Runtime safety
for AI agents.

Name: Savants
Author: Savants

344 verified agent incidents. 65 data deletions. 28.6M leaked secrets.
Savants stops the next one.

Get Started View Docs

Guard

Block dangerous actions.

Deterministic rules enforced outside the LLM. IDE + SDK.

Intelligence

3.6x fewer tokens per task, on average.

Semantic search. Call graphs. Blast radius. Benchmarked on 40 agents.

Knowledge

Remembers. Enforces. Grows.

Personal knowledge graph across every session, every tool. MCP-native.

curl -fsSL savants.sh | sh | pip install savants-guard | npm install @savants.dev/guard

30-second install. No API keys. No cloud required.

Scroll

Savants Guard

This already happened. To real companies.

Prompt instructions are suggestions. CLAUDE.md is a suggestion. .cursorrules is a suggestion. Savants is deterministic enforcement. The LLM cannot see it, reason about it, or bypass it.

April 2026

Cursor agent wiped a production database

PocketOS lost 3 months of customer data in 9 seconds. The agent deleted the database AND every backup. Savants Guard would have blocked it.

April 2026

28.6M secrets leaked via AI agents

AI coding tools doubled the rate of credential exposure. Anthropic rated the prompt injection attack CVSS 9.4 Critical. Savants blocks secret exfiltration.

2026

Regulators are catching up to MCP risks

NIST, NSA, and CISA have flagged AI agent tooling as a supply chain risk. Enterprises need audit trails before compliance frameworks harden. Savants provides audit-ready evidence.

~/.savants/guard-rules.json syncs from cloud every 30s

[
  "when command contains 'rm -rf /' then block",
  "when command contains 'git push --force' then rewrite 'git push --force-with-lease'",
  "when command contains 'npm publish' then ask 'Publishing is permanent'",
  "when command contains 'chmod 777' then suggest 'Use chmod 755 instead'",
  "when file_path contains '.env' then block"
]

Without Savants (real incident)

Agent> DELETE FROM reservations; DROP TABLE backups;
3 months of customer data gone. 9 seconds.

With Savants

Agent> DELETE FROM reservations; DROP TABLE backups;
BLOCKED. Database safe. Audit logged.

Works with Claude Code, Cursor, Windsurf, and any MCP-compatible tool. Rules sync from cloud across your whole team.

Install Savants Add to your SDK

curl -fsSL savants.sh | sh | pip install savants-guard | npm install @savants.dev/guard

Savants Intelligence

Your AI agent burns 3.6x more tokens than it should.

We tested 40 zero-context agents on 10 real code tasks. Without savants, investigations average $3.53 on Opus. With savants, $1.21. Same or better accuracy. Fewer tokens. Lower bill.

without savants

You: "What calls handleTokenExpiry and what breaks if I change it?"

Your AI agent:

→ grep -rn "handleTokenExpiry" src/ (47 matches)

→ reads auth/tokenRefresh.ts (3,200 tokens)

→ reads middleware/session.ts (2,800 tokens)

→ grep -rn "tokenRefresh" src/ (28 matches)

→ reads 6 more files...

→ grep -rn "import.*token" src/ (93 matches)

✗ 21 tool calls. 968K tokens. $3.53.

✗ Missed 2 indirect callers.

968K tokens vs 266K.
Same answer. 3.6x fewer tokens.

with savants

You: "What calls handleTokenExpiry and what breaks if I change it?"

Your AI agent:

→ savants callers handleTokenExpiry

→ savants blast handleTokenExpiry

7 callers across 4 services, blast radius = 12 files

Including indirect callers via SessionManager.refresh()

✓ 7.6 tool calls. 266K tokens. $1.21.

✓ Complete caller chain. Every indirect caller found.

Savants traces callers, callees, and blast radius across your entire codebase.

Benchmarked on 40 zero-context agents. 10 real tasks.

3.6x

fewer tokens on Claude Opus (Anthropic's top model). 968K → 266K average per investigation. Up to 9.4x on complex caller tracing.

