DeepThonk

thonk harder, not richer.

DeepThonk implements the OpenDeepThink algorithm (Zhou et al., 2026, arXiv:2605.15177) as a budget-friendly, provider-neutral "deep think / pro mode" wrapper, with first-class DeepSeek support. See Acknowledgments.

DeepThonk runs a population of candidate answers through pairwise judging, Bradley-Terry ranking, critique-guided mutation, elite preservation, and a final dense ranking pass. The CLI and MCP server both call the same TypeScript core engine.

Designed for agents. Every algorithm dimension — population shape (n, k, t, m), regularization (lambda), per-phase temperatures, prompt style, and per-phase prompt templates — is reachable inline through MCP arguments and CLI flags. CLI can load prompt files with --prompts or inline JSON with --prompts-json; MCP accepts inline structured prompt args. Every intermediate artifact (config, candidates, populations, comparisons, scores, per-call usage, status) is exposed as an MCP resource so an agent can inspect any step. See Customization for the full agent-composable surface.

Use it for hard, verifiable reasoning, coding, planning, and synthesis where breadth plus judgment can beat one expensive single shot. Avoid it for highly subjective tasks where judge noise dominates.

Quickstart

Requires Node ≥ 22.13.

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

Or install globally:

npm install -g deepthonk
deepthonk plan --profile paper

The paper profile plans 285 model calls and 8 sequential rounds. Confirm budget and provider pricing before pointing it at paid models. The short alias dt is installed alongside deepthonk. Develop from source: see Development.

DeepThonk is an independent TypeScript reimplementation and practical integration layer; the published algorithm is the load-bearing part. Cost guarantees rely on provider pricing being present in config. Trace-v2 resume can reuse validated per-item receipts after a crash, while older traces replay an incomplete phase wholesale. The MCP HTTP transport remains loopback-only. The included acceptance smoke uses the deterministic fake provider and a toy task — it is not a CF-73 / HLE reproduction; see docs/ and Acknowledgments for the canonical paper and Python reference.

Related MCP server: Deepseek MCP Server

Setup

Create a reusable local config:

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

By default this writes ~/.config/deepthonk/config.yaml. deepthonk run loads that file automatically when --config is not supplied. If you pass --api-key, setup stores it in ~/.config/deepthonk/env; otherwise it uses the named environment variable from your shell.

DeepSeek

export DEEPSEEK_API_KEY=...
deepthonk run \
  --task task.md \
  --profile paper \
  --provider deepseek \
  --generator-model deepseek-v4-flash \
  --mutator-model deepseek-v4-flash \
  --judge-model deepseek-v4-pro \
  --out runs/task-paper

DeepSeek is implemented as an OpenAI-compatible profile using https://api.deepseek.com/v1. DeepThonk ships default USD pricing for deepseek-v4-flash and deepseek-v4-pro from the official DeepSeek pricing page, including cache-hit/cache-miss input rates. Model names and prices are still editable config because both can change.

Before paid runs, inspect cost shape and resolved config:

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

Start with --profile quick and consider --max-concurrency, --max-calls, --max-input-tokens, --max-output-tokens, --max-usd, and --request-timeout-ms before larger paid profiles. max_calls is reserved before dispatch and counts logical model invocations, including failed calls and invalid-JSON retries; provider-internal HTTP retries are reported separately. Token/USD totals are known only after responses and can overshoot by at most the active concurrency window. Plans keep nominal calls separate from finalizer and retry headroom in worst_case_calls.

OpenAI-Compatible Providers

export DEEPTHONK_API_KEY=...
deepthonk run \
  --task task.md \
  --profile balanced \
  --provider openai-compatible \
  --base-url https://provider.example.com/v1 \
  --api-key-env DEEPTHONK_API_KEY \
  --generator-model cheap-model \
  --judge-model strong-model

The driver calls POST {base_url}/chat/completions and requests JSON mode for comparisons when supported. If a provider returns 400 or 422 with a body that mentions response_format or json, the driver retries without JSON mode, coordinates that capability probe across concurrent calls, and remembers the result for the process. Set supports_json_mode: false in YAML to disable JSON mode up front.

