Why Amodal

The thesis behind the product. Why configuration beats intelligence, why the runtime is free, and why this pattern keeps repeating.

Configuration is orthogonal to intelligence.

Models get smarter every quarter. But even a model that knows everything still needs to be told what to do, how to do it, what it's allowed to do, and what your specific business looks like. That's not intelligence. That's configuration. And it gets more valuable as models improve.

1

Your methodology (Skills)

"How we triage deals at Acme" isn't in any training data. Your sales cycle is 45 days for mid-market, 90 for enterprise. Deals go stale at different rates per stage. Enterprise deals get 2x thresholds. None of this is something a model learns. You configure it.

skills/deal-triage.md
# Deal Triage
 
### 1. Load the deal
Pull from Salesforce. Check size, stage.
### 2. Check decision-maker
VP or above by Proposal stage.
### 3. Score and prioritize
Enterprise gets 2x thresholds.
2

Your system access (Connections)

Claude knowing the Salesforce API schema doesn't mean it has your OAuth tokens. It doesn't know which 5 of 47 endpoints matter to your sales team. It doesn't know your instance URL, your custom fields, or your rate limit quota.

Salesforce
OAuth2 · 5 endpoints · acme.salesforce.com
Gong
API key · 3 endpoints · calls + sentiment
HubSpot
OAuth2 · 4 endpoints · contacts + engagement
3

Your private data (Knowledge)

Your baselines ("normal login volume is 2,000-3,000/day"). Your patterns ("this vendor always resubmits invoices late in Q4"). Your false positives ("the 2am activity in Zone B is scheduled maintenance"). Private. Specific. Not in the training data.

Pattern

"Deals silent 14+ days in Negotiation close at 8%"

Source: 47 sessions over 3 weeks

False positive

"Acme vendor resubmits on the 15th, not a duplicate"

Flagged 3x, resolved identically

4

Your guardrails (Rules & Access)

Field-level access control: the agent reads opportunity amounts but never exposes margin data below VP. Confirmation tiers: writes require human review. Rate limits. PII handling. A smarter model doesn't make guardrails less important. It makes them more important.

Confirmation required
Update Salesforce: Proposal → Negotiation
Notify #sales-ops in Slack
Rate limit: 10 writes/hr · Audit: full context logged · PII: email role-gated
5

Your presentation (Output & Brand)

Which widgets render. What language to use. What the agent's personality is. What gets emphasized vs buried. This is UX and brand. It changes per customer, per channel, per audience.

PropertyPro
A
3 units have overdue maintenance...
LegalDesk
A
Clause 4.2 deviates from standard...
Counterintuitive

Smarter models need more configuration, not less.

The assumption is that as models get smarter, you need less tooling around them. The opposite is true.

More capabilities = more to govern.

A model that can only read needs simple guardrails. A model that can read, write, delete, send emails, create Jira tickets, and update CRM records needs a comprehensive policy layer. Every new capability is a new surface area for misconfiguration.

More intelligence = more context required.

A smarter model can do more with the right context. But "the right context" is your specific sales process, your specific API endpoints, your specific compliance requirements. The smarter the model, the more valuable precise configuration becomes.

More adoption = more diversity of use.

When every team in the company has an agent, the configuration surface explodes. Sales ops wants deal triage with their pipeline stages. Finance wants reconciliation with their chart of accounts. HR wants onboarding with their specific policies. Same runtime, different configuration.

Rules are the new code.

A skill is a methodology file. Your sales playbook, your investigation runbook, your compliance checklist. Domain experts author them. Engineers don't need to be involved.

The repo structure IS the product. Skills define what the agent does. Connections define where data lives. Knowledge defines what the agent knows. Evals define how you know it works. All version-controlled. All reviewable in a PR. All testable in CI.

This is the Terraform insight applied to AI: the configuration is more valuable than the runtime. You don't build infrastructure by writing Go code for the AWS API. You write .tf files. You don't build agents by writing Python glue around an LLM. You write .md files.

skills/deal-triage.md
# Deal Triage
 
## Steps
 
### 1. Load the deal
Pull the deal from Salesforce. Check size,
stage, owner, last activity date.
 
### 2. Check decision-maker
Is a VP or above in the contact roles?
Enterprise deals require this by Proposal stage.
 
### 3. Calculate velocity
Days in current stage vs historical average.
Enterprise gets 2x thresholds.
 
### 4. Score and prioritize
Flag stale deals. Rank by close probability.
Output top 5 with reasoning.

You bring the domain. We bring the infrastructure.

Your repo contains the domain-specific parts. Everything else is the platform.

You write
Amodal provides
Connections (credentials + API docs)
Secret management, encrypted at rest
Skills (methodology as markdown)
Reasoning engine that follows them
Knowledge (context as markdown)
On-demand loading, learning flywheel
Channels config (Slack, Teams, web)
Integrations, delivery, streaming
Evals (YAML test suites)
CI quality gates, regression detection
git push
Hosting, scaling, zero-downtime deploys
Conversation history per user
Admin dashboard (usage, costs, sessions)
Access control (who sees what)
Observability (is it working well?)
Model experiments (cheaper model? better model?)

