Name: Model Finder
Author: indigloo

Model Finder MCP

Find the cheapest AI model that passes your tests: audit repo models, build evals, routing, sizing

Grade A Local · stdio No auth needed 10 tools v1.3.6 · 3 Nov 2026 MIT

5.0(1) by indiglooPyPI 647 Runs on your machine, nothing hosted

100/100

trust score

downloads in 30d

local

runs on your machine

tools declared

What it does

Audits the models your repo uses, drafts evals from your prompts, ranks results, generates cheapest-first routing and sizes local models. Runs on your machine; nothing is sent to Indigloo.

Quickstart

# 1 — add it to Claude Code
claude mcp add --scope user model-finder-mcp '--' uvx model-finder-mcp

# 2 — ask your agent for something these tools do
#   modelfinder_audit_models, modelfinder_draft_eval_examples, modelfinder_create_eval_config and 7 more

Model Finder costs nothing on mcprush: there is no plan to choose here, no cap to set and nothing mcprush can bill you for.

Publisher

indigloo1 server

1 category

All 1 server

Tool surface

What the model actually sees. Descriptions are diffed on every release — see version history.

Only tools that write or reach the network

List the AI model IDs (quoted literals) and AI SDKs a repository uses, with file:line locations, provider family, route hints (Bedrock, Vertex AI, Azure, OpenRouter), usage counts and whether each ID is hard-coded or an environment-variable default. Use first, to see which models a project depends on before comparing or replacing them; use modelfinder_draft_eval_examples to find test inputs instea

Parameter	Type	Required	Notes
path	string	no	Directory inside the workspace, relative to the project root (e.g. "services/api"). Defaults to the project root. If a call fails because the project is unknown, pass the absolute path of the project
limit	integer	no	Items per page (1-200, default 50).
offset	integer	no	Items to skip; use next_offset from the previous page.
provider_family	string	no	Only models from this family, e.g. "anthropic".
hard_coded_only	boolean	no	Only IDs not read from configuration or environment.
response_format	string	no	markdown (default) for a readable summary, json for the full machine-readable result.

Draft eval examples from the repository's prompt files and fixtures (prompts/, evals/, fixtures/, *.prompt.md, and JSON/JSONL with input, prompt or chat messages). Use after modelfinder_audit_models when the user has no examples yet. Drafts have empty checks: agree checks with the user, then pass them to modelfinder_create_eval_config. Read-only and local: no network calls, no files written. Drops

Parameter	Type	Required	Notes
path	string	no	Directory inside the workspace, relative to the project root (e.g. "services/api"). Defaults to the project root. If a call fails because the project is unknown, pass the absolute path of the project
limit	integer	no	Items per page (1-200, default 30).
offset	integer	no	Items to skip; use next_offset from the previous page.
response_format	string	no	markdown (default) for a readable summary, json for the full machine-readable result.

Build an eval config from candidate models, examples with checks, and settings, then validate it and screen the examples for personal data and secrets. Use once examples and checks are agreed; to check a config the user already has, use modelfinder_validate_eval_config instead. Read-only and local: no network calls, no files written. Returns the config for the user to save as JSON and run with eva

Parameter	Type	Required	Notes
task	string	no	One-line description of what the model must do.
candidates	array	yes	1-10 models to compare.
examples	array	yes	Inputs with the checks each reply must pass.
settings	object	no	max_output_tokens (default 1024), timeout_s (default 60), temperature, system (system prompt).
pass_threshold	number	no	Share of calls that must pass every check (0-1).
repeats	integer	no	Runs per example (3+ for noisy tasks).
response_format	string	no	markdown (default) for a readable summary, json for the full machine-readable result.

Check an existing eval config before a paid run: schema, no keys in the config, HTTPS endpoints, no template placeholders, valid checks. Use on a config the user wrote or edited; to build a new one use modelfinder_create_eval_config. Read-only and local: no network calls, no files written. Returns valid, problems, the number of API calls the run would make, required environment variable names, unp

Parameter	Type	Required	Notes
config	object	yes	The eval config object (as saved for eval_models.py).
response_format	string	no	markdown (default) for a readable summary, json for the full machine-readable result.

