Power your agent
การติดตั้งจากผู้เผยแพร่ ไม่ใช่เอนด์พอยต์ของ PlugMCP
URL แบบรีโมตหรือคำสั่งบนเครื่องของคุณ? การติดตั้งสองแบบต่างกันอย่างไร
รันบน PlugMCP
A new server you create. It is not this listing's official endpoint. Use the publisher install when you want their URL and their credentials.
เปิดจากผู้เผยแพร่
Experience OS for AI employees.
Praxos is not a generic memory layer. It is the flight recorder and learning system for AI agents that do real work.
When an agent completes a task, gets corrected, violates a policy, or produces a bad outcome, Praxos turns that episode into reusable experience: lessons, action checks, policies, and regression evidence.
Humans gain experience. Agents should too.
MCP-powered memory, policy, and experience layer for safer AI agents.
Companies are moving from chatbots to AI employees that answer customers, update CRMs, triage tickets, write code, and operate internal tools. These agents still repeat the same mistakes because their experience is not captured as an operating asset.
Praxos records:
The result is an experience layer that makes agents safer and better over time.
Start with customer-facing AI agents for B2B SaaS.
The first painful workflow:
"Before this support or success agent replies to an enterprise customer, check whether the answer contradicts past commitments, escalations, product decisions, or human corrections."
Praxos can warn or block risky actions before they touch customers.
Episode
A real task attempt: task, action, outcome, feedback, sources.
EvidenceReceipt
Proof for a memory: source URI, snippet, observed_at, confidence, metadata.
Lesson
Reusable experience generated from episodes and corrections.
Policy
A company rule that can warn or block future agent actions.
Business Context
Account, Customer, Commitment, Escalation, Decision.
ActionCheck
A pre-flight decision: allow, warn, or block, with evidence.
ReviewItem
Human approval queue for newly compiled lessons.
pip install -e .
praxos demo --story
By default Praxos stores its ledger in ~/.praxos/praxos.db. In locked-down
environments where the home directory is read-only, it falls back to
.praxos/praxos.db in the current project. You can also set PRAXOS_DATA_DIR
or pass --db explicitly.
The demo writes to an isolated demo workspace each run, so repeated demos stay clean instead of duplicating policy matches.
The story demo prints the investor-facing moment:
Praxos story demo
1. A support agent is about to promise Enterprise A that Feature X ships by Friday.
2. A human previously corrected this exact failure.
3. Praxos compiled that failure into a lesson, evidence receipts, and a review item.
4. Praxos checks the future action before it reaches the customer.
Decision: BLOCK
Record an agent failure and turn it into a lesson:
praxos record \
--agent support-agent \
--task "Reply to Enterprise A asking when Feature X ships" \
--action "Tell them Feature X will ship by Friday" \
--outcome failure \
--feedback "Never promise delivery dates. Product moved Feature X to Q3 in the April 12 escalation."
Add a hard policy:
praxos policy add \
--name "No delivery promises" \
--trigger "ship by friday delivery date promise eta" \
--severity block \
--instruction "Do not promise delivery dates unless there is an approved product source."
Check a future action before the agent sends it:
praxos check \
--task "Reply to Enterprise A about Feature X" \
--action "Say we can deliver Feature X by Friday"
Praxos uses mcp-use to expose its real MCP server. We use mcp-use because it provides a full-stack MCP framework for Python, including server tools via the MCPServer API, while staying compatible with MCP clients.
Reference: https://docs.mcp-use.com/python/server
Start Praxos as an MCP server over stdio:
praxos mcp
# or
praxos-mcp
Start Praxos as an MCP server over streamable HTTP:
praxos mcp --transport streamable-http --host 127.0.0.1 --port 8766
MCP tools exposed through mcp-use:
get_experience
check_action
record_outcome
learn_from_feedback
Example Claude/Cursor-style local MCP config:
{
"mcpServers": {
"praxos": {
"command": "praxos-mcp",
"args": ["--db", ".praxos/praxos.db"]
}
}
}
For debugging without an MCP client, Praxos also ships a small local JSON tool server:
praxos server --host 127.0.0.1 --port 8765
JSON tool endpoints:
GET /tools
POST /tools/get_experience
POST /tools/check_action
POST /tools/record_outcome
POST /tools/learn_from_feedback
Example:
curl -X POST http://127.0.0.1:8765/tools/check_action \
-H "Content-Type: application/json" \
-d '{"task":"Reply to Enterprise A about Feature X","action":"Promise Friday delivery"}'
Example result:
{
"decision": "block",
"reasons": [
"Policy matched: No delivery promises",
"Relevant lesson: Reply to Enterprise A asking when Feature X ships"
]
}
from praxos import ExperienceLedger
ledger = ExperienceLedger(".praxos/praxos.db")
episode = ledger.record_episode(
agent_id="support-agent",
task="Reply to Enterprise A asking when Feature X ships",
action="Tell them Feature X will ship by Friday",
outcome="failure",
human_feedback="Never promise delivery dates. Product moved Feature X to Q3.",
)
check = ledger.check_action(
task="Reply to Enterprise A about Feature X",
action="Say we can deliver Feature X by Friday",
)
print(check.decision)
print(check.reasons)
For the first wedge, Praxos models B2B SaaS customer-facing context:
praxos business account --name "Enterprise A" --external-ref crm://enterprise-a
praxos business commitment --account acct_... \
--description "Do not promise Friday delivery for Feature X." \
--source-uri crm://enterprise-a/commitments/feature-x
praxos business escalation --account acct_... \
--summary "Feature X timing caused a prior customer escalation." \
--severity high
praxos business decision --account acct_... \
--decision "Product moved Feature X to Q3; avoid near-term delivery promises."
When an agent calls check_action, matching commitments, escalations, and decisions can warn or block future action.
Every automatically compiled lesson enters a review queue.
praxos review list
praxos review approve rev_...
praxos review reject rev_...
Rejecting a lesson archives it. Approved lessons remain active and auditable.
By default Praxos uses dependency-free hybrid matching:
For semantic matching, plug in any reranker:
export PRAXOS_RERANK_URL=http://localhost:9000/rerank
export PRAXOS_RERANK_TOKEN=optional-token
The reranker endpoint receives {"query":"...","documents":["..."]} and returns {"scores":[0.0,0.9]}.
Praxos is not trying to remember everything.
It is trying to make AI workers learn from work:
This repo is a clean MVP:
mcp-useNext product steps:
MCP-powered memory, policy, and experience layer for safer AI agents.
เครื่องมือ
ยังไม่มีรายการเครื่องมือ เซิร์ฟเวอร์แบบรีโมตจะได้รายการหลังการตรวจสอบสถานะครั้งแรก
FAQ
คัดลอกคอนฟิก stdio สำหรับ Cursor, Claude Desktop, Claude Code หรือ VS Code ได้ในหน้านี้ คอนฟิกนี้ใช้การติดตั้งของผู้เผยแพร่ ไม่ใช่พร็อกซีของ PlugMCP
ระเบียนนี้ไม่ได้ระบุตัวแปรสภาพแวดล้อมที่จำเป็น
Praxos ทำงานแบบโลคัลผ่าน stdio
ตรวจสอบล่าสุดเมื่อ 2026-10-28 MCP-powered memory, policy, and experience layer for safer AI agents.
