How TachiBot thinks.
From intent to a verified answer: how routing, planning, deliberation and verification fit together — and what happens on every single tool call.
Smart routing & tachi
The tachi auto-router reads intent from your query, picks a mode, and assembles the right tool chain — so you never have to memorize 67 tool names.
Generate a fix, then have a different model verify it.
Pull live, cited facts, then reason over them.
Diverge into many ideas, then refine the best.
Parallel jurors + a judge for any “which is best?”.
A 4–16 agent swarm for system-level design.
Check claims against evidence with adversarial eyes.
The planner pipeline.
planner_maker runs a 7-step council across six labs — each step handled by the model best at it — then planner_runner executes the plan against goal-oriented verification gates that catch drift early.
planner_maker is a coordinator: instead of one model writing the whole plan, it returns one tool at a time and routes each stage to the model that's best at it. A single model would carry its own blind spots through every step; a relay lets a fresh perspective correct the previous one.
- 0%Gemini 4 Argon “Sherlock”
Six deduction questions — are we even building the right thing?
- 10%Grok 4.7
Early drift detection — kill a wrong approach before it spreads.
- 25%GPT-6 Astra
Strategy validation — and can amend the plan if needed.
- 50%Qwen3.8 Max
Goal alignment with a genuinely different reasoning style.
- 80%Kimi K3
Decomposes the remaining work so nothing is left dangling.
- 100%GPT-6 Astra + Gemini 4 Argon
Two-judge final verdict + Reflexion Lite against the evidence.
How seven minds think together.
A structured reasoning pipeline — ground in data, decompose, explore alternatives, stress test for holes, then judge.
Search real-time data from 4 providers. No thinking starts without facts.
Decompose into atomic parts. Map dependencies, constraints, and execution order.
Generate alternative approaches from different training data and perspectives.
Attack assumptions. Find holes, blind spots, and failure modes in every path.
Synthesize the best elements from every model. Resolve conflicts. Score everything. Not 10? Here's why — and how to fix it.
Every deduction comes with a reason and a fix. No mystery scores.
Each model was chosen for a specific strength. Different training data, different benchmarks, different blind spots.
Faster than the 235B model it replaces, at half the cost and more depth — picked in a live head-to-head across five Qwen candidates. 1M context, multimodal, configurable reasoning effort.
AIME — American Invitational Math Exam. Top math competition, one level below the International Math Olympiad.
GPQA Diamond — PhD-level science questions written by domain experts. Tests deep reasoning, not memorization.
SWE-bench Pro leader — long-horizon, multi-file coding on real GitHub issues. 1T-param MoE, open-weights.
Token context — MSA sparse attention (~1/20 compute at 1M). Embedded SCoT, reflexion, and ReAct techniques with per-task temperatures.
SWE-Bench — agentic multi-file code generation. 80B/3B MoE at 262K context, and ~16x cheaper than the reasoning tier, so it stays the codegen model.
Five models. One answer.
Real GPT-5.5 → GPT-5.6 migration analysis, ~3 minutes
"Should I migrate from GPT-5.5 to GPT-5.6? Differences, breaking changes, migration steps."
- 1Version discoveryPerplexity
GPT-5.6 shipped Jul 9 — sol / terra / luna tiers replace pro and mini
- 2Feature comparisonGrok 4.5
1.05M context, 128K output, 54% more token-efficient than 5.5
- 3Migration impactGPT-5.6
Breaking: the $30/$180 pro tier is gone — use sol + high reasoning effort
- 4Performance analysisGemini
terra matches 5.5 at half the price; luna is cheapest but weaker on long context
- 5Final recommendationKimi K3
Low-risk: default to sol, route code to terra, keep luna off huge inputs
6 profiles.
Load only the tools you need. Set TACHIBOT_PROFILE to scale the surface from a lean 14 to the full 67 — fewer tools, less context spent on schemas.
