Available Tools - Complete Schemas

Log reasoning thoughts in a dedicated scratchpad. Provides structured thinking space for step-by-step problem solving. Foundation of the 54% performance improvement from Anthropic research.

Multi-model collaborative reasoning with 10+ specialized modes (deep-reasoning, code-brainstorm, architecture-debate, etc.). Coordinate different AI models to solve problems together. Also available as 'tachi' alias.

Sequential thinking with multi-model execution, context distillation, and auto-judgment. Chain reasoning across models with smart context management.

Setup diagnostic: which API keys are detected, which tools are available vs hidden (and why), the active profile, and a suggested first step. Zero-cost, needs no API key. Call it when tools seem missing.

Live web search using Grok 4.3 with advanced filtering. Configure sources (web/news/x/rss), domain restrictions, recency filters, and max results. IMPORTANT: Costs $5 per 1000 sources searched.

Deep logical reasoning with Grok using first principles thinking. Break down complex problems to fundamental truths and build solutions from the ground up.

Code analysis, optimization, debugging, review, and refactoring with Grok 4.3. Specialized for technical tasks requiring deep code understanding.

Deep debugging assistance with context-aware analysis. Grok examines errors, code, and context to find root causes and suggest fixes.

Systematic bug triage (Grok 4.3): RANKED root-cause hypotheses with likelihoods, the cheapest discriminating check for each, and the minimal fix for the top candidate. Unlike grok_debug's single-shot answer, this narrows a hypothesis space like real debugging does.

System architecture and design for small/medium/large/enterprise scale. Design scalable systems with proper constraints and requirements.

Creative brainstorming with Grok 4.3 for deep creative thinking. Generate innovative ideas with configurable constraints and quantity.

Web search with real-time information using Perplexity Sonar. Filter by recency (hour/day/week/month/year) and search domains (general/academic/news/social).

Complex reasoning with Perplexity Sonar Reasoning Pro. Analytical, creative, systematic, or comparative approaches for deep problem solving.

Generate ASCII visualization of workflow structure showing steps, tools, models, parallel execution, and conditional logic. Perfect for understanding workflow composition.

Start a workflow in step-by-step streaming mode. Execute one step at a time with session management for long-running workflows. Returns session ID for continuing execution.

Continue executing the next step of a streaming workflow session. Use the session ID returned from workflow_start or previous continue_workflow calls.

Check the progress and status of a running streaming workflow session. Shows completed steps, current step, remaining steps, and latest output.

Execute multi-step tool sequences from YAML/JSON files. Variable interpolation, parallel execution, unlimited steps. Build complex automated reasoning processes.

List all available workflow templates and custom workflows in your project directory.

Create custom workflows from templates (code-review/brainstorm/debug/research/custom). Define your own tool sequences with YAML/JSON.

Validate workflow YAML/JSON content for correctness. Checks syntax, interpolation references, tool names, and circular dependencies. Returns detailed error messages with fix suggestions.

Validate a workflow file from filesystem. Same validation as validate_workflow but reads from file path. Checks syntax, tool names, variable references, and dependencies.

Creative brainstorming with GPT-5.5 series. Choose model variant, reasoning effort (none/low/medium/high/xhigh), and style.

Mathematical and scientific reasoning using GPT-5.5. Excels at logical analysis, mathematical proofs, scientific reasoning, and analytical problem-solving.

Comprehensive code review with GPT-5.5. Focus on security, performance, readability, bugs, or best practices. Get detailed feedback and improvement suggestions.

Explain complex concepts at different levels (beginner/intermediate/expert) using various styles (technical/simple/analogy/visual). Perfect for learning and teaching.

Web search with OpenAI. Retrieve real-time information grounded in web search results using GPT models with search capabilities.

Collaborative ideation with Gemini. Multi-round brainstorming where Gemini builds upon previous thoughts to generate deeper insights.

Code quality, security, performance, and bug analysis with Gemini. Focus on specific aspects or get general comprehensive analysis.

Text analysis for sentiment, summary, entities, key points, or general insights. Extract structured information from unstructured text.

Qwen3-Coder-Next (80B/3B MoE, 262K context, SWE-Bench >70%) for code generation, review, optimization, debugging, refactoring, and explanation. Cost-efficient at $0.14/$0.80 per M tokens.

Qwen3-Competitive for competitive programming, algorithmic challenges, and LeetCode-style problems.

Generate runnable tests with a coding-specialized model (Qwen3-Coder-Next). Enumerates edge cases and failure modes first, then emits complete test code matching your framework and existing conventions.

Heavy reasoning with Qwen3-235B-Thinking (235B MoE, 98% HMMT math competition). Excels at complex multi-step reasoning, mathematical proofs, and analytical problem-solving.

Algorithm analysis with Qwen3-235B-Thinking (235B MoE, LiveCodeBench 91.4, HMMT 98%). O(1)-first optimization, complexity profiling, competitive programming patterns, and data structure selection.

Advanced agentic reasoning with Moonshot AI's Kimi K2.7-Code (1T MoE, 32B active, 256k context). Multimodal (vision/video) with Agent Swarm (100 sub-agents). Excels at long-horizon reasoning, multi-step analysis, and complex problem-solving. Tops SWE-Bench.

Task decomposition with Kimi K2.7-Code's Agent Swarm reasoning. Breaks complex tasks into structured subtasks with IDs, dependencies, acceptance criteria, and parallel execution hints. Feeds into planner synthesis for ordered implementation plans.

