Claude Code
Anthropic
Anthropic’s terminal coding agent reads repositories, edits files, runs commands, supports subagents and hooks, and connects external tools through MCP.
claude
Auto-config available
Rust 1.95+
MCP rmcp 1.6
run_thinking_mode
Auto · Quick · Explore · Map · Sketch · Test
fail-closed verification
Apache-2.0
Turn AI Thinking into
Structured Protocols
ReasonKit Think is a Model Context Protocol server that gives your AI coding agents practical Thinking Modes — Auto, Quick, Explore, Map, Sketch, and Test — plus fail-closed verification and governance.
OS
Shell
Alt
curl -fsSL https://get.reasonkit.sh/think | bash
Apache 2.0 Licensed
No Vendor Lock-in
GDPR Compliant
100% Rust — Zero GC pauses
Self-Hostable
The ReasonKit Advantage
Why ReasonKit Think?
Most AI agents reason shallowly (fast but wrong) or deeply but opaquely (slow and unverifiable). ReasonKit Think bridges that gap with structured, auditable, governable reasoning.
Anti-Hallucination HARD GATE
Strict evidence-based verification gates. Claims without proof are marked
DATA_DEFICIT and fail closed — never silently passed through.Traceable Auditability
Every thought, branch, and decision is logged in a stateful DAG for post-mortem analysis. Replay any reasoning session from first principles.
Dynamic Paradigms
Seamlessly switch between CoT, ToT, and GoT on the fly. Your agent picks the right reasoning mode for the complexity level of each task.
Skeleton-of-Thought
For multi-aspect problems, route into outline-first skeleton branches, elaborate them in parallel, then merge via GoT checkpoints.
Honest Provider Boundaries
Heuristic outputs are labeled as heuristic. Model-backed analysis fails closed until real provider bindings exist. No fake confidence.
State Survival
Local snapshots survive process restarts. Telemetry resources expose tool latency, verification outcomes, pruning, and graph-depth metrics.
Assumption Ledger
Critical premises are tracked and can block convergence until resolved. Unresolved high/critical assumptions gate consensus and synthesis.
Replayable Convergence
Agents can replay session timelines, run checkpoints, and converge only after quality, evidence, and assumption gates pass.
Reasoning Modes
Six Thinking Modes, One MCP Server
Start with Auto by default. Switch modes only when the problem needs more branching, mapping, or verification depth.
Mode 01
Auto
Adaptive default
Routes your task to the safest practical workflow, executes it, and returns an auditable summary of what happened.
Best for: Most requests when you do not want to pick tools manually.
"Use Auto mode. Compare options, verify risky claims, and return an auditable recommendation."
Mode 02
Quick
Concise linear audit
A short CoT-style pass with assumptions, checks, and conclusion. Fast and practical without hidden reasoning dump.
Best for: First-pass debugging, lightweight reviews, quick triage.
"Use Quick mode and give a concise audit of likely root causes."
Mode 03
Explore
Branch and prune
Tree-of-Thoughts branching for alternatives. Score branches, prune weak paths, and keep the strongest options.
Best for: Tradeoffs, migration choices, incident strategy options.
"Use Explore mode. Generate three strategies, score them, and prune the weakest."
Mode 04
Map
Graph synthesis
Graph-of-Thoughts mode for dependencies, conflicts, and merged plans with explicit rationale.
Best for: Architecture planning and multi-system rollout design.
"Use Map mode to merge compatible branches into one final plan."
Mode 05
Sketch
Outline-first reasoning
Skeleton-of-Thought flow: generate structure first, then elaborate sections under suitability gates.
Best for: Big proposals, multi-part plans, independent lenses.
"Use Sketch mode. Build an outline first, then expand each section with checkpoints."
Mode 06
Test
Verification-first
Runs CoVe and governance checks, classifies evidence quality, and returns go/no-go route decisions.
Best for: High-risk changes where correctness and evidence matter most.
"Use Test mode. Extract claims, mark evidence gaps, and decide whether we can proceed."
Verification Engine
Reasoning is Useless
Without Evidence
ReasonKit Think enforces CoVe (Chain of Verification) and Triangulation. Every deep deliberation automatically builds a Verification Matrix.
1
Extract Claims
Pulls out factual or architectural assertions from the AI's reasoning chain automatically.
2
Evaluate Evidence
Checks if claims are supported by Tier 1 sources: docs, codebase, verified logs, and authoritative references.
3
Fail Closed
Conflicting sources or missing data trigger GATHER_MORE_EVIDENCE instead of outputting potentially unsafe code.
VERIFIED
3 Tier-1 sources confirmed
HEURISTIC
Structural proxy, review recommended
DATA_DEFICIT
Claim blocked — gather evidence first
SOURCE_CONFLICT
Conflicting sources — cannot synthesize
ASSUMPTION_RESOLVED
All premises validated, convergence allowed
Autopilot Mode
Speak Human. Think Structured.
