Consult7 is a Model Context Protocol (MCP) server that enables AI agents to consult large context window models via OpenRouter for analyzing extensive file collections - entire codebases, document repositories, or mixed content that exceed the current agent's context limits.
Why Consult7?
MCP server to consult a language model with large context size
Estrelas
296
Ferramentas
—
Disponibilidade
—
Consult7 enables any MCP-compatible agent to offload file analysis to large context models (up to 2M tokens). Useful when:
Agent's current context is full
Task requires specialized model capabilities
Need to analyze large codebases in a single query
Want to compare results from different models
"For Claude Code users, Consult7 is a game changer."
How it works
Consult7 collects files from the specific paths you provide (with optional wildcards in filenames), assembles them into a single context, and sends them to a large context window model along with your query. The result is directly fed back to the agent you are working with.
ULTRA = Run GPTT, GROT, and FABT in parallel (3 frontier models)
FUSE = Fusion: a frontier panel deliberates and a judge synthesizes, in one call
These mnemonics make it easy to reference model+mode combinations in your queries.
Note on Fable 5.1.anthropic/claude-fable-5.1 is Anthropic's most capable model but priced at a premium (~2× Opus 4.8). It does not replace Opus 4.8 as the everyday Claude choice for single calls. Since v3.11.0 it holds the Anthropic seat in the ULTRA panel (which is meant for hard questions anyway). Unlike Opus 4.8 (adaptive thinking only), OpenRouter honors Fable's effort scale, so mid/think map to effort=high/effort=xhigh.
Featured: Fusion (multi-model analysis)
Consult7 supports OpenRouter's Fusion (openrouter/fusion) — a single call where a panel of frontier models (Opus, GPT, Gemini Pro) answers your query in parallel and a judge model synthesizes their responses into one answer. Reach for it on hard questions where multiple perspectives help and the cost of being wrong outweighs a few extra completions.
Context: 128K — smaller than the 1M–2M single models, so it's best for hard questions on moderate input, not giant file bundles.
Mode → research depth:fast / mid / think map the panel's web-search/fetch budget to max_tool_calls of 2 / 8 / 16.
Mnemonic:FUSE = openrouter/fusion.
Trivial prompts answer directly (no panel); the panel fires only when the question warrants deliberation. Fusion is billed per panel run, so it costs more than a single-model call.
Installation
Claude Code
Simply run:
claude mcp add -s user consult7 uvx -- consult7 your-openrouter-api-key
gemf = google/gemini-3.8-flash + fast (Gemini 3.8 Flash, fast and cheap)
ULTRA = call GPTT, GROT, and FABT IN PARALLEL (3 frontier models for maximum insight)
FUSE = openrouter/fusion (one call: a frontier panel deliberates, a judge synthesizes; mode sets web-research depth)
You can use any OpenRouter model ID (e.g., deepseek/deepseek-r1-0528). See the full model list. File size limits are automatically calculated based on each model's context window.
Performance Modes
fast: No reasoning requested - quick answers, simple tasks. GPT-6 Astra, Grok 4.7 and Fable reason by design, so on them fast means their own default level (billed), shown in the footer as reasoning: model default. Gemini 4 Argon, 3.8 Flash and 3.1 Pro get an explicit effort in every mode: fast = low, mid = medium, think = high
think: Maximum reasoning - security audits, complex refactoring
File Specification Rules
Absolute paths only: /Users/john/project/src/*.py
Wildcards in filenames only: /Users/john/project/*.py (not in directory paths)
Extension required with wildcards: *.py not *
Mix files and patterns: ["/path/src/*.py", "/path/README.md", "/path/tests/*_test.py"]
Common patterns:
All Python files: /path/to/dir/*.py
Test files: /path/to/tests/*_test.py or /path/to/tests/test_*.py
Multiple extensions: ["/path/*.js", "/path/*.ts"]
Automatically ignored:__pycache__, .env, secrets.py, .DS_Store, .git, node_modules. Wildcards skip them; naming one explicitly is an error.
Fail fast: a relative or missing path, a directory, a wildcard that matches nothing, an explicitly named ignored file, a binary file (NUL bytes in the first 8 KB; .pdf excepted), or files over the model's size budget fail the whole call before anything is sent to the model (no cost). The error lists every problem.
PDFs (v3.12.0):.pdf files are sent to the model as native documents through OpenRouter (the model sees the pages, including figures), next to the text files in the same call. Models that do not list file input on OpenRouter are rejected before the call. Typeset math and handwriting are read from the rendered page, not from the PDF text layer (verified on equations from a LaTeX paper and on a scanned handwritten submission). A page costs roughly 500 input tokens on Gemini, 1,500 on GPT and 2,500-3,000 on Claude and Grok, so for a long PDF on a premium model, convert it to text first (pdftotext) and send the .txt. PDFs have their own 32 MB wire-size budget; Anthropic models take at most 600 pages per request. PDFs are always forwarded natively, never through OpenRouter's OCR parsers, so no PDF leaves the chosen model's provider (this holds with zdr too).