Run deterministic checks (contains, not_contains, exact, regex, json_keys, max_chars) on one model reply and report which failed. Use it to design and debug eval checks or routing guards against a sample reply before spending money on a run; to score a whole eval run use modelfinder_rank_eval_results. Read-only and local: no network calls, no files written. Returns passed, failures (check and reas

Parameter	Type	Required	Notes
text	string	yes	The model reply to test.
checks	array	yes	Deterministic checks; all must pass. Types: contains, not_contains, exact (value, ignore_case?), regex (pattern), json_keys (keys), max_chars (value). Example: [{"type": "json_keys", "keys": ["categor
response_format	string	no	markdown (default) for a readable summary, json for the full machine-readable result.

Rank the models in a finished eval run (the results.json object written by eval_models.py) and recommend the cheapest one that met the pass threshold, ranked by cost per passing call. Use after an eval run; to test a single reply use modelfinder_check_output, and to turn the ranking into a fallback chain use modelfinder_generate_routing. Read-only and local: no network calls, no files written. Ret

Parameter	Type	Required	Notes
results	object	yes	The full results.json object written by eval_models.py.
response_format	string	no	markdown (default) for a readable summary, json for the full machine-readable result.

Turn a finished eval run (results.json object) into a cheapest-first routing.json, with quality guards and a personal-data policy, plus a LiteLLM proxy config whose fallbacks fire on API errors only (it does not enforce the quality guards or personal-data policy). Use after modelfinder_rank_eval_results when the user wants a fallback chain. Read-only and local: no network calls, no files written.

Parameter	Type	Required	Notes
results	object	yes	The full results.json object written by eval_models.py.
guards	array	no	Checks applied to every production reply; a failure escalates to the next model.
chain	array	no	Explicit model order (candidate ids), overriding cost order.
include_unknown_cost	boolean	no	Append passing models whose price is unknown.
pii_policy	string	no	block: never send prompts with personal data or secrets; warn: send and report; off: no check.
response_format	string	no	markdown (default) for a readable summary, json for the full machine-readable result.

Estimate monthly token cost from requests per month, average tokens per request and USD-per-million prices, with optional cached-input share and pass rate (cost per accepted result, counting retries). Use for planning or to compare models before running an eval. Read-only and local: no network calls, no files written. Prices are not looked up: pass them from the provider's official pricing page. R

Parameter	Type	Required	Notes
requests_per_month	number	yes	e.g. 100000
input_tokens	number	yes	Average input tokens per request.
output_tokens	number	yes	Average output tokens per request (include reasoning tokens).
input_price_per_million	number	yes	USD per 1M input tokens.
output_price_per_million	number	yes	USD per 1M output tokens.
cached_fraction	number	no	Share of input tokens billed at the cached rate (0-1).
cached_input_price_per_million	number	no	Required when cached_fraction > 0.
pass_rate	number	no	Optional eval pass rate; failures are assumed retried.
response_format	string	no	markdown (default) for a readable summary, json for the full machine-readable result.

Estimate the memory needed to run an open-weight model locally: weights at a given precision plus KV cache for a context length and concurrency (standard MHA/GQA attention only), with headroom against available memory. Use when choosing a local model or GPU. Give all four of layers, kv_heads, head_dim and context to include the KV cache; without them only weights are counted. Read-only and local:

Parameter	Type	Required	Notes
params_b	number	yes	Total stored parameters in billions (for MoE, all experts).
bits	number	yes	Effective bits per weight: 16, 8, about 4.5 for common 4-bit formats.
overhead_pct	number	no	Allowance on top, in percent.
layers	integer	no	num_hidden_layers from the model config.
kv_heads	integer	no	num_key_value_heads (or num_attention_heads if absent).
head_dim	integer	no	head_dim, or hidden_size / num_attention_heads.
context	integer	no	Context tokens per sequence actually used.
concurrency	integer	no	Sequences served at once.
kv_bits	number	no	KV cache precision in bits.
available_gib	number	no	GPU or unified memory available, in GiB.
response_format	string	no	markdown (default) for a readable summary, json for the full machine-readable result.

