Everything you need to build agentic systems.
The agent, the MCP servers it works with and the runtime that runs it. One open-source framework, made to work together from the first line.
Built together, not stitched together.
Most teams stitch an agent, its tools and a runtime together from separate libraries. Promptise ships all three in one install.
Glue code at every joint.
Each library has its own interfaces. Your team writes the code between them and keeps it working.
One install. One set of interfaces.
The agent, the interface and the harness are made to work together from the first line.
Three parts, and what's inside each.
The agent does the work. The interface connects it to your systems. The harness runs it.
Smart and reliable.
Our reasoning and execution engines turn any model into an agent that thinks before it acts. It works out the right approach for each task and checks its own result. You set the limits. It finds the way.
Any system, agent-ready through MCP.
Build MCP servers for your own systems and connect to any that exist. Turn the API you already have into one an agent can use.
Autonomous and under control.
Our agent runtime starts agents on a schedule or an event and keeps them working around the clock. They stay inside the limits you set, and you can see everything they did.
Every module, and what it does.
Twelve modules in three parts. Start with the one you need and add the others later. Pick a part, then a module.
The Promptise Agent
One function returns a production agent. Any model, tools discovered from your MCP servers, memory searched before every turn, and guardrails, caching and tracing switched on by a flag.
- 3
- memory providers
- 4
- conversation stores
- 4
- tool-selection levels
- Any model: a provider string or any LangChain chat model
- MCP tool discovery, with semantic tool selection
- Memory in Chroma, Mem0 or in-memory
- Conversations persisted to SQLite, Postgres or Redis
- Per-request caller identity through cache, guardrails and tracing
from promptise import build_agentfrom promptise.config import HTTPServerSpecfrom promptise.memory import ChromaProvider agent = await build_agent( model="openai:gpt-5-mini", servers={"crm": HTTPServerSpec(url="http://localhost:8000/mcp")}, memory=ChromaProvider(), guardrails=True, observe=True,)Production-ready from the first line.
Eight capabilities come built in. You switch them on when you need them. Nothing has to be added later.
- 1Governance & controlsBudgets · Approval gates · Health checks · Missions
- 2AutonomyTriggers · Journals · Crash recovery
- 3SafetyInjection · PII · Credentials · Content safety
- 4Identity & auditAttribution · Verifiable credentials · Audit log
- 5Tools & integrationMCP client · MCP servers · OpenAPI import
- 6ObservabilityTracing · Metrics · Eight destinations
- 7ContextMemory · Conversation stores · Token budgets · Semantic cache
- 8ReasoningReasoning graph · Strategies · Per-node models
What each capability does+
- Context
- Answers grounded in what the agent knows, without overflowing the context window or the budget.
- Reasoning
- Reasoning designed for the job, so results are predictable instead of left to one generic loop.
- Tools & integration
- Connect to any system through the open MCP standard, and publish your own APIs the same way.
- Safety
- Prompt injection, personal data and leaked credentials are caught before the model sees them, and again on the way out.
- Governance & controls
- Spending limits and approvals enforced in code, not in the prompt, with a person signing off where it counts.
- Autonomy
- Agents run on their own schedule and pick up where they left off when something breaks.
- Identity & audit
- Know which agent acted, and prove it to someone who was not there.
- Observability
- Every model call, tool call and retry traced, and sent to the tools your team already uses.
Watch one request run through the system.
A customer writes in at three in the morning. Follow the ticket step by step: what wakes the agent, what it knows, how it reasons, what it calls, who signs off and what it records.
Open the full walkthroughStep 01 / 14 · 03:02:11 · Agent Runtime
A ticket wakes the agent
A helpdesk webhook fires. The runtime wakes the support agent's process and hands it the ticket.
An illustrative run. Names and timings are examples; what each step describes is what the framework does. See every module in the full walkthrough →
Install it. Run your first agent.
Python 3.10 or newer. No scaffolding and no project generator: it works inside the code you already have.
$ pip install promptise # or, with vector memory, ML guardrails and exporters$ pip install "promptise[all]"promptise · verb
Let's promptise it. Whatever it is.
Like digitising something, promptising it changes how the work gets done. An agent thinks it through, works in your real systems, runs on its own and stays inside your rules, and you can see every step it took. Anything can be promptised.
What promptised meansPromptised means it
- It thinksreasoning you design
- It knowsmemory and your documents
- It actstools, APIs and MCP
- It works with othersagents that delegate
- It runs on its owntriggers and recovery
- It stays safeguardrails, approvals, budgets
- It shows its worktraces and a journal
- It shipsone command to production
Built it? Two ways to go further.
The framework is yours to run. When you want more, there are two ways to get it.
Deploy with one push.
The platform will deploy, run and monitor what you build, in our cloud or on your own servers.
Want it built with you? Hire our engineers.
The people who build Promptise work with your team, from the first prototype to production.
Build with the framework. Run it on the platform. Launch it with our engineers.
The framework is open source and free. The platform is coming, with early access now. Our engineers are ready when you are.