SWE-focused code analysis with Kimi K2.7-Code (SWE-bench Pro leader). Specialized for code review, bug detection, refactoring suggestions, and implementation planning. Agent Swarm spawns sub-agents for different aspects of code analysis.

Long-context analysis with Kimi K2.7-Code's full 256K token window. Analyze entire codebases, long documents, or extensive conversation histories. Ideal for cross-file analysis and large-scale code review.

Code fix, review, and optimization with MiniMax M3 (1M context, MSA sparse attention). Embedded SCoT, reflexion, and rubber_duck techniques. Per-task temperatures for optimal output.

Multi-step task decomposition with MiniMax M3. ReAct + least-to-most protocol with HALT criteria. Breaks tasks into ordered steps and executes with verification.

Deep reasoning with DeepSeek V4 Pro (open-weight frontier — top AIME/GPQA math & logic). Excels at long reasoning chains, proofs, and competitive-programming-grade analysis.

Algorithmic code review with DeepSeek V4 Pro (top AIME/CodeElo): correctness, complexity/Big-O, edge cases, data-structure choice, optimization. The strongest algorithmic reviewer in the roster.

Dedicated security audit (DeepSeek V4 Pro): taint/data-flow analysis, OWASP/CWE-mapped findings with severity, exploitability sketch, and concrete fixes. For code you are authorized to review.

Agentic reasoning & tool-use planning with Zhipu GLM-5.2 (SWE-Bench Pro leader). Strong at multi-step software-engineering plans and tool orchestration reasoning.

Efficient deep reasoning with StepFun Step 3.7 Flash (196B — high AIME/SWE-Verified at flash-tier cost). Near-frontier reasoning when budget matters.

Broad-knowledge reasoning with Baidu ERNIE 4.5 VL (424B/47B-active MoE — human-preference/arena strength). Best for questions needing wide world knowledge.

Query a local open-weight model (Ollama / LM Studio / llama.cpp / vLLM) — zero-cost, offline, private. Ollama is called via its native /api/chat endpoint so num_ctx is honored. Configure with LOCAL_LLM_BASE_URL / LOCAL_LLM_MODEL / LOCAL_LLM_NUM_CTX.

Google Search grounding with Gemini. Uses dynamic retrieval to ground responses in real-time Google Search results. Get factual, sourced answers with search citations.

Science-backed LLM-as-a-Judge evaluation (Gu et al., arXiv:2411.15594). Synthesize, evaluate, rank, or resolve multiple AI perspectives into a unified verdict. Uses chain-of-thought, first-principles, and adversarial reasoning.

Multi-model jury panel. Runs your question through configurable AI jurors in parallel (12 available: grok, openai, qwen, qwen_reason, kimi, perplexity, minimax, deepseek, glm, stepfun, ernie, local — 'hermes' is a legacy alias of local), then Gemini Judge synthesizes a unified verdict. Based on 'Replacing Judges with Juries' (Cohere, arXiv:2404.18796). The local juror runs whatever LOCAL_LLM_MODEL points at, e.g. a Nous Hermes build via Ollama. (The Hermes agent itself is model-agnostic — backed by GPT, Claude, Gemini, or self-hosted models — so 'hermes' never implied distinct jury weights.)

Multi-model diff-aware code review: 2-3 lab-diverse reviewers (Kimi K2.7-Code, DeepSeek V4 Pro, GPT-5.5) scoped to the changed lines, deduplicated and severity-ranked by a Gemini judge. Verdict: MERGEABLE / MERGEABLE WITH FIXES / DO NOT MERGE. Needs GEMINI_API_KEY plus OPENROUTER_API_KEY and/or OPENAI_API_KEY (reviewers self-drop if their key is missing).

Adversarial red-team of an existing plan (from any source): multi-model pre-mortem, hidden-assumption audit, ranked risks with mitigations, concrete plan edits, verdict. Complements planner_maker (builds) and planner_runner (executes). Needs GEMINI_API_KEY plus OPENROUTER_API_KEY, XAI_API_KEY, or OPENAI_API_KEY (critics self-drop if their key is missing).

Multi-model council for creating verified implementation plans. 6-step process: Grok searches for ground truth, Qwen analyzes feasibility, GPT-5.5 critiques gaps, Gemini scores quality. Returns confidence-scored plans.

Step-by-step execution of implementation plans with verification checkpoints. Tracks progress, validates each step, and provides 50% and 100% verification milestones.

Browse recent implementation plans. View plan summaries, status, and confidence scores. Filter by status or search by keywords.

Turn a loose feature request into a reviewable spec (GPT-5.5): user stories, Given/When/Then acceptance criteria, non-functional requirements, explicit out-of-scope, open questions. For sign-off BEFORE planning — feed the approved spec to planner_maker.

Discover prompt engineering techniques. Defaults to the ~9 core techniques that still help 2026 reasoning models (output-structure contracts like scot, pre_mortem, bdd_spec). Pass all=true for the full 31, or use preview_prompt_technique with technique="auto" if you don't know which to pick.

Preview how a technique enhances your prompt WITHOUT executing (returns an execution_token). Or pass technique="auto" with just a query to get the top techniques recommended for that task, with reasons.

Apply a research-backed prompt technique to any AI tool. Enhances your query with structured prompting patterns, then routes to the best model for that technique. Composable building blocks for thinking.

Opt-in prompt improver (cheap/fast model): restructures a raw query into a goal-first brief and shows its work — refined prompt + what changed + open questions. Never executes anything; you review, then feed the brief to any tool. Never injects reasoning instructions like "think step by step".