Detect Aadhaar (Verhoeff checked), PAN, payment cards (Luhn checked), email addresses, UPI IDs, Indian mobile numbers and API or private keys in text. Use before sending prompts or eval examples to a provider; modelfinder_create_eval_config already runs this on its examples. Read-only and local: no network calls, no files written. Returns categories, counts and a clean flag, never the matched valu

Parameter	Type	Required	Notes
text	string	yes	Text to scan, up to 1,000,000 characters.
response_format	string	no	markdown (default) for a readable summary, json for the full machine-readable result.

Free

$0 forever

Every tool, no call limit
No card, and mcprush charges nothing for it
Source published under MIT
Runs on your machine — nothing of it reaches our gateway
Nothing to cap, because mcprush bills nothing

Nothing here is billable: mcprush charges nothing to install Model Finder or to call it, at any volume.

ATrust score 100 / 100

Release 1.3.6 · scanned 3 Nov 2026

Two independent axes, because powerful and malicious are different questions. The grade is threat only. The capability level is blast radius, and it is never a penalty on the grade — it is priced as one subtract-only term in the score, where you can see it.

Threat score100

Client score100

Findings0 threat · 0 capability

Capabilityminimal

Coveragesource

Verificationsource

Surface digestc115f654c9e6837ff22eeee94569be5764500cacaebe90ed2ccaf6c6e47a235d

Score vectorthreat 100 → client 100

Term	Level	What it prices	Points
capability-exposure	minimal	capability blast radius (minimal) — client exposure if the model is manipulated	−0
verification-discount	source	publisher verification (provenance) — cryptographic build provenance ties the artifact to its source	−0
coverage-honesty	source	inspection depth (source) — how much of the target the scan could see	−0

score = clamp(round(threatScore − E_cap − E_ver − E_cov), 0, 100). No model, no opinion, no network call at scoring time: the same methodology version against the same bytes always produces this number, and every line above is a field of the report rather than a summary of it.

What the scan could actually read

A grade is only as meaningful as its coverage, so the scanner publishes its own depth before it publishes its result.

Tool surfacenot available

Implementation sourceread

Package metadataread

Live transportnot available

No tools were enumerated, so prompt-injection, capability and toxic-flow analysis had no tool surface to inspect. To grade a package’s real runtime tools, scan the running server: --command "npx -y <package>".

Capability — what it could do if the model were manipulated

Tags derived from each tool’s schema and the implementation, not from what the tool calls itself. minimal is the level these add up to.

Tool	Capability tags	Why the tag was assigned
No per-tool rows for this release — see the note below.

No tool surface was enumerated, so there is nothing to attribute a capability to.

Toxic-flow graph

The lethal trifecta, checked as a graph rather than as a checklist: untrusted input, a sensitive source and an external sink have to meet before there is a path worth worrying about.

The public result for this release does not print the flow graph, so there is nothing to show here. That is not the same as "no paths were found": what the scan did read is above, under coverage.

Supply chain and provenance

Integrityfirst-seen

Surface digestc115f654c9e6837f

Scanned2026-11-03

This is the first scan of this surface here, so there is nothing yet to compare it against.

Every result on this tab comes from one deterministic pass over the published package — offline, rule by rule, and auditable line by line above. Same methodology version, same bytes, same score.

Where are you running it?

Tool surface

Release history

Model Finder MCP

What it does

Quickstart

Where are you running it?

Tool surface

What counts against your monthly calls

What the scan could actually read

Capability — what it could do if the model were manipulated

Toxic-flow graph

Supply chain and provenance

Release history

See also