Join our Slack community
Governance-as-Code for AI agents

Govern your agents in prod.

The open source policy layer for AI agents. Deterministic enforcement, real human approvals, tamper-evident audit. No LLM watching another LLM.

Open source. Self-hosted or cloud. Works with LangGraph, CrewAI, AutoGen, Claude Agents SDK, MCP, and 8 more.

faramesh • runtime
stripe.refund(amount=5000)
DEFER
awaiting approval · amount > $1000
How it fits

Every action, checked before it runs.

Faramesh sits between your agents and the tools they call. Before any action runs, your policy decides what happens. Run it. Block it. Hold it for a human.

Your agents can't bypass it. There is no ungoverned path to your tools.

Non-bypassable
Deterministic
Auditable
Agent action
Attempting
stripe.refund(12000)
Faramesh
Policy check
PERMITaction runs
DENYblocked by policy
DEFERheld for approval
How policies are written

Policy as code, not prompts.

FPL, our native policy language for AI agents. Deterministic, versioned, and enforced on every tool call.

policies/default.fplFPL
1
2
3
4
5
6
7
8
9
agent coding-bot {
default deny
framework "langchain"
 
rules {
default deny
 
deny! shell/* reason: "never shell in prod"
permit git/read
permit git/commit when branch != "main"
}
}
→ evaluate
action
evaluated in 2msfaramesh v0.4.2
How it runs

Built for production.

Every tool call goes through five stages of defense. Scroll through a refund request as it flows through the runtime.

Stage 01 of 05

Your agent calls a tool

Your agents keep using their framework as normal. No wrappers, no code changes.

$ agent.invoke("refund #4821")
→ stripe.refund(4821)
Stage 02 of 05

Faramesh intercepts

Before the call reaches your tool, our runtime catches it. Non-bypassable.

02
intercept
captured before execution
Stage 03 of 05

Policy evaluates

Your FPL policy runs deterministically. Microseconds, not seconds.

03
policy.evaluate
rules/refund.fpl
matched: amount >= 500 → defer
Stage 04 of 05

Verdict fires

Permit, deny, or defer. The call either runs, is blocked, or waits for a human.

04
verdict
DEFER
awaiting human approval
Stage 05 of 05

Audit trail written

Every decision is hashed and chained. Tamper-evident record you can replay any time.

05
audit.log
sha256:a1f9c3…e7b4
signed & immutable
68 ns
Engine eval

per simple permit, zero allocations

58 µs
Full pipeline

median latency per decision

13
Frameworks

auto-patched at runtime, zero code changes

Framework support

Works with what you already use.

Drop in to any agent framework. No code changes. No lock-in.

Supported frameworks

13 frameworks. One integration point.

Auto-patches your framework's tool-call methods at runtime. Your agent code stays untouched.

LangGraph·
LangChain·
CrewAI·
AutoGen·
Pydantic AI·
Google ADK·
LlamaIndex·
AWS Strands·
OpenAI Agents SDK·
Smolagents·
Haystack·
Deep Agents·
AWS Bedrock AgentCore·
MCP Servers·
LangGraph·
LangChain·
CrewAI·
AutoGen·
Pydantic AI·
Google ADK·
LlamaIndex·
AWS Strands·
OpenAI Agents SDK·
Smolagents·
Haystack·
Deep Agents·
AWS Bedrock AgentCore·
MCP Servers·
Install

Choose the installation path that fits your environment.

Multiple platforms supported
Integration cost
0
code changes to your agent

Faramesh intercepts at the framework layer, not the application layer. Your agent stays exactly as it is.

Latency
~57us
per policy decision

Production-grade latency. Faster than a network round trip. Inline with every tool call.

What your team ships

From prototype to production.

The parts of shipping agents that aren't about the agent.

SHIP

Graduate agents from prototype to production.

Refund flows. Database writes. External APIs.

AUDIT

Produce board-ready evidence for every agent action.

Signed, immutable, and replayable. The trail your reviewers want.

OBSERVE

Know exactly why an agent did what it did.

Deterministic verdicts, not probabilistic guesses. The policy that ran. The rule that matched. The reason the verdict fired.

APPROVE

Route high-impact actions to a human before they run.

Large refunds. Production deploys. Wire transfers. Before, not after.

Pricing

Start free. Scale when you're ready.

Open source runtime. Managed platform. Enterprise-grade infrastructure. Choose where you start, and move up when your team does.

Open source
Free
Forever, for everyone
Faramesh Core

Self-hosted runtime enforcement for your agents.

  • Full FPL runtime
  • All 13 framework integrations
  • Non-bypassable enforcement
  • Tamper-evident audit log
  • Elastic License 2.0
  • Community support
View on GitHub
Recommended for teams
Pay as you go
Starts at
$0.80/ 1K decisions
No seats. No minimums.
Faramesh Cloud

Managed infrastructure with visual governance.

  • Everything in Core
  • Visual governance workspace
  • Managed control plane
  • Credential broker
  • Approval workflows (HITL)
  • Shared audit chain
  • Email support
Early access
Enterprise
Custom
Annual contracts
Enterprise

Org-wide deployment with compliance and controls.

  • Everything in Cloud
  • SSO and SCIM
  • Private VPC or on-prem
  • SOC 2, HIPAA, PCI-DSS
  • Multi-approver routing
  • Priority support and SLAs
  • Dedicated engineering contact
Contact us

Every tier includes the full open source runtime. You never lose access to what you started with.

Get started

Ship your agents to production.

Every agent you deploy deserves a policy, an audit log, and a human in the loop.

Open source. Self-hosted or cloud. No credit card required.