2.9x

lower cost. $3.53 → $1.21 per task on Opus. $0.80 → $0.31 on Claude Sonnet. Same or better accuracy.

Equalizer

Claude Sonnet scores 55.5% accuracy without savants. With savants, it jumps to 65.5% — closing the gap with Opus at 89%.

Measured with fresh agents, no codebase context, no prior knowledge. Read the full methodology →

The model equalizer

Cheaper models score 55%.
With savants, they score 65%.

Claude Sonnet (Anthropic's mid-tier model) explores less efficiently and hits dead ends on complex tasks. Savants gives it the same structured code analysis as expensive models. Result: 2.6x cost reduction ($0.80 → $0.31/task) while accuracy jumps 10 percentage points. Opus (the top model) sees even larger savings: 2.9x ($3.53 → $1.21/task).

Sonnet without savants

55.5%

accuracy · 1.03M tokens · $0.80/task

Sonnet with savants

65.5%

accuracy · 365K tokens · $0.31/task

Real problems. Measured results.

Three scenarios from our 40-agent benchmark.

semantic_search "Find the function that retries failed payments"

✗ Without savants ~1M tokens

grep "retry" → 23 matches across config files, test mocks, comments. Reads them all.

✓ With savants ~265K tokens

handleTransactionWithBackoff() in payment-service.ts:142 — 3.8x fewer tokens

callers "What calls this function? Will my change break anything?"

✗ Without savants 1.17M tokens

grep the function name — miss indirect callers, dynamic imports. 9.4x more tokens on complex traces.

✓ With savants 125K tokens

7 confirmed callers across 4 services, blast radius = 12 files — 9.4x reduction

file_skeleton "Show me this file's structure without reading 2,000 lines"

✗ Without savants ~630K tokens

Read entire file, burn thousands of tokens on boilerplate. 4.2x more tokens on large function xrays.

✓ With savants ~151K tokens

34 functions with signatures. Just the structure you need — 4.2x reduction.

Every API doc. Already indexed.
Zero setup.

Your AI doesn't need to hallucinate Stripe's API or guess Kubernetes syntax. Savants indexes official documentation so your agent gets the right answer.

without docs

You: "How do I verify a Stripe webhook signature?"

AI: "I think you use crypto.createHmac('sha256', secret)...

actually let me check... maybe it's hmac.update()..."

Hallucinated. Partially wrong. You debug for 20 minutes.

with savants docs

You: "How do I verify a Stripe webhook signature?"

AI → savants docs search stripe "webhook signature"

Resolve webhook signature verification errors

docs.stripe.com/webhooks/signatures

"Use stripe.webhooks.constructEvent(body, sig, secret)"

Correct. From official Stripe docs. Zero hallucination.

Stripe

Kubernetes

React

Next.js

Fastify

PostgreSQL

Redis

Docker

AWS

Cloudflare

Pre-indexed and free. Plus upload your private docs: 1 credit per 10K tokens.

60 seconds. 3.6x fewer tokens on average.

Step 01

Install

Single binary, zero dependencies. No bash, no Python, no runtime needed. macOS, Linux (including NixOS), and Windows.

curl -fsSL savants.sh | sh

Windows: irm releases.savants.dev/latest/install.ps1 | iex

Step 02

Index

Savants builds a map of every function, every caller, every import. Locally.

savants reindex

Step 03

Done

Your AI agent automatically uses savants instead of grep. Nothing else to configure.

savants search "auth middleware"

We found a real bug in Fastify.

Root cause identified, call chain traced, PR submitted - in 141 seconds. Without savants, the same task took 17 minutes of manual grep and file reading.

Read the case study →

Knowledge Graph

The longer you use it, the smarter it gets.

Savants builds a personal knowledge graph from every session. It remembers across tools, enforces across teams, and follows you from Claude Code to Cursor to Windsurf.

Day 1

Install. Guard blocks. Intelligence saves tokens.

Savants Guard stops dangerous actions deterministically. The code graph gives your agent 3.6x fewer tokens per task. Zero configuration — just install and go.

Day 30

Knows your codebase, preferences, and team.

