Council of High Intelligence
Put consequential decisions through independent analysis, direct challenge, and a verdict that preserves dissent.
- Source repo
- 0xNyk/council-of-high-intelligence
- Stars
- ★ 4.9k
- Last updated
- 6d ago
- License
- MIT
- Primary language
- Shell
- FA score
- 63/100 · Some gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformCodex · Claude Code
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- An engineering lead choosing between microservices and a monolith wants the maintenance and system-level tradeoffs challenged.
- Not a fit if
- Users who only need quick factual lookups or cheap, reversible experiments
- Teams that do not want multi-round cross-examination
- Website
- nyk.dev/oss/council-of-high-intelligence ↗ set on GitHub
- Maintenance
- Commits in the last 30 dayslast commit 2026-11-16latest release v1.2.0 (2026-07-04)+3.6 ★/day over 7 days
- Source review
- 63/100 · Some gaps
What does this agent do, and when should you use it?
Council of High Intelligence is a deliberation skill for Claude Code, Codex, Gemini CLI, and OpenCode. Users invoke /council for a full council, a quick panel, a two-person debate, or a named triad focused on a domain. Its protocol moves from independent analysis to cross-examination, final positions, and synthesis, using 18 analytical personas with documented methods and blind spots. The verdict records unresolved questions, dissent, kill criteria, and a concrete next action, while labeling evidence as fact, inference, assumption, or unknown. It installs into supported coding clients and is aimed at users weighing complex engineering or product choices; the README does not specify model credentials or costs.
After installation, a user submits a decision with /council and selects --full, --quick, --duo, or --triad <domain>, optionally choosing members with --members. In the full protocol, members restate the question and analyze independently, cross-examine one another, state final positions, and enter synthesis; checks look for premature agreement, repeated claims, missing dissent, and unsupported confidence. A detection script checks available provider paths and distributes seats across them, separating opposing positions across model families when possible. Users can preview routing with --dry-route or supply a seat map with --models <path>. The verdict includes open questions, acceptable compromises, kill criteria, and a next step; users can record a prediction, owner, and review date to track the outcome.
- An engineering lead choosing between microservices and a monolith wants the maintenance and system-level tradeoffs challenged.
- A product team deciding whether to ship can use
--triad shippingor a full council to examine timing and risk. - A founder considering open-sourcing a framework wants strategic, incentive, and long-term risk perspectives in one discussion.
- An AI product team choosing its next capability can use
--triad ai-productto bring relevant lenses together. - A team facing a decision defined by one central tension can use
--duoto compare two positions.
How do you install or deploy this agent?
Claude Code users can install through the plugin marketplace. For Codex, Gemini CLI, or OpenCode, clone the repository and run the matching installer. Restart the target client after installation. The README does not specify required API credentials, runtime versions, or external services.
How do you use this agent?
In a supported client, start with /council followed by the decision question. Choose full, quick, duo, or a named triad; the command remains /council in Codex, Gemini CLI, and OpenCode.
What are this agent's strengths and limitations?
- The full mode defines rounds for independent analysis, cross-examination, final positions, and synthesis instead of jumping straight to consensus.
- It includes 18 analytical personas, each with a primary lens and a named counterweight.
- Verdicts retain dissent, unresolved questions, kill criteria, and next actions; a split tally is returned as a split.
- It documents support for Claude Code, Codex, Gemini CLI, and OpenCode, with provider detection and seat-routing options.
- Installation requires the Claude Code marketplace or an installer script; users of other clients must clone the repository.
- Full deliberation uses multiple analysis and challenge rounds, so it takes more effort than a direct answer or quick mode.
- The docs describe several client and provider paths but do not specify required credentials, model charges, or precise provider-failure limits.
- The README provides no empirical evaluation of outcomes compared with single-model decisions.
How does this agent compare with similar options?
The README presents full council, quick mode, duo debate, direct answers, primary documentation, and low-cost experiments as different choices. It recommends direct answers or primary sources for factual lookups, experiments for cheap reversible choices, quick mode when breadth matters but cross-examination will not change the outcome, and duo mode when one tension dominates.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Council of High Intelligence This agent | 63 · Some gaps | Agent plugin / skillFree + model costs | ★ 4.9k | 6d ago | Shell | Codex · Claude Code |
| Plannotator | 80 · Good | CLIFreemium | ★ 14k | today | TypeScript | Codex · Claude Code |
| Ordewell | 87 · Good | CLIFree + model costs | ★ 211 | today | TypeScript | Codex · Claude Code · OpenAI API · Claude API |
| EOC — Everything OpenAI Codex Workflow OS | 64 · Some gaps | Agent plugin / skillFree + model costs | ★ 77 | 5mo ago | JavaScript | Codex |
How does FollowAgents rate this agent?