Size budget: one total for all files together, no per-file limit: (model context − output reserve − reasoning reserve) × ~4 bytes per token, e.g. ~4 MB for 1M-context models, ~2 MB for Grok 4.7 (500K), ~8 MB for Grok 4.20 (2M). A second check on the estimated token count keeps a 10% safety margin. A size error states the budget and the total requested.
Tool Parameters
The consultation tool accepts the following parameters:
files (required): List of absolute file paths or patterns with wildcards in filenames only
query (required): Your question or instruction for the LLM to process the files
model (required): The LLM model to use (see Supported Models above)
mode (required): Performance mode - fast, mid, or think
output_file (optional): Absolute path to save the response to a file instead of returning it
The path is checked before the model is called, so a bad path costs nothing
If the file exists, it will be saved with _updated suffix (e.g., report.md → report_updated.md, then report_updated_1.md, ...)
When specified, returns "Result has been saved to /path/to/file" plus the metadata footer
Useful for generating reports, documentation, or analyses without flooding the agent's context
zdr (optional): Enable Zero Data Retention routing (default: false)
When true, routes only to endpoints with ZDR policy (prompts not retained by provider)
Not available: Claude Fable 5.1 and 5, GPT-5.6 Sol, Grok 4.20 (returns error)
Usage Examples
Via MCP in Claude Code
Claude Code will automatically use the tool with proper parameters:
{
"files": ["/Users/john/project/src/*.py"],
"query": "Explain the main architecture",
"model": "google/gemini-3.8-flash",
"mode": "fast"
}
Via Python API
from consult7.consultation import consultation_impl
result = await consultation_impl(
files=["/path/to/file.py"],
query="Explain this code",
model="google/gemini-3.8-flash",
mode="fast", # fast, mid, or think
provider="openrouter",
api_key="sk-or-v1-..."
)
Testing
# Test OpenRouter connection
uvx consult7 sk-or-v1-your-api-key --test
Uninstalling
To remove consult7 from Claude Code:
claude mcp remove consult7 -s user
Version History
v3.13.1
PDF budget stated up front: the tool description and the refusal now say about 24 MB of PDF files per call (the 32 MB limit is on the base64 payload) and that scans run 1-2 MB per page.
A mid-stream upstream timeout error says it is transient and that consult7 does not retry, so a caller knows to rerun.
The package and consultation_impl docstrings state that the MCP tool is the only supported interface; batch work is the calling agent's job (one tool call per item, output_file for the answers). consult7.__version__ now reports the real version.
v3.13.0
PDF input..pdf files in files go to the model as native documents (OpenRouter file content parts, native engine only, no OCR fallback). Mixed with text files in one call; page count and PDF count shown in the footer; token estimate per page by vendor; models without file input, oversized PDF payloads (32 MB base64) and more than 600 pages on Anthropic models fail before the call.
GPT-6.1 Terra (openai/gpt-6.1-terra) added as the cheap GPT option: 1M context, $1/$5 per million tokens, effort-based reasoning (mid = high, think = xhigh), no ZDR. GPT-6 Astra stays the default GPT.
v3.12.0
Gemini 4 Argon (google/gemini-4-argon) is the flagship Gemini and the gemt model. Gemini 3.8 Flash (google/gemini-3.8-flash) replaces Gemini 3 Flash for gemf. Both support ZDR.
Gemini reasoning levels fixed. Gemini 4 Argon, 3.8 Flash and 3.1 Pro reason even without a reasoning parameter, so the old mapping (fast sent nothing, mid sent effort low) made mid reason less than fast (about a third as much on 3.1 Pro). These models now get an explicit effort in every mode: fast = low, mid = medium, think = high.
Removed old Gemini models from the tuned list: Gemini 3 Flash, 3.1 Flash Lite, 3 Pro (no longer on OpenRouter), 2.5 Pro and 2.5 Flash. They still run as unknown models, without reasoning parameters.
consult7 --help prints usage instead of being taken as the API key. Works on Python 3.15 (dependencies relocked, unused anthropic dev dependency removed).
v3.11.2
No per-file size limit. The files share one total budget, so a single large file can use all of it (before, one file was capped at half the budget). Size errors report the total requested and the budget, and token errors mention the 10% safety margin.
Binary files fail fast (NUL bytes in the first 8 KB) instead of being sent as garbage tokens.
Unknown model IDs: when a model is not in OpenRouter's model list, errors say so, name the assumed 128K context, and suggest the closest listed IDs. Variant IDs such as model:nitro use the base model's context size.
