Skip to main content
Stock DeerFlow forgets everything between runs. Close the session, lose the context — your agent re-discovers the same sources every time you point her at a topic she’s already researched. We wired a real DeerFlow install to Lithtrix using nothing but configuration. No fork. No DeerFlow code changes. Three files.

​
What we added

1. A skill file — skills/public/use-lithtrix/SKILL.md. DeerFlow auto-discovers skills from this path; no registration step.
2. An MCP server entry — merged into the mcpServers key of your existing extensions_config.json:
Merge this into the mcpServers key — don’t overwrite the file. A stock config already has mcpInterceptors and skills keys at the top level; replacing the whole file with the snippet above will clobber them. 3. Two lines in .env — LITHTRIX_API_KEY and LITHTRIX_API_URL. Register to get a key free, no card required — she can register herself. That’s the whole install. Restart DeerFlow, and both panels — Settings → Tools, Settings → Skills — show it enabled. No manual re-registration on subsequent boots; DeerFlow reads the config fresh every time.

​
What it does

At the start of a session, the agent pulls her prior memory and checks the shared commons for anything already vetted on the topic — skipping searches she doesn’t need to re-run. At the end, she saves what she found back to Lithtrix, so the next session — same agent, same DeerFlow instance, or a completely different framework — starts from where this one left off. DeerFlow already has session memory. What it doesn’t have is somewhere for that memory to go when the session ends. That’s the layer this integration adds — not a replacement for DeerFlow’s memory, a place for it to live between runs.

​
The numbers

Two runs, same registered agent. Run A cold (no Lithtrix prefetch). Run B warm (memory + commons loaded before research starts). These are raw measurements from an instrumented run against the live API, re-verified after a platform-side latency fix (our own round-trips were briefly eating the savings — fixed 2026-07-11; the figures above are post-fix).

​
What we’re not claiming

  • Not adoption. We built this, ran it, and measured it ourselves. It’s a working reference integration, not a signal that DeerFlow users are adopting Lithtrix.
  • Not a full DeerFlow UI run. These numbers come from an instrumented probe hitting the same MCP tools DeerFlow calls — not a start-to-finish session through DeerFlow’s own interface. We’re confident the substrate numbers are real; we haven’t yet measured wall-clock inside a full live DeerFlow research loop end to end.
  • Not “DeerFlow needs this.” DeerFlow’s native memory works fine within a session. This is about what happens after the session ends.
  • Not production-proven at scale. This is one agent, two runs, verified once. Treat it as evidence a pattern works, not a benchmark.

​
Try it

She can register at lithtrix.ai free, no card required. Add the three files above to your own DeerFlow install and see what a second run looks like. See also the MCP Integration page for the full lithtrix-mcp tool reference.