Provider names are flexible. For a custom OpenAI-compatible endpoint, use any provider label with --base-url, --api-key-env, and role-specific model flags. For OpenRouter:

export OPENROUTER_API_KEY=...
deepthonk run \
  --task task.md \
  --profile balanced \
  --provider openrouter \
  --generator-model openrouter/auto \
  --mutator-model openrouter/auto \
  --judge-model openrouter/auto

For mixed-provider runs, use YAML config and override individual roles under providers, especially judge.

Optional finalizer_model / --finalizer-model can post-process the ranked winner. Leave it unset when you want the raw ranked answer as the final artifact.

MCP

The MCP server exposes the same engine the CLI runs. Once wired into an MCP host, the host can plan budgets, kick off background runs, poll status, fetch winners, and stream structured trace artifacts — all through MCP tools, resources, and prompts.

DeepThonk is listed on the MCP Registry as io.github.linxule/deepthonk (since v0.2.1). Hosts that install by registry name resolve it to npx deepthonk serve-mcp automatically. For hosts that don't, the explicit configs below do the same thing by hand.

Provider API keys come from the host process's environment, not from DeepThonk's config alone. Each host handles env passthrough slightly differently — see the concrete patterns below. No key is required to start the server: with a sampling-capable host you can run provider: "sampling", and the fake provider needs no network at all.

MCP Sampling is supported as provider: "sampling" when the connected host advertises the MCP sampling capability. Blocking run, rank, and mutate work over stdio and stateful Streamable HTTP. HTTP background start rejects Sampling because nested Sampling must remain attached to the initiating request; stdio background runs and direct-provider HTTP background runs remain available. Sampling is not available from standalone CLI runs.

Claude Code

claude mcp add deepthonk \
  -e DEEPSEEK_API_KEY=$DEEPSEEK_API_KEY \
  -- npx -y deepthonk serve-mcp --transport stdio

Use -s user for cross-project scope or -s project to commit registration into .claude/. Verify with claude mcp list. See claude mcp --help for the authoritative flag set.

Claude Desktop

Config path: ~/Library/Application Support/Claude/claude_desktop_config.json (macOS), %APPDATA%\Claude\claude_desktop_config.json (Windows), ~/.config/Claude/claude_desktop_config.json (Linux). Add:

{
  "mcpServers": {
    "deepthonk": {
      "command": "npx",
      "args": ["-y", "deepthonk", "serve-mcp", "--transport", "stdio"],
      "env": {
        "DEEPSEEK_API_KEY": "sk-..."
      }
    }
  }
}

Restart Claude Desktop after editing.

Cursor

Project-scoped: .cursor/mcp.json in the workspace. User-global: ~/.cursor/mcp.json. Same JSON shape as Claude Desktop:

{
  "mcpServers": {
    "deepthonk": {
      "command": "npx",
      "args": ["-y", "deepthonk", "serve-mcp", "--transport", "stdio"],
      "env": { "DEEPSEEK_API_KEY": "sk-..." }
    }
  }
}

Other MCP hosts (stdio)

Any host that speaks MCP stdio launches:

npx -y deepthonk serve-mcp --transport stdio

with the relevant provider env vars set on the process. If you'd rather install globally, npm install -g deepthonk then use command: "deepthonk" directly. Refer to the host's MCP registration docs for the configuration shape.

Streamable HTTP

For local web hosts, or when stdio isn't available:

deepthonk serve-mcp --transport http --port 3333 \
  --max-active-jobs 2 --max-queued-jobs 32

The server binds 127.0.0.1:3333 only and exposes stateful POST, GET, and DELETE at http://127.0.0.1:3333/mcp. Sessions use cryptographic IDs, expire after 30 idle minutes when no request is active, and are capped at 64. DNS rebinding protection is on (CVE-2025-66414): requests with Host headers outside 127.0.0.1:3333 / localhost:3333 are rejected. The wrapper also rejects non-JSON POSTs, non-loopback Origin headers, and Sec-Fetch-Site: cross-site before reading the body. It has no bearer auth. Do not expose this port through a reverse proxy without re-evaluating that trust boundary.

