The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the DeepThonk listing page.
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.
Requires Node ≥ 22.13.
Run without installing:
Or install globally:
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.
Create a reusable local config:
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 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:
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.
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:
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.
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.
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.
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:
Restart Claude Desktop after editing.
Project-scoped: .cursor/mcp.json in the workspace. User-global: ~/.cursor/mcp.json. Same JSON shape as Claude Desktop:
Any host that speaks MCP stdio launches:
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.
For local web hosts, or when stdio isn't available:
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.
| Tool | Purpose |
|---|---|
deepthonk.plan | Estimate calls and sequential rounds for a profile (no model calls). Use before paid runs. |
deepthonk.start | Start a run in the background; returns run_dir, job_id, and job-scoped artifact resources. |
deepthonk.status | Poll job status from a run_dir. |
deepthonk.result | Return final summary + winner once a job is complete. |
deepthonk.cancel | Request cancellation by writing cancel.json into the run directory. |
deepthonk.lock_inspect / deepthonk.lock_reclaim | Inspect lock ownership and explicitly reclaim only an exact fingerprint. |
deepthonk.repair_budget | Replace legacy [redacted] numeric budget fields with explicit original values. |
deepthonk.run | Blocking convenience: start + await completion in one call. Prefer start + polling for long-running jobs. |
deepthonk.rank | Rank a user-supplied candidate set with pairwise judging + Bradley-Terry (skip generation). |
deepthonk.mutate | Mutate one supplied candidate with critique (one-shot). |
deepthonk.resume | Detect whether a run can be resumed; with continue: true, replay from the last validated phase boundary. |
deepthonk.export | Export run summary or full trace in JSON or markdown. |
deepthonk.profile_list | List saved named profile bundles. |
deepthonk.profile_show | Show one saved profile; manually edited secret-shaped values are rejected on load. |
deepthonk.profile_save | Save a reusable named profile bundle. |
deepthonk.profile_delete | Delete a saved named profile bundle. |
All tools accept inline provider/model fields, or config_path pointing at a DeepThonk YAML config.
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.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.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.
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.
Every algorithm dimension is reachable through MCP arguments and CLI flags:
n, k, t, m — override profile defaults inline.lambda, sample_temperature, mutate_temperature, judge_temperature.general or paper-programming.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.generate, judge, mutate.provider_max_concurrency sets the recovery ceiling for the shared adaptive provider-route limiter.all-pairs or k-regular, with an explicit logical call cap.Example agent call (MCP), no YAML file required:
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.
A successful run writes:
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.
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:
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.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.
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.