Louis Choquel (@lchoquel) header
Louis Choquel (@lchoquel) profile picture

Co-founder & CTO at Pipelex • Declarative language for composable Al workflows. Devtool for agents and mere humans. 🚀 https://teapaw.org/t.co/OjTMGCReBL

Paris
Joined October 2007
Photo by Louis Choquel (@lchoquel)Photo by Louis Choquel (@lchoquel)Photo by Louis Choquel (@lchoquel)Photo by Louis Choquel (@lchoquel)Photo by Louis Choquel (@lchoquel)Photo by Louis Choquel (@lchoquel)Photo by Louis Choquel (@lchoquel)Photo by Louis Choquel (@lchoquel)Photo by Louis Choquel (@lchoquel)
Louis Choquel retweeted
Something quietly important happened in agent tooling this week and it is not another model release. Tom Dorr posted about Pipelex, a project that lets you declare AI procedures in a typed file format called .mthds instead of writing free form prompts or brittle glue code. Each step is explicit, every input and output is typed, and the pipeline is deterministic outside the parts where you actually want the model to think. Agents write the .mthds file, a human reviews it, Pipelex runs it. That review step is the whole point. A dozen teams are independently converging on this same problem: agents need a format they can reliably generate and re-execute, not vibes based prompting that drifts every time you touch it. LangGraph solved a version of it with graph based state machines and passed CrewAI in GitHub stars this year on the strength of audit trails and rollback points for compliance heavy workflows. Temporal and Dagster solved an earlier version for plain software pipelines, durable execution that survives a crash halfway through. Atlassian shipped Agentic Pipelines inside Bitbucket this year for the same reason, chores nobody wants to hand babysit, wrapped in something repeatable. What is different about Pipelex is the packaging. MTHDS is proposed as an open standard, with a language for typed concepts and pipes plus a package system with lock files and a federated registry, the same shape as npm or PyPI but for AI methods instead of code libraries. I keep coming back to a pattern from the early API economy: nobody won by having the best individual endpoint, they won by having the best schema everyone else agreed to build against. Stripe won payments on developer experience and a stable contract, not on being the only company that could move money. The bottleneck agent tooling has right now is not capability, it is repeatability. My read is that whoever defines the format teams standardize on for declarative agent workflows ends up owning more of the stack than whoever ships the flashiest autonomous demo. teapaw.org/tom_doerr/status/21029…
Pipelex executes typed AI procedures declared in .mthds files. github.com/Pipelex/pipelex
5
2
1
6
671
We're no longer the only intelligent species on this planet
2
1
53
Louis Choquel retweeted
Agent workflows get harder to maintain when the logic is trapped in prompts. Pipelex is a declarative language and tool for builders who want to define reusable, typed AI methods. It helps you build repeatable AI workflows by declaring inputs, outputs, prompts, and pipeline steps in human-readable `.mthds` files, then executing them with Pipelex. Key features: • Typed methods – define semantic input and output types so workflow connections are explicit • Composable pipelines – chain pipes into sequences and nest methods to build larger procedures • Model routing + structured parsing – Pipelex handles model routing and structured-output parsing for methods • Multiple run targets – run the same `.mthds` file from the CLI, Python, REST API, MCP, or n8n • Agent and IDE support – the README documents Claude Code and Codex setup plus a VS Code extension for syntax highlighting and flowcharts It’s open-source (MIT license). Link in the reply 👇
6
26
133
5,932
Louis Choquel retweeted
Build and run AI methods as executable pipelines. github.com/Pipelex/pipelex
1
8
51
6,478
The problem when AI rewrites your prose with 'The blah blah nobody's talking about' is that the AI didn't get off its ass and do a search to check whether anybody is talking about it.
1
69
4 months of heads-down building. Biggest Pipelex release yet. New open language spec called MTHDS. A Claude Code plugin that writes, tests, and visualizes AI workflows for you. One API key for all models. A package manager to share it all. Full drop 🧵
2
4
1
4
419
And the long tail: Pydantic structures generated from inline definitions in .mthds files PDFs and docs directly in LLM prompts Deterministic compose mode (no LLM needed) Image-to-image generation S3 & GCS storage Langfuse & OpenTelemetry observability OpenRouter, Hugging Face, Scaleway, Portkey backends Python 3.14 support
1
1
1
54
The bet we're making: AI workflows should work like software, not like art projects. Typed. Testable. Shareable. Composable. That's what MTHDS is. We're around all day if you want to try it. What would you build first? mthds.ai mthds.sh app.pipelex.com
1
1
22
Louis Choquel retweeted
Massive divide I’m seeing: A) Startups where the founder hands-on, building with the latest AI tools and best models, sees first-hand what this means and championing everyone to use it, not caring about $$$ B) founder not engaged, devs still think AI (aka Copilot) is “meh”
107
100
21
1,760
210,504
That's why we built @PipelexAI : the declarative language for AI workflows
Palantir CEO Alex Karp says off-the-shelf LLMs fail in enterprise. “Once you build a software layer to orchestrate and manage the LLMs in a language your enterprise understands, you actually can create value.”
1
67
Without Know-How, your agent is a zombie 🧟‍♂️. A Knowledge Graph isn't enough. → We need a Know-How Graph ‼️ ⏩ A network of executable methods that AI can actually reuse. Read our Viewpoint 👇 ➡️ knowhowgraph (dot) com
1
2
3
143
Launch Week from SF: That's a wrap 🚀 We shipped: • Pipelex language + OSS API server • MCP server • n8n node • Agent Rules Results: 🏆 DevHunt #1 Tool of the Week 📰 HackerNews frontpage ⭐ 500+ GitHub stars 👥 50+ devs at our first hackathon, 15 projects built in 2.5hrs Huge thanks to everyone who supported us. This milestone means everything to us. Major features dropping soon 👀 ⭐ Star our repo (link below) 👇
2
2
1
5
768