Reasoning in steps, not guesses.
TachiBot runs your prompt through structured reasoning — council, jury, planner, step-by-step — so models debate, verify, and reason deeper before they answer. One answer you can trust.
- "Have Grok check Twitter for that error message"
- "Ask Perplexity what changed in React 19 this week"
- "Get Gemini to brainstorm, then have Kimi K3 and GPT-5.6 both analyze it"
- 12
- Providers
- 67
- Tools
- 88
- Prompt techniques
Think, remember, act.
Three open-source pieces that work alone or together. Orchestrate many models, give agents lasting memory, and run them in a loop.
Stop trusting. Start verifying.
Same question. Different approach. Better answer.
- ×Generic advice, might miss recent updates
- ×No sources or documentation links
- ×Could confuse React 18 vs 19 features
- ×No way to verify accuracy
The Council & Jury
One question fans out to a panel of independent models in parallel. Their perspectives are aggregated, then a Gemini judge synthesizes a single verdict — the “replace judges with juries” pattern, made real.
A single evaluator. It scores, ranks, resolves conflicts, or merges perspectives into one answer — the synthesis primitive the jury and council both lean on.
Many jurors answer in parallel, then a judge synthesizes. One tool call, fast — the default for any “which is right / which is best?” question.
Research → adversarial debate → sequential reasoning → synthesis, iterated over rounds. For architecture calls where one round isn't enough.
Where each model is strongest.
Lab-diverse on purpose — different training data, different blind spots. Pick jurors per question; Gemini presides.
Opus-class flagship — real-time info, configurable reasoning effort, lowest hallucination rate.
Open-weight frontier reasoning — top AIME & CodeElo math/CP scores.
2.8T open-weight MoE — the largest open model shipped. Long-horizon agentic coding, 1M context, natively multimodal.
Agentic workflows & rock-solid tool-use reliability. 54% more token-efficient than 5.5.
Long-context multimodal reasoning & science — the presiding judge.
Long-horizon coding — SWE-Bench Pro 62.1, Terminal-Bench 81.0; agentic planning juror.
Near-frontier AIME/SWE-Verified reasoning at flash-tier cost.
Broad world knowledge & human-preference/arena strength.
High-speed frontier agentic coding & grounded search, 1M context.
Heavy mathematical & formal reasoning at configurable effort, 1M context, multimodal.
1M-token context + long-horizon agents, cost-efficient.
Free offline juror via Ollama / LM Studio — zero token cost, fully private. Runs any local model, incl. Nous Hermes builds. (The Hermes agent itself is model-agnostic — 300+ backends from GPT and Claude to self-hosted.)
Mix and match your reasoning pipeline.
Bulletproof implementation plans from a council of six models — ground truth, code analysis, decomposition, critique, synthesis, and a judged final plan with verification checkpoints.
- 01GrokSearches for ground truth
- 02QwenAnalyzes code
- 03Kimi K3Decomposes into ordered subtasks
- 04GPT-5.6Critiques gaps
- 05Qwen3.8 MaxDrafts the synthesis
- 06GeminiJudges with quality scores
A lab-diverse panel of jurors in parallel — 12 to choose from, including a free offline local juror. Gemini synthesizes the verdict.
Research-backed patterns — reflexion, SCoT, ReAct, pre_mortem and more. Embedded in tool prompts; preview before executing.
Chain unlimited steps with variables, dependencies, auto-distillation and an optional AI judge. Debates, research pipelines, code reviews.
Every tool, filterable.
Filter the full toolset by capability or provider — or search by name. Every tool follows one design rule: a single required string param, all enums optional with defaults.
Deep first-principles reasoning with optional heavy mode.
Mathematical & high-effort reasoning with GPT-6 Astra at xhigh reasoning effort.
Heavy mathematical reasoning (Qwen3.8 Max, 1M context).
Multi-perspective deliberation: optimist, pessimist, expert, contrarian.
Multimodal reasoning with a 100-subagent swarm.
Reasoning grounded in live web search.
Open-weight frontier reasoning — top AIME/GPQA math & logic chains.
Agentic reasoning & tool-use planning (SWE-Bench Pro leader).
Efficient deep reasoning — high AIME/SWE-Verified at flash-tier cost.
Broad-knowledge reasoning with human-preference/arena strength.
Query a local model via Ollama/LM Studio/llama.cpp/vLLM — free, offline, private.
Sequential thinking chain with optional per-step model execution.
Lightweight internal reasoning log.
Agentic code generation. SWE-bench >70%.
Algorithm analysis: complexity & optimization tiers.
Strongest algorithmic review: correctness, Big-O, edge cases, data structures.
Competitive-programming problem solving.
SWE-focused code generation & bug-fixing leader.
Single-pass code operations, cost-efficient (1M context).
Code analysis & optimization.