Production infrastructure, included.

Everything you need to run AI agents at scale, out of the box.

Cost tracking

Per-session, per-skill, per-tenant cost breakdown. See exactly what every agent run costs. Compare models on your actual workloads. Set budget alerts before you get a surprise bill.

This month
deal-triage (Haiku)$18.40
compliance-check (Sonnet)$22.10
pipeline-review (Haiku)$6.73
Total$47.23
Switch deal-triage to Haiku. Save ~$8/mo.

Eval suites

Every deploy runs your eval suite. Skills that regress don't reach production. PRs show pass rates. Compare quality across model swaps before committing.

deal-triage
✓ 14/14 passed (100%)
compliance-check
✓ 11/12 passed (91.7%)
tenant-screening
⚠ 8/10 passed (80%)
Deploy blocked. tenant-screening needs review.

Session replay

Full transcript of every conversation. What the user asked, which skills fired, what tools were called, what actions were taken, and what the agent said back. Aggregate query patterns across tenants to find what users actually need.

Session 8842
User: "Move NovaTech to Negotiation"
Skill: deal-triage
Tools: request (Salesforce), request (Gong), request (HubSpot)
Actions: update_deal, set_next_step, notify_slack
Status: confirmed by sarah@acme.com
Cost: $0.008 (Haiku)

Team and governance

Role-based access. Domain experts edit skills in the browser. Engineers manage connections in git. SSO and audit logging for enterprise. Multi-tenant isolation so each customer's data stays separate.

Team
SK
Sarah Kim
Admin
MC
Marcus Chen
Skill Editor
DP
Dana Park
Viewer
SSO via Okta ✓ · Audit logging ✓ · 4 tenants isolated

This pattern has played out before.

Every major platform develops a configuration layer. The core technology provides the engine. The configuration layer is how teams actually use it in production.

Platform
Configuration layer
What it became
Cloud (AWS, GCP, Azure)
Terraform
How teams configure cloud infrastructure
Node.js
npm
How developers distribute and share packages
Containers (Docker)
Kubernetes + Helm
How teams orchestrate and deploy services
Linux
apt, yum
How you configure and manage a system
LLMs (Claude, GPT, Gemini)
Amodal
How teams configure AI agents

What the configuration layer handles.

Model providers focus on intelligence. The configuration layer handles everything around it: how you connect systems, define methodology, manage access, and share across teams.

What the model provides
What Amodal provides
Reasoning and generation
Methodology frameworks scoped to your process
Tool calling
Curated connection packages with auth + endpoints + entities
Context windows
Tenant-specific knowledge that compounds with use
Function execution
Field-level access policies and confirmation tiers
Fine-tuning
Marketplace of community skills and connections
Embeddings and retrieval
Override model: install a package, customize what's yours

npm doesn't replace Node.js. It's the layer that makes Node useful in production. Amodal is the same layer for AI agents.

Every ecosystem gets a package manager.

Node got npm. Python got pip. Rust got Cargo. Infrastructure got Terraform providers. AI agents will get one too. We think it should be open.

Connections are the killer content. When Company A publishes a Salesforce connection package with curated endpoints, entity descriptions, and access policies. Company B installs it and customizes only what's specific to them. The catalog compounds.

Terraform went from 20 providers to 3,000+. Each one made the platform more valuable for everyone. We're building the same dynamic for agent configuration.

Terminal
$ amodal install @amodalai/salesforce
 
Installing from registry.amodalai.com...
+ connections/salesforce.json
+ knowledge/salesforce-surface.md
+ knowledge/salesforce-entities.md
 
$ amodal install @amodalai/skill-deal-triage
 
+ skills/deal-triage.md
+ evals/deal-triage.eval.yaml
+ knowledge/deal-stage-definitions.md
 
Installed 2 packages. Run `amodal dev` to test.

We want a hundred companies building on us for free.

A hundred companies will build AI agents for their vertical: security, compliance, finance, HR, customer success. If we charge a platform tax on the runtime, they'll evaluate us against building their own. They have VC money. They'll build their own. We get zero.

If the CLI is free forever, they build their agents as Amodal skills. Ship in weeks, not years. Their customers use our skill format. The format becomes the standard. The ecosystem compounds.

We don't make money from the next vertical agent company directly. We make money because they proved our format works for security, which makes the next company choose it for compliance, which makes the next one choose it for finance.

Free forever (Apache 2.0)
CLI runtime (ReAct loop, tools, skills)
The skill / connection / automation format
Public registry (read and publish)
Self-hosting your own agent
Local development and testing
Platform (paid)
Multi-tenant hosting and isolation
Admin UI (marketplace, KB review, team config)
Chat widget for embedding in products
Enterprise auth (SSO, RBAC, audit logs)
Managed infrastructure and automations

Ready to start building?

Open source. Free to start. The platform scales when you need it.

Get Started