Why each dimension lost points
README, SECURITY.md, and SKILL.md describe local installation, local CLI detection, and environment-based keys kept out of argv, providing some visibility into permissions and data flow, so most related scores are 2. Deductions reflect that full mode defaults to 18 members and decision material should be handled carefully, while confirmation, external effects, and rollback are not fully specified. Security policy and contribution review describe intended source checks, but the publisher is unverified and neither identity nor review outcomes can be independently confirmed from this packet; attribution scores 2.
README, CLAUDE.md, and the workflows describe protocol checks, roster validation, and installer dry-runs. However, the supplied SKILL.md is truncated, so full cross-host consistency cannot be checked; no execution results or dependency-availability guarantees are shown. Provider fallback is described in README, but concrete error-message behavior is thin, so these criteria score 1.
README and SKILL.md cover full, quick, duo, triad, member, and configuration-based choices, with scenario guidance based on decision stakes and problem type. They also direct factual lookups and cheap reversible decisions to lighter methods, which supports clear boundaries. Deductions reflect that host-specific routing still depends on each CLI, configuration, and local tools, and the packet cannot establish full compatibility; environment fit and trigger precision score 2.
README provides a repository map, installation commands, mode examples, limitation guidance, and maintenance links. The LICENSE file is complete and explicitly MIT, earning 3. Deductions reflect that the supplied SKILL.md is truncated, leaving installation details, long-term naming stability, and release history only partly checkable. A maintainer, support policy, and security reporting route are named, but evidence of ongoing maintenance is limited.
The protocol calls for unresolved questions, dissent, kill criteria, and a concrete next step; README also gives prediction and review fields, making outputs usable for later decision tracking. Multiple perspectives and independent analysis may add value, but the packet does not quantify benefit over a single model or invocation costs. The 18-member full mode may be costly, so marginal value and cost-benefit score 2.
README distinguishes facts, inferences, assumptions, and unknowns, and locates protocols, configurations, and validation scripts, making design claims partly traceable. Deductions reflect that script contents and validation results are not included, and README claims about protocol and routing have limited independent corroboration. The fact/inference labeling rules are clearly stated and earn 3.
- Full mode can involve many members and multiple providers. Before supplying sensitive decision material, check the selected host, routing configuration, and provider data-handling practices.
- This review uses only the supplied static material. SKILL.md is incomplete; installation and runtime were not executed, and publisher identity, compatibility, and script behavior were not independently verified.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
Release notes excerpt
Added - Claude Code plugin-marketplace support (#45) — .claude-plugin/plugin.json + marketplace.json; install with /plugin marketplace add 0xNyk/council-of-high-intelligence then /plugin install council@council-of-high-intelligence. Root SKILL.md remains the single source of truth via a skills/council/ symlink; a new Asset Resolution section resolves agents/scripts/configs from either the install.sh layout or ${CLAUDE_PLUGIN_ROOT}. …View on GitHub ↗ -
Release notes excerpt
Added - — dedicated Codex council coordinator - Codex install support in (, flags) with reliability hardening - NVIDIA NIM provider support () — auto-detection via - Round 2 cross-examination anonymization — peer Round 1 outputs are masked behind stable labels in full and quick modes (Choi et al., arXiv:2510.07517; Karpathy ) - Anti-conformity directive — Round 2 prompts in all three modes now require members to name the specific flaw in their earlier argument before updating; defends correct prior positions against social pressure (Cui et al., Free-MAD, arXiv:2509.11035) …View on GitHub ↗ -
v1.0.0 v1.0.0 — Multi-Provider Council
Release notes excerpt
Council of High Intelligence v1.0.0 The first stable release. 18 historical thinkers deliberate on your hardest problems — now across multiple AI providers. What's New Multi-Provider Auto-Routing The council automatically detects installed LLM providers (Anthropic Claude, OpenAI via Codex, Google Gemini, Ollama local models) and distributes members across them for genuine model diversity. Zero config required. - Auto-detection script probes for available providers and outputs structured JSON …View on GitHub ↗
FAQ
Can I focus on one domain without convening the full council?
--triad <domain> to select three relevant lenses, such as --triad risk or --triad strategy.Does it require one specific model provider?
--no-auto-route keeps native-host defaults. Specific credential requirements are not stated.What happens when members still disagree?
Is it suited to factual lookups or a discussion where the decision is already made?
What does it cost to use?
Is Council of High Intelligence open source?
Is Council of High Intelligence safe to use?
Is Council of High Intelligence still maintained?
Who makes Council of High Intelligence?
What is Council of High Intelligence's official website?
Thanks — we use this to improve the page.