The knowledge graph learns which services you own, how you prefer to debug, and who your teammates are. Context injection happens automatically via MCP — every AI tool you use gets smarter.

Day 90

Enforces team decisions. Recalls incident history.

Guard rules attach to knowledge nodes: "staging before prod for payments-service" becomes an enforced policy, not a suggestion. Your agent remembers last quarter's auth outage when you touch the same code.

Day 360

Organizational intelligence. Switching costs a year of context.

A year of decisions, patterns, ownership maps, and incident history — structured as a graph, not flat memory. Nodes are code entities, people, events, and policies. Edges are real relationships: owns, changed, caused, depends-on.

Not flat memory

Mem0 and Zep store facts per user. Savants stores a graph — code entities, people, events, and policies connected by real relationships. That's why it can enforce, not just recall.

Not per-tool

Because Savants is MCP-native, the knowledge graph follows you across Claude Code, Cursor, Windsurf, and any MCP-compatible tool. One graph, every session, every IDE.

Sub-50ms recall

Graph traversal + vector search on Cloudflare's edge. D1 reads in 5ms, Vectorize queries in 15ms. Context selected and injected before the LLM sees your message.

Local tools are free.

Cloud tools add production intelligence. Pay per call, no commitments.

diagnose-error 50 credits

Root cause file + line in 0.7s. Upstream trace, git blame, Slack context.

Example

Input: "PrismaClientKnownRequestError: Foreign key constraint violated on ProcessingJob_userId"

Root cause: sendFailureNotification() at resume-review.activities.ts:838

Upstream trace:

← triggered by generateAIReview() (LLM validation failed)
← triggered by Temporal worker activity

Fix: Verify userId exists before prisma.create()

Confidence: 85% (enriched with Sentry breadcrumbs)

pr-risk 20 credits

8-check risk analysis per PR. Blast radius, affected consumers, test gaps.

Example

Input: PR #847 - "Refactor payment retry logic"

Risk: Medium (6/10)

Blast radius: 12 files, 3 services affected

Warnings:

⚠ Removed error retry - was this intentional?
⚠ paymentService.ts has no test coverage for this path
✓ No breaking API changes detected

diff-impact 10 credits

Full downstream impact analysis across services and modules.

Example

Input: "What breaks if I rename processOrder()?"

Affected: 18 files across 4 services

api-gateway: 3 routes call processOrder()
billing-service: 2 event handlers
notification-service: 1 webhook
12 test files reference this function

Rename safely with these 18 changes.

radar 10 credits

What-did-I-miss digest. Surfaces drift between your code and production.

Example

Input: "What changed in the last 24 hours?"

3 items need attention:

● Sentry: 12 new errors in auth-service (started 3h ago)

● PR #892 merged without review (config change)

● Deployment v2.4.2 healthy (all checks passed)

10 free cloud calls/month. No card required. Start free → | View pricing →

$ savants connect

Works with your tools

Claude Code

Cursor

Windsurf

Indexes TypeScript, JavaScript, Python, Rust, Go, Java, C, C++, C#, Ruby, PHP, Swift, Kotlin, and Scala.

Runtime safety for AI agents.

This already happened. To real companies.

Cursor agent wiped a production database

28.6M secrets leaked via AI agents

Regulators are catching up to MCP risks

Your AI agent burns 3.6x more tokens than it should.

Benchmarked on 40 zero-context agents. 10 real tasks.

Cheaper models score 55%. With savants, they score 65%.

Real problems. Measured results.

Every API doc. Already indexed. Zero setup.

60 seconds. 3.6x fewer tokens on average.

Install

Index

Done

The longer you use it, the smarter it gets.

Install. Guard blocks. Intelligence saves tokens.

Knows your codebase, preferences, and team.

Enforces team decisions. Recalls incident history.

Organizational intelligence. Switching costs a year of context.

Local tools are free.

diagnose-error 50 credits

pr-risk 20 credits

diff-impact 10 credits

radar 10 credits

Works with your tools

Runtime safety
for AI agents.

Cheaper models score 55%.
With savants, they score 65%.

Every API doc. Already indexed.
Zero setup.