Quick tasks, low token budget
Deep investigation, multi-source
Software development, SWE tasks
General-purpose, mixed workflows
Max code tools + agentic workflows
Everything enabled
How a tool call flows.
From your prompt to a frontier model and back, every request passes through the same six layers. The request flows down; the answer streams back up.
- 1MCP ClientClaude Code · stdio (Model Context Protocol)
You ask in natural language; Claude emits a structured tool call and streams the result back to you.
- 2FastMCP Serverserver.ts · safeAddTool()
Registers tools by active profile + available API keys, then wraps every call with input validation, usage/cost tracking and a heartbeat.
- 3Toolsrc/tools/*.ts · execute()
The tool's own logic runs — e.g. jury, planner_maker, grok_reason. Simple tools call a provider directly; orchestrating tools drop into the next layer.
- 4OrchestrationToolExecutionService · ModelProviderRegistry · ToolRouter · WorkflowEngine
Resolves model aliases → real tools, builds parameters, routes by intent, and drives multi-step / multi-round flows.
- 5Provider callcallGrok · callOpenAI · callGemini · callOpenRouter · callPerplexity
Normalizes the request, tries the OpenRouter gateway first, falls back to the direct API, applies per-model timeouts & reasoning effort.
- 6Model APIxAI · OpenAI · Google · OpenRouter · Perplexity
The actual frontier model runs and returns its answer — which travels back up the stack to you.
Who parses “ask grok & perplexity, then let gemini judge”?
- 1“ask grok … latest on the MCP spec”→grok_search
- 2“& perplexity”→perplexity_ask
- 3“then let gemini judge”→gemini_judgegets 1 + 2
How it holds together.
Registry-driven, SOLID, and obsessive about token economy.
tachi + focus read intent and orchestrate the right tools, modes and models for you.
ModelProviderRegistry maps 40+ aliases to tools; FocusModeRegistry adds reasoning modes without touching the core (OCP).
The planner returns a single nextTool at a time — fully user-interruptible, with visible progress.
2.5k chars between steps, 6k for synthesis, truncated on clean boundaries — beating lost-in-the-middle recall loss.
Variable interpolation, dependency resolution, parallel steps, retries and live manifest artifacts.
Every call wrapped by safeAddTool() — token tracking, validation, heartbeat to keep MCP alive.
tachi vs focus.
Both orchestrate many models — they differ in how much you decide. One picks the tools for you; the other hands you the controls.
Describe the goal in plain language. tachi reads your intent, picks the right task mode, and runs a pre-baked tool chain — zero config.
- →Auto-routes by keyword priority (ties resolve to the higher-stakes mode)
- →Each mode is an outcome recipe — generate, then verify
- →One-shot & stateless — returns the answer plus what it ran
› tachi "debug this null-pointer error"› tachi "microservices vs monolith for 10M users"› tachi "which is best — React, Vue or Svelte?"You pick the reasoning strategy and the panel. focus runs a controlled, multi-round deliberation exactly the way you specify.
- →You choose a process mode — how the models think together
- →Knobs: domain · rounds · models · temperature · ping-pong style
- →Multi-round & stateful — sessions resume via continue_focus
› /focus architecture-debate Redis vs Memcached› /focus research React 19 new features› /focus deep-reasoning scale to 10k connections| tachi | focus | |
|---|---|---|
| You provide | a query | a query + a chosen mode |
| Picks the mode | automatically (intent routing) | you do, explicitly |
| Mode type | task outcomes — solve, judge… | reasoning processes — debate, deep… |
| Rounds | one-shot recipe | multi-round (default 5), ping-pong |
| Config knobs | almost none | domain · rounds · models · temp · style |
| State | stateless | session-based, resumable |
| Reach for it when | “just handle this — pick the tools” | “reason this way, with these models” |
Built in public. Actively maintained.
New releases ship regularly. If it helps your workflow, a star helps others find it.
Report a bug or request a feature →