You don't need to remember MCP tool names. The reasoning_intent_router and
reasoning_autopilot understand natural language. Just talk to your agent normally.
CoT / Quick
"step by step"
"quick think"
"linear reasoning"
"first pass"
"quick CoT"
ToT / Branching
"explore options"
"compare options"
"multiple paths"
"score by constraints"
GoT / Synthesis
"join branches"
"merge paths"
"refine assumptions"
"distill a plan"
Skeleton-of-Thought
"outline first"
"parallel branches"
"multi-aspect answer"
"skeleton-of-thought"
Verification
"fact-check"
"cross-check"
"validate claims"
"verification failure"
Governance
"go/no-go"
"decision gate"
"should we proceed"
"checkpoint"
Assumption Ledger
"record this assumption"
"premise gate"
"must be true"
Memory / Replay
"replay the trace"
"export memory snapshot"
"remember this reasoning"
Compatibility
Works With Your Favorite Tools
Connect ReasonKit Think to the coding tools you already use. Each card shows the current setup path; automatic configuration is only claimed where the shipped installer is verified today.
Claude Code
Anthropic
Gemini CLI / Antigravity CLI
Google
Google's terminal agents support MCP, code and shell tools, extensions, and multi-agent workflows; Antigravity serves individual users while Gemini CLI remains available for enterprise and API-key access.
Gemini: Enterprise / API key / VertexIndividual Google AI Pro, Ultra, and free users moved to Antigravity CLI on June 18, 2026.
gemini / agy
Gemini: Auto-config
Antigravity: Update in progress
Codex CLI
OpenAI
OpenAI’s local coding agent inspects repositories, edits files, runs commands in configurable sandboxes, supports subagents and plugins, and connects MCP servers.
codex
Auto-config available
Cursor
Anysphere
Anysphere’s terminal agent plans, searches, edits, and automates codebases interactively or headlessly, sharing rules and MCP configuration with the Cursor editor.
agent
Installer update in progress
GitHub Copilot CLI
GitHub
GitHub’s standalone terminal agent edits code, runs tools with permissions and sandboxing, and supports custom agents, hooks, plugins, and configurable MCP servers.
copilot
Installer update in progress
VS Code
Microsoft
Microsoft’s extensible code editor combines language tooling, debugging, source control, terminals, and agent chat with local or remote MCP servers.
code
Auto-config available
OpenCode
Anomaly Innovations
Anomaly’s open-source terminal coding agent supports multiple model providers, editable agents, skills, permissions, plugins, and local or remote MCP servers.
opencode2
Installer update in progress
Qwen Code
Alibaba Cloud · Qwen Team
Alibaba’s terminal coding agent plans and edits features, debugs repositories, runs shell and web tools, supports subagents, automation, and MCP extensions.
qwen
Installer update in progress
Zed
Zed Industries
Zed Industries’ collaborative code editor emphasizes responsiveness, integrated AI agents, terminal and debugging workflows, and local or remote MCP context servers.
zed
Installer update in progress
Windsurf / Devin Local
Cognition
Cognition’s Devin Local runs inside Devin Desktop and shares the Devin CLI harness, local files, subagents, sandboxing, permissions, skills, and MCP.
devin
Installer update in progress
Grok Build
SpaceXAI
SpaceXAI’s extensible coding agent offers interactive and headless modes, plan review, sandboxed commands, plugins, hooks, skills, parallel subagents, and MCP.
grok
Manual setup
Hermes Agent
Nous Research
Nous Research’s open-source, self-improving agent runs locally across terminal and messaging channels, with skills, memory, scheduled tasks, tool execution, and MCP integration.
hermes
Manual setup
Get Started
Up and Running in 30 Seconds
1
Install the MCP Server
One command installs the binary, detects your AI clients, and patches their MCP configs automatically. No manual JSON editing.
curl -fsSL https://get.reasonkit.sh/think | bash
2
Restart Your Client
Restart Claude Code, Cursor, VS Code, or whichever AI client you use. The MCP server starts automatically via stdio transport.
claude # or cursor, code, gemini…
3
Think Differently
Ask your agent to use Auto, Quick, Explore, Map, Sketch, or Test mode depending on the depth and verification you need.
"Use Auto mode and return an auditable recommendation."
Make Your AI Agent
Actually Think
Stop trusting shallow reasoning. Install ReasonKit Think in 30 seconds and give your agent a structured, auditable, governable mind.
curl -fsSL https://get.reasonkit.sh/think | bash
ReasonKit is built by Len P. van der Hof as a product of LPH98.ventures. The same operator writing on evaluation, override doctrine, and versioning an agent OS lives at Lenvanderhof.com.