DrawloomMost problems Drawloom faces have been solved before. Before we build anything, we look at how others solved them.
How we investigate
- Find the best examples. We look for open-source projects that solve the same problem, ideally in different ways.
- Read the code. We study how they actually work, down to the code.
- Run a small experiment. We try the idea in a separate prototype, called a spike, before it goes anywhere near Drawloom.
- Decide. We choose what fits Drawloom best. See Decisions.
We’ve published our studies of agent harnesses and workbenches, knowledge and memory, evaluation and access control.
Thank you
Drawloom is built on work that other people chose to share. Thank you.
What Drawloom runs on
- Temporal: keeps long-running work going through crashes and restarts.
- Tauri: the Mac app itself.
- Node.js and pnpm: run and package the app.
- Svelte, shadcn-svelte and Bits UI: every screen.
- Model Context Protocol: connects agents to tools.
- SQLite and sqlite-vec: storage and search, on your own computer.
- Zod: checks data wherever it enters.
- Vitest and Biome: testing and formatting.
- Astro: this website.
Agent harnesses and workbenches
- OpenAI Codex: the agent Drawloom works with today.
- DeepSeek Harness: how to keep a full conversation history alongside a lighter view for the screen.
- Open Design: how to build a rich product around agents you don’t own.
- MCP Apps: the standard Drawloom uses for plugin screens.
Knowledge and memory
- Hindsight: keeping evidence separate from summaries.
- Graphiti: marking a fact as out of date without erasing it.
- Letta Code: knowing when work has safely finished.
- Mem0: checking that every save actually worked.
- A-Mem: notes that update related notes.
- HippoRAG: removing a source along with everything learned from it.
- LongMemEval: testing long-term memory.
- Sleep-time Compute: preparing context while the agent is idle.
- MINJA: how shared memory can be attacked.
- SEPIO and W3C PROV-O: recording where a fact came from.
Evaluation
- Promptfoo: running and comparing test cases locally.
- Arcade MCP: testing whether an agent picks the right tool.
- DeepEval: measures for retrieval, agents and conversations.
- Langfuse: comparisons, feedback and background evaluation.
- Braintrust and Autoevals: the scoring Drawloom uses.
- LangSmith: linking scores to what the agent actually did.
Access control and monitoring
- NIST SP 800-162: the model behind how Drawloom decides who sees what.
- OpenID AuthZEN: a standard way to ask “is this allowed?”.
- Cedar and Casbin: policy engines we tested side by side.
- OpenTelemetry: seeing what the agent did, without recording your conversations.
- Archify: our architecture diagrams.