Tools

All tools accept inline provider/model fields, or config_path pointing at a DeepThonk YAML config.

Resources

• deepthonk://runs — JSON index of all runs in runs/.

• deepthonk://runs/{run_id}/summary — run summary (JSON).

• deepthonk://runs/{run_id}/config — redacted run config (JSON).

• deepthonk://runs/{run_id}/{candidates|comparisons|scores|usage|trace} — per-phase and per-call NDJSON.

• deepthonk://runs/{run_id}/population/{generation} — population snapshot for a generation (JSON).

• deepthonk://runs/{run_id}/{winner|final} — text artifacts.

• deepthonk://runs/{run_id}/status — run state (JSON).

• deepthonk://runs/{run_id}/prompts/{sha256} — verified content-addressed rendered prompt (JSON).

• deepthonk://runs/{run_id}/trace-v2 — bounded phase index; follow its manifest, commit, checkpoint-index, and checkpoint URIs to inspect every receipt.

• deepthonk://jobs/{job_id}/{status|result|config|candidates|comparisons|scores|usage|trace|final|winner}?run_dir=... — job-scoped lookup; the run_dir query param is required.

• deepthonk://jobs/{job_id}/population/{generation}?run_dir=... — job-scoped population snapshots before or after completion.

• Job results also include prompt_template and trace_v2_template for the job-scoped equivalents.

• deepthonk://runs/{run_id}/{resource}/page/{cursor} and job equivalents — opaque bounded pages. Whole reads are limited to 1 MiB; pages contain at most 1,000 records and 1 MiB.

Prompts

Four templates mirror the core loop: deepthonk/generate, deepthonk/compare, deepthonk/mutate, deepthonk/finalize. Hosts can render them directly to drive the algorithm by hand without invoking the tool surface.

Limits

deepthonk.resume reports trace state by default. With continue: true, it validates phase order, artifacts, run/provider/model identity, scores, usage, deterministic pair schedules, trace-v2 receipts, and phase commits. Valid incomplete-phase receipts are reused without another provider call. MCP Sampling model hints are preferences and token usage may be unavailable. Its shared adaptive limiter starts at 4; direct providers start at 8.

Customization

Every algorithm dimension is reachable through MCP arguments and CLI flags:

• Population shape: n, k, t, m — override profile defaults inline.

• Algorithm constants: lambda, sample_temperature, mutate_temperature, judge_temperature.

• Prompt style: general or paper-programming.

• Per-phase prompt templates: override generate, compare, mutate, or finalize with custom system/user templates and variable substitution ({task}, {rubric}, {candidate}, {candidateA}, {candidateB}, {critique}). CLI supports --prompts <yaml> and --prompts-json <json>; MCP supports inline prompts. Unknown variables throw a fail-fast error at run-start.

• Concurrency: per-phase caps for generate, judge, mutate.

• Provider backpressure: provider_max_concurrency sets the recovery ceiling for the shared adaptive provider-route limiter.

• Per-call output bounds: independent generation, mutation, judge, and finalizer token caps.

• Final ranking schedule: seeded all-pairs or k-regular, with an explicit logical call cap.

• Critique bound: deterministic deduplication and a configurable aggregate-character cap before mutation.

Example agent call (MCP), no YAML file required:

{
  "task": "Draft a concise non-solicitation clause for a senior sales employee.",
  "profile": "balanced",
  "n": 6, "t": 1,
  "provider": "deepseek",
  "judge_model": "deepseek-v4-pro",
  "prompts": {
    "generate": { "system": "You are an experienced employment-law attorney." },
    "compare":  { "system": "Prefer enforceable clauses under California law. Return strict JSON only." }
  }
}

CLI accepts the same surface. Use --prompts <yaml> for reusable prompt files or --prompts-json <json> for one-off inline overrides. MCP and CLI both merge over any --config/config_path YAML defaults.

