Stop your agents
from going rogue
Prismor helps enterprises control their agents' identity, access, and runtime actions
Prismor/one tool call, start to finish
“wire the docs page up to the database and deploy it”
From the first agent you find to the audit you pass
Four jobs, one console. These are the screens your security team works in, and the CLI your developers already run feeds every one of them.
Dashboard/Shadow AI
Dashboard/Approvalscodex · Linux desktop · headless
also in #infra-approvalsmcp__postgres__query on prodcustomer_piiAlice Chen · 2m
send_email to an outside domainuntrusted inputsupport-bot · 6mDashboard/Overviewcursor4claude-code4cursor6
langchain-api2cursor2Dashboard/Proof Pack
Datadog·every row links to its session trailFind every agent before you write a rule
Each enrolled machine reports what it actually runs: the coding agents installed, the MCP servers they connect to, and the skills, plugins and hooks loaded into them. The fleet view splits what Prismor governs from what it does not, and a machine that has never been scanned shows as unknown, not as clean
Agents, MCP servers and extensions across the fleetDecide what each agent can reach, and for whom
Policy binds to the person, the agent and the device. Scope tools and paths per identity, front MCP servers with one gateway, and allowlist where agents may send traffic. Rules live in git and ship as a signed device policy. When a headless agent reaches for something risky it pauses, and an approver sees the exact command and the conversation behind it
Identity-bound policy, egress allowlists, step-up approvalsStop the bad call, not the whole session
Every tool call is checked before it executes. Prismor blocks destructive commands, catches untrusted input flowing into a critical action, and cloaks secrets so the model only ever handles placeholders. The agent is told why it was stopped, so it can take a safer path instead of failing the task. Works with Claude Code, Cursor, Codex, Copilot, LangChain and the OpenAI Agents SDK
Every tool call intercepted, cloaked or blocked before it runsGive your auditor evidence, not a slide
Every session keeps its full tool-call trail: arguments, verdict, rule, user and device. The Proof Pack turns a date range into a report for a customer security review or your board, covering enrollment, interventions and approvals, each mapped to the OWASP Agentic Top 10. The same events stream to your SIEM as OCSF
Session trails, Proof Pack exports, OCSF to your SIEMOne security layer for every agent your team uses
Connect Prismor once and keep the same runtime policy, secret protection, and audit trail across coding agents, SDKs, and custom MCP workflows.
A closed loop that gets safer with every decision
Export agent activity to your observability stack, involve a human when policy needs context, then turn approvals, denials, and incident outcomes into better policy recommendations.
Send traces, policy outcomes, and audit events into the tools your security team already monitors.
Datadog
OpenTelemetry
New Relic
Splunk
GrafanaNotify the right owner, request approval, or route an exception without stopping the entire workflow.
SlackPrismor uses guided decisions and recurring patterns to recommend tighter rules, reviewed before they are applied.
What developers are saying
AI agents can run commands before you get a chance to review them. This repo adds a checkpoint
Your agent is about to call a tool and you have no pause button in the session
Your guardrail can warn you and still wave the command through. Prismor's docs make the distinction explicit: observe only logs; enforce can block, but warn/log actions still allow the call
a seatbelt for claude code
Looking for a promising tool: Prismor. Tell me something is wrong in the agentic coding landscape without telling me something's wrong
open-source tooling that answers a real bill question always finds users
if you use Codex or Claude Code, you probably need this
bookmarked, thanks for sharing
Coding agents can run shell commands, install packages, call MCP tools and touch your secrets. Maybe we should watch them
Love the MCP gateway angle. One policy layer across everything is the right abstraction
bro this is actually clean, i've had claude code almost run rm -rf on my server once. policy checks would've saved me
the observe mode is the part i'd want tested: how often does it flag legitimate shell commands, and who maintains those policies when the repo changes?
Honestly, AI trading bots need this more than humans
gm, open source control plane that gates every tool call sounds like the missing guardrail. how heavy is the latency?
enforce mode saves wallets tbh
the enforce vs observe switch is smart! teams can see the blast radius before
This is a real gap. The agent part gets all the attention, but visibility and control are what make it usable in production
nice take on agent iam, the guardrail scoping feels solid. how do you handle secret rotation across delegated sub agents?
Local first security for agents is a strong direction. Secret redaction is table stakes, but policy enforcement around tool calls and package installs is where this gets really interesting
this is genuinely needed. most agent stacks rn are a black box once you deploy them, and the "we'll add observability later" crowd finds out the hard way when something exfiltrates a token
Open source is the way to go! Superb demo and great product
super cool product!
Tried it for 3 of my repositories. Super awesome man!
This is awesomeee
I think this is a very solid approach
too good
Hats off for providing us with this!!
security is key especially as models become increasingly capable!
As enterprises are adapting agents into their workforce, security becomes very crucial part of the workflow. Looking forward for great work
so cool!
From agent chaos to auditable control
Connect
Wrap your agent framework with Prismor in one step. MCP, the OpenAI Agents SDK, LangChain, CrewAI. Your existing agents keep working.
Intercept
Every tool call routes through Prismor before it executes. The agent has no path around it.
Enforce
Your active policy decides each call: allow, block, or mask. Secrets never reach the model. Destructive commands never run.
Audit
Read the full timeline of every agent action in the dashboard: arguments, outcome, user, session. Export it when compliance asks.
Improve
Update policies from the CLI as your conventions change. Every rule is versioned and reviewed by a teammate, same as infra config.
