAMA_HOST	No	Local Ollama endpoint for routing assessment

How It Works

Install once. Every AI agent session becomes cost-aware:

• Before a task — pick_model recommends the cheapest viable model
• After a task — log_request tracks what was spent
• Over time — credibility score builds, recommendations improve

No configuration needed for basic cost tracking. Add INFERLANE_API_KEY for platform features.

Data & Privacy

This package reads some files on your machine to do its job. Everything below stays local and is never transmitted — the watcher does zero network I/O — unless you set INFERLANE_API_KEY to opt into cloud sync. Without that key, InferLane has no visibility into any of it.

What it reads locally (read-only, on by default):

• Your Claude Code transcripts. On startup the server reads Claude Code's JSONL transcript files at ~/.claude/projects/**/*.jsonl to pull the per-message token-usage blocks (message.usage: input / output / cache token counts + model name) so the fuel gauge reflects actual spend with no API key and no manual logging. It only reads these files — it never modifies them — and it does not send their contents anywhere unless you enable cloud sync. This watcher is on by default; opt out by setting INFERLANE_NO_CC_WATCH=1 (it also stays inactive if ~/.claude/projects doesn't exist).

What it writes locally:

• ~/.inferlane/state.db — a SQLite database holding your request log, model ratings, credibility scores, and budget config so they survive restarts. It lives only on your machine. If SQLite can't initialize, the server falls back to in-memory mode and writes nothing.

Optional persistence installer (separate tool, opt-in):

A separate, related InferLane tool — the persistence installer shipped as the inferlane-persist / inferlane-unpersist commands in the @inferlane/mcp-server package — can make the cost-awareness guidance persist across agent sessions. It is opt-in and prompts for confirmation before changing anything: a [y/N] prompt that defaults to No, so an unattended or piped run installs nothing unless you pass an explicit --yes flag. When you consent, it modifies only:

• ~/.claude/CLAUDE.md (and, if you choose them, the equivalent per-project instruction files it detects: project CLAUDE.md, .cursorrules, .github/copilot-instructions.md, .gemini/styleguide.md, CONVENTIONS.md, AGENTS.md) — it inserts a small, clearly marker-delimited activation block (<!-- BEGIN INFERLANE vN --> … <!-- END INFERLANE vN -->). Your existing content is left intact.
• ~/.claude/settings.json — with a separate explicit confirmation, it merges a single PreToolUse spend-guard hook without clobbering any other hooks or keys.

Every file that installer touches and every hook it wires is recorded in a manifest at ~/.inferlane/persist-manifest.json so the uninstaller can remove exactly what was added and nothing else.

This mirrors §1.3 ("Local-only data") of the InferLane Privacy Policy.

License

Apache-2.0