Save reusable bundles as named profiles at ~/.config/deepthonk/profiles/<name>.yaml and load them with --profile-name <name> (CLI) or profile_name: "<name>" (MCP). A named profile is a standalone bundle that replaces the main config file for that run; CLI flags and MCP inline arguments still override fields inside it. See examples/profiles/legal-drafting.yaml for the shape. The Customization guide covers listing, showing, saving, and deleting named profiles from both CLI and MCP.

See the Customization guide for the complete variable contract, the compare-phase JSON safety rule, the named-profile schema, and three worked examples. Per-role provider routing (providers.judge.provider = openrouter) and per-model pricing remain YAML-only — they're nested structured config, not inline ergonomics.

Trace Files

A successful run writes:

runs/{run_id}/
  config.json
  events.jsonl
  candidates.jsonl
  comparisons.jsonl
  scores.jsonl
  usage.jsonl
  population-0.json
  population-{generation}.json
  manifests/*.json
  checkpoints/*/*.json
  commits/*.json
  summary.json
  artifacts/winner.txt
  artifacts/final.txt
  artifacts/prompts/<sha256>.json  # only when prompt capture is enabled

status.json is written when the run is launched through MCP deepthonk.start (so async clients can poll). cancel.json is written only when cancellation is requested. A run.lock file is held for the duration of the run.

Trace rows are batch-appended through a single writer and flushed before lifecycle boundaries. Per-item output and usage receipts are written atomically before a phase commit. If a run is killed mid-tournament, trace-v2 resume validates and reuses completed receipts, re-materializes their JSONL rows, and calls the provider only for missing or invalid work.

usage.jsonl carries one row per provider call (generator, judge, mutator, finalizer) with phase/role/provider/model/token/USD/latency/retry. It contains no prompt content. jq over usage.jsonl is the simplest way to break a run's cost down by role.

Prompts and raw model outputs are off by default. When prompt capture is enabled, candidate/comparison rows carry a promptRef and identical rendered prompts share one content-addressed blob under artifacts/prompts/; the full prompt is not repeated in every JSONL row or receipt. API keys are never written. Candidate answers, critiques, scores, and final artifacts are trace data and are stored/exported by design, so do not use a shared run directory for sensitive tasks unless that is acceptable.

Development

pnpm install
pnpm run build
pnpm test
pnpm run lint

This repository is source-only. Generated run traces, paper notes, build output, coverage, logs, and node_modules are intentionally ignored.

Release publishing is tag-triggered through npm Trusted Publishing. See docs/release.md before bumping versions or pushing a v* tag.

Known limits for this release:

• Trace-v1 resume replays an interrupted phase wholesale; per-item reuse requires a trace-v2 run created by DeepThonk v0.2.

• HTTP Sampling is blocking-only. Use stdio or a direct provider for background deepthonk.start work.

• maxUsd requires known model pricing; add explicit prices in YAML for custom providers and model IDs. Optional fields longContextThresholdTokens, inputUsdPerMillionLong, and outputUsdPerMillionLong enable tiered pricing for models with a long-context surcharge (e.g., Gemini at 200K input tokens). Cache hit/miss rates remain flat.

Acknowledgments

DeepThonk is an independent TypeScript reimplementation of the OpenDeepThink algorithm. It is not a fork; no source code from the reference implementation is vendored. Credit for the algorithm, the benchmark methodology, and the empirical results belongs to the paper authors.

• Paper: Shang Zhou, Wenhao Chai, Kaiyuan Liu, Huanzhi Mao, Qiuyang Mang, Jingbo Shang. OpenDeepThink: Parallel Reasoning via Bradley–Terry Aggregation. arXiv:2605.15177, 2026. https://arxiv.org/abs/2605.15177

• Reference Python implementation: https://github.com/ZhouShang0817/open-deep-think (MIT). This is the authors' code release and the canonical implementation for reproducing paper results (CF-73, HLE).

DeepThonk's contribution is an integration layer: a provider-neutral TypeScript core, an MCP server, a CLI, structured trace artifacts, and budget enforcement. The algorithm itself is theirs.