Footer: wall-clock time, the mode always shown ([fast] too), zdr shown when on. Both token estimates now use the same prompt text, and a rejected fast call no longer claims reasoning disabled.
v3.11.1
Fail fast before any paid call. A missing file in a list, a wildcard that matches nothing, an explicitly named ignored file (.env, secrets.py, ...), or files over the model's per-file or total size limit now fail the call with a clear error before anything is sent. Before, these were reported only inside the prompt (or dropped), and the paid call went through on partial input.
output_file is checked before the call. A relative or unwritable path fails at no cost. If saving still fails after the call, the response is returned instead of being lost.
Errors set isError=true in the MCP result, so clients and wrappers can detect failures without parsing text.
Upstream error messages no longer include the OpenRouter account user_id.
Footer: 1 file (not 1 files). On fast, models that always reason (GPT-6 Astra, Grok 4.7, Fable) show reasoning: model default.
Shorter tool description (under 2,000 characters, file rules first). Claude Code truncated the old one before the file rules.
v3.11.0
New ULTRA panel: GPTT + GROT + FABT (3 models in parallel). Gemini 3.1 Pro leaves the panel (gemt and all Gemini models stay available); Opus 4.8 (oput/opuf) stays available but its ULTRA seat goes to Fable.
gptt → GPT-6 Astra (openai/gpt-6-astra, 1M context, $10/$50 per M) — also the model used by consult7 <key> --test. Reasoning is mandatory on Astra, so mid/think now map to effort=high/effort=xhigh (GPT-5.6 Sol used medium/high). ZDR supported.
grot → Grok 4.7 (x-ai/grok-4.7, 500K context; grok 4.6 had been the default since v3.10.0). Same effort mapping (high/xhigh). ZDR supported.
fabt/fabm → Claude Fable 5.1 (anthropic/claude-fable-5.1, 1M context). Same effort mapping. ZDR not supported.
Superseded IDs (openai/gpt-5.6-sol, x-ai/grok-4.6, anthropic/claude-fable-5) keep working with their previous settings.
v3.9.0
New default GPT: GPT-5.6 Sol (openai/gpt-5.6-sol) — the latest top-tier GPT, ~1M context / 128K output, effort-based reasoning (mid → effort=medium, think → effort=high). Replaces GPT-5.5 as the gptt default; GPT-5.5 stays available as a legacy model. ZDR is not supported on GPT-5.6 Sol (GPT-5.5 still is).
Grok 4.5 not added:x-ai/grok-4.5 is region-restricted on OpenRouter (returns a 403 "not available in your region") and could not be verified against the real API, so it was not integrated. Grok 4.20 remains the grot default.
v3.8.0
Added Claude Fable 5 (anthropic/claude-fable-5) — Anthropic's most capable model, 1M context. Premium price (~2× Opus 4.8), so it's reserved for specifically hard problems and is not part of the ULTRA panel; it does not replace Opus 4.8 as the default Claude model. New mnemonics fabt (think) / fabm (mid). Unlike Opus 4.8 (adaptive thinking only), OpenRouter honors Fable's effort scale, so mid/think map to effort=high/effort=xhigh (max intentionally not exposed — it tends to overthink at ~2× token cost). ZDR not supported (Fable requires 30-day retention).
Response-length prompt tuned: the system prompt now asks the model to match answer length to the task (thorough when the question needs depth, concise otherwise) instead of a blunt "be concise".
v3.7.1
Surface mid-stream API errors: when OpenRouter sends an error as a streaming data chunk (after the initial 200), the call now returns that error message instead of a misleading "No content received".
v3.7.0
Added Fusion (openrouter/fusion) — a multi-model panel plus a judge in one call; mode maps to web-research depth (fast/mid/think → max_tool_calls 2/8/16). New FUSE mnemonic.
Upgraded Claude Opus 4.7 → 4.8 (1M context, adaptive thinking); oput/opuf now point to 4.8, and 4.7 is kept as a legacy ID.
The response footer now reports the call cost in USD (from OpenRouter usage accounting), e.g. cost: $0.0923.
v3.6.1
Toggle-reasoning footer now distinguishes mid vs think for adaptive models (Opus, Grok)
Friendlier error message when a model has no Zero Data Retention endpoint
output_file return now includes the metadata footer so callers can verify what ran
v3.6.0
Upgraded models: GPT-5.5, Claude Opus 4.7, Grok 4.20
Claude Opus 4.7 (1M context) uses adaptive thinking — reasoning.enabled=true
Ainda não há lista de ferramentas. Servidores remotos recebem uma após a primeira verificação de saúde.
Perguntas frequentes
Perguntas
Como instalo Consult7?
Copie nesta página a configuração stdio para Cursor, Claude Desktop, Claude Code ou VS Code. Ela usa a instalação do publicador, não um proxy do PlugMCP.
Consult7 precisa de uma chave de API?
Este registro não lista nenhuma variável de ambiente obrigatória.
Qual transporte Consult7 usa?
Consult7 roda localmente via stdio.
Quando Consult7 foi verificado pela última vez?
Última verificação: 2026-10-28. MCP server to consult a language model with large context size