DeepThonk: OpenDeepThink for Agents

When one answer is too brittle, make the model argue with alternatives.

DeepThonk implements the OpenDeepThink algorithm as a provider-neutral CLI and MCP server. It runs a population of candidate answers through pairwise judging, Bradley-Terry ranking, critique-guided mutation, elite preservation, and a final dense ranking pass. The point is not speed. It is spending controlled test-time compute on tasks where breadth plus judgment can beat one expensive single shot.

The package is deepthonk on npm, with dt as the short CLI alias. Current source version: 0.3.0. It requires Node.js 22.13 or newer. Algorithm source: OpenDeepThink, Zhou et al. (arXiv:2605.15177).

Why it matters

DeepThonk spends many model calls by design; plan the budget before running.

Tools

The MCP server is intentionally inspectable. An agent can:

The resume path is conservative: it validates manifests, provider and model identity, artifacts, usage, receipts, and phase commits. Trace-v2 can reuse valid per-call output and usage receipts from an interrupted phase without another provider call, restoring their accounting exactly once. Trace-v1 replays the whole incomplete phase.

Host support

DeepThonk exposes the same engine through a shell CLI and an MCP server. Use the CLI when you want a visible run directory. Use MCP when you want an agent to plan, start, poll, inspect, and export runs from inside its own workspace.

Provider API keys come from the host process environment. DeepSeek is first-class; OpenAI-compatible endpoints, OpenRouter, and host-provided MCP Sampling are also configurable.

Sampling carries three constraints:

• It needs a client that advertises the capability. Standalone CLI runs and mixed Sampling/direct routes are unsupported.
• Blocking run, rank, and mutate work over stdio or stateful Streamable HTTP; HTTP background start rejects Sampling, which must stay attached to the request that initiated it.
• Call caps still apply. Token and USD caps do not, because clients do not report standardized usage.

Install

DeepThonk is published in the MCP Registry as io.github.linxule/deepthonk. Registry-aware hosts resolve that entry to the npm package. Manual MCP configuration must include the serve-mcp argument: the bare deepthonk binary has no default action and exits after printing help.

Run without installing:

npx -y deepthonk plan --profile paper
npx -y deepthonk run --provider fake --profile quick \
  --task "Find the smallest positive integer divisible by 3, 4, and 5." \
  --out runs/test-quick
npx -y deepthonk inspect runs/test-quick

For paid providers, configure a reusable profile first:

deepthonk setup \
  --provider deepseek \
  --api-key-env DEEPSEEK_API_KEY \
  --fast-model deepseek-v4-flash \
  --judge-model deepseek-v4-pro

Then plan before you run:

deepthonk plan --config ~/.config/deepthonk/config.yaml
deepthonk run --task task.md --config ~/.config/deepthonk/config.yaml --profile quick --dry-run

DeepThonk separates phase concurrency from a process-shared provider ceiling. Direct-provider routes start at eight concurrent calls and Sampling starts at four. A provider 429 halves the live ceiling; successful calls restore it gradually, while queued calls remain FIFO and cancellation-aware. Phase and provider concurrency inputs are capped at 1,024.

Every model call also carries an output cap: 4,096 tokens by default for generation, mutation, and finalization, and 1,024 for judging. Final all-pairs ranking schedules above 100 judge calls require an explicit sufficient rank cap; invalid schedules fail before generation spends calls.

For MCP hosts, prefer deepthonk.start plus deepthonk.status/deepthonk.result for long jobs. deepthonk.run is the blocking convenience path. If you need HTTP transport for a local tool bridge, use deepthonk serve-mcp --transport http --port 3333. It binds loopback only and rejects non-loopback request origins. It has no bearer-auth boundary. Do not expose it through a proxy without re-evaluating that trust model.

Part of Research Memex

DeepThonk sits next to Sequential Thinking, not underneath it. Sequential Thinking gives a single model a structured scratchpad. DeepThonk creates and ranks many attempts. Use Sequential Thinking when you want one transparent reasoning chain; use DeepThonk when you want search, mutation, judging, and a trace of alternatives.

It also pairs naturally with Vox. Vox gives you access to many models. DeepThonk gives you a repeatable protocol for spending more compute on a hard task.

Carrel status: Carrel does not install DeepThonk today. Add it manually when a project needs OpenDeepThink-style test-time compute.