Hunch MCP server

Read from the package source without running it. The installed server may list more.

• hunch_blast_radius

  Given a file you're about to change, return its dependency blast radius (files whose code depends on it) AND any invariants reached THROUGH that radius — 'near-violations' you could break indirectly without touching their own scope. Call before editing a widely-depended-on file. Mirrors `hunch check --blast`. Not for a bare dependent list (hunch_get_dependents) or constraints scoped to the target

• hunch_bug_lineage

  Given a symptom description or a symbol, return matching bugs with their lineage (introduced → fixed → recurrence) so the agent doesn't re-discover past root causes.

• hunch_capture_decision

  Start a decision-capture interview: returns the grilling protocol (interrogate ONE question at a time until the decision tree is resolved) plus a capture-session token. Grill the human, then commit via hunch_record_decision with the token + confirmed topic. Use for '/capture', 'record this decision', 'grill me on this'. The token proves the write is the tail of an interview, not a silent guess. Re

• hunch_change_identity

  Content-address the exact Git tree delta between two revisions. Commit messages, authors and squash metