Multiple AI agents analyze your problem independently,
critique each other's solutions, and iterate until
they reach consensus—or an orchestrator resolves.
The value isn't individual outputs.
It's the meta-cognition: agents reviewing agents.
Architecture decisions fail at integration points. A single perspective—human or LLM—optimizes for one dimension while missing others.
Schema normalized for queries, but connection pool exhaustion under load
Microservices boundary clean, but distributed transaction hell
Auth flow secure, but latency budget blown on token validation
Cache invalidation "solved" until eventual consistency bites
Single-agent LLMs echo back your framing. Peer review surfaces the friction points.
The skeptical question: "Why not just prompt a single LLM to consider multiple perspectives?"
Each agent has a focused system prompt. A database agent isn't trying to also think about security—it reviews the security agent's work instead. Specialization without context pollution.
When Agent A objects to Agent B's solution, you see the specific critique. Single-LLM "multi-perspective" prompts tend to smooth over conflicts. Peer review makes friction visible.
Agents revise based on peer critique, not just their own re-reading. The revision incorporates external signal, not just self-consistency checks.
73% consensus is information. "I think this is good" isn't. The approval scores tell you where disagreement lives—and that's often where bugs hide.
Not "how similar are outputs?" but "would each agent approve the others' work?"
# 3 agents = 6 pairwise reviews Database → Backend: APPROVE (1.0) Database → Infra: CONCERNS (0.7) Backend → Database: APPROVE (1.0) Backend → Infra: OBJECT (0.0) Infra → Database: CONCERNS (0.7) Infra → Backend: APPROVE (1.0) ───────────────────────────────── Aggregate: (1.0 + 0.7 + 1.0 + 0.0 + 0.7 + 1.0) / 6 = 73% Threshold: 80% Result: Iterate or orchestrator resolves
Beyond the core workflow—features that make Consult practical for real engineering work.
Run the same query across Anthropic, OpenAI, and Google simultaneously. Agents from different providers critique each other's solutions. Claude reviews GPT's architecture. Gemini challenges Claude's assumptions. Disagreement across model families surfaces blind spots that single-provider analysis misses.
consult -p "..." -m team
Before burning tokens on full analysis, a lightweight pre-flight detects ambiguous queries. Asks only high-impact questions: scope boundaries, constraints, success criteria. Skips clarification for follow-ups clearly scoped by prior context. Explains why each question matters.
Automatic—triggers when ambiguity detected
Follow-up queries preserve full context. "Now add rate limiting to that design"—without re-explaining your schema, constraints, or prior decisions. Session state persisted to ~/.consult/sessions/. Resume conversations across terminal sessions.
Just keep typing in TUI, or use session flags in CLI
Drop in your schema.sql, architecture diagrams, error logs. PDFs automatically converted to images for providers lacking native support. Provider-specific size limits enforced gracefully. Conversion cached—no redundant processing across workflow phases.
F key in TUI, or --attach in CLI
Long conversations get AI-summarized when context window fills. Preserves: original question, final solution, key insights and constraints. Discards: intermediate back-and-forth, superseded ideas. Inspired by Claude Code's context management strategy.
C key in TUI, or automatic when threshold exceeded
Cost-optimized defaults: Haiku, GPT-4o-mini, Gemini Flash for expert agents. SOTA model (Opus) reserved for meta-review synthesis only. Override any model via environment variables. Switch providers mid-session in TUI.
ANTHROPIC_MODEL=claude-sonnet-4-20250514 consult ...
Prefer scripts and pipelines? The CLI does the same workflow without the interface.
$ consult -p "Order service: event sourcing vs state-based. 100k orders/day, \
need audit trail, eventual consistency for reads, strong for inventory" \
-e "database_expert,backend_expert,software_architect"
$ consult -p "SaaS auth: OAuth2 + SAML, tenant isolation, session management \
across subdomains, SOC2 compliance. Pain point: JWT token bloat" \
-e security_focused -i 3
$ consult -p "Product cache: 50k SKUs, prices from ERP every 15min, inventory \
real-time. Redis 5min TTL showing stale prices. Avoid cache stampede" \
-e "database_expert,backend_expert,performance_expert"
$ consult -p "Migrate user table UUID→ULID: 200M rows, 50+ FK references, \
zero-downtime. Aurora PostgreSQL. Evaluate dual-write vs shadow vs CDC" \
-e architecture -t 0.9
Not a web dashboard. A proper terminal UI with collapsible workflow visualization, live consensus tracking, and keyboard-driven navigation.
Those are code editors—they write code in your IDE. Consult is for architecture decisions before you write code. Use Consult to decide what to build, then your editor to build it. Complementary, not competitive.
You still need to understand your problem. Consult enhances analysis, it doesn't replace your judgment.
It's LLMs with structure. Better than raw ChatGPT for architecture decisions, but still LLM-powered with LLM limitations.
Full reasoning chains visible. Every approval/objection shows rationale. You see exactly why agents agreed or disagreed.
Not a chat wrapper. A structured peer review workflow that surfaces disagreement between perspectives.
BYOK model — you bring your own API keys, pay providers directly. We don't upcharge.
$9 USD/month
$90 USD/year
Save $18 — 2 months free
Your API keys and data never leave your machine. Here's exactly how.
Keys read from ~/.consult/.env or environment variables. Loaded into memory at runtime, never written to logs, never sent over network except directly to the provider (Anthropic/OpenAI/Google) via HTTPS.
Verify: tcpdump or mitmproxy shows only outbound connections to api.anthropic.com, api.openai.com, generativelanguage.googleapis.com.
Zero outbound connections to our servers. License validation is cryptographic signature verification performed locally—no network call required. We don't know who's using Consult, how often, or for what.
Verify: Block all outbound traffic except LLM providers. Consult continues to work.
Session history stored in ~/.consult/sessions/. Logs in ~/.consult/logs/. Both are local filesystem—encrypt at rest with FileVault/LUKS if your policy requires. No cloud sync, no external backup.
Verify: ls -la ~/.consult/ shows all persisted data. Delete anytime.
API keys pattern-matched and redacted in logs (sk-ant-***, sk-proj-***). Full prompt content logged only at DEBUG level (disabled by default). Production logging shows workflow events, not payload content.
Verify: grep -r "sk-" ~/.consult/logs/ returns zero matches.
We can't prevent Anthropic/OpenAI/Google from seeing your prompts—that's how LLMs work. If your compliance requires on-prem inference, Consult isn't the right tool. We're transparent about this boundary.
Mitigation: Use providers with data retention opt-outs. Anthropic API has zero-retention by default.
Each agent has domain-specific prompts that shape analysis. Peer review catches what individual perspectives miss.
Curated combinations for common scenarios. Use -e set_name in CLI.
# Install $ pip install getconsult # Configure API key (at least one required) $ mkdir -p ~/.consult $ echo 'ANTHROPIC_API_KEY=sk-ant-...' > ~/.consult/.env $ chmod 600 ~/.consult/.env # Or export directly $ export ANTHROPIC_API_KEY=sk-ant-... # Verify setup $ consult --status # Run your first query $ consult -p "Design a user authentication system" -e essentials
Supports Anthropic, OpenAI, and Google models.
Default uses cost-optimized models (Haiku, GPT-4o-mini, Gemini Flash).