Specialized debugging assistance.
Ranked root-cause hypotheses with likelihoods, a discriminating check per hypothesis, and the minimal fix for the leader.
Comprehensive code review with focus areas.
Generate runnable tests via Qwen3-Coder-Next — edge cases enumerated first.
Multi-model diff-aware review — Kimi K3 + DeepSeek + GPT-6 Astra scoped to changed lines, Gemini-judged verdict.
Web search with current information (Sonar).
Live web search with Grok 4.6.
Cheaper live search on grok-4.3 — high-volume lookups and jury fan-outs.
Real-time web search via GPT-5.6.
Web search with Google Search grounding.
Code quality, security & performance analysis.
Rhetorical analysis: bias, fallacies, persuasion.
Long-context analysis across a 1M window.
Clear explanations for complex topics, any level.
OWASP/CWE security audit via DeepSeek V4 Pro — taint/data-flow analysis, severity per finding, concrete fixes.
Contrarian first-principles idea generation.
Find the 3rd, 4th & 5th alternative approaches.
Convergent synthesis — clusters ideas into themes.
Multi-model jury panel → Gemini judge synthesis.
Multi-perspective evaluation & verdict synthesis.
Council-based plan creation via coordinator pattern.
Execute plans against goal-oriented verification gates.
List recently created plans.
Structured task decomposition via agent swarm.
Multi-step task decomposition & execution.
System architecture & design via 4–16 agent swarm.
Adversarial multi-model plan red-team — pre-mortem, hidden assumptions, ranked risks, Gemini-judged verdict.
Loose request → reviewable spec via GPT-6 Astra: user stories, Given/When/Then, out-of-scope, open questions. Sign off before planning.
Researched long-form in one call — persona interview for the angle, then plan/draft/restructure/line-edit passes (storm + re3).
Execute YAML-based multi-step workflows.
Begin a workflow session with variables.
Resume an interrupted workflow session.
Discover available workflows.
Create a new YAML workflow definition.
Check progress of a running workflow session.
Validate workflow YAML: syntax, tool registry, dependency graph.
Validate a workflow file on disk before running.
Render a workflow's step graph for inspection.
Discover techniques — ~9 core by default, all=true for the full 31.
Preview an enhanced prompt before executing, or recommend techniques with technique="auto".
Run a query with a chosen technique applied.
Opt-in prompt improver — raw query → goal-first brief + what-changed + open questions. Never executes.
Smart auto-router: research, solve, verify, create, judge.
Mode-based multi-model reasoning (deep, debate, analyze).
Continue a focus session by its session ID.
View or reset tool usage & cost statistics.
Setup diagnostic — detected API keys, visible vs hidden tools (and why), active profile, suggested first step. Zero-cost.
12 providers.
One key, or yours.
Agentic reasoning, code review & brainstorming on the GPT-6 Astra flagship. GPT-5.6 tiers stay explicit — sol for explains, terra for code & search, luna for fast calls.
Multi-perspective analysis, brainstorming & the presiding jury judge on 3.1 Pro; search runs the agentic 3.7 Flash.
First-principles reasoning, architecture swarms & contrarian brainstorming. Opus-class flagship with configurable reasoning effort.
Largest open-weight model shipped — 2.8T MoE, natively multimodal. Task decomposition, long-horizon agentic coding & 1M-context analysis.
Algorithm analysis & heavy mathematical reasoning with configurable effort; agentic code generation on Qwen3-Coder-Next.
Cost-efficient coding & long-horizon agentic execution. MSA sparse attention, multimodal, 1M context.
Live web search & research grounded in current sources. Cost-optimized on Sonar.
Open-weight frontier reasoning — top AIME/CodeElo math & competitive programming. Default juror & strongest algorithmic reviewer.
Long-horizon coding flagship from Zhipu — SWE-Bench Pro 62.1, Terminal-Bench 81.0, usable 1M context.
Efficient deep reasoning — near-frontier AIME/SWE-Verified scores at flash-tier cost.
Broad world knowledge & human-preference strength. Multimodal 424B/47B-active MoE.
Any OpenAI-compatible local server — Ollama, LM Studio, llama.cpp, vLLM. Zero-cost, private, offline jurors.
- >70%
- Qwen3-Coder-Next · SWE-bench
- 62.1
- GLM-5.2 · SWE-Bench Pro
- 2.8T
- Kimi K3 · open-weight MoE
- −54%
- GPT-5.6 sol · tokens vs 5.5
- 1M
- Qwen3.8 Max · context + effort
Get started in minutes.
Add to Claude Code, Claude Desktop, or any MCP client.
npx -y -p tachibot-mcp tachibot initPrints the exact config · never writes or echoes keys
Give your agent a council.
New releases ship regularly. If it helps your workflow, a star helps others find it.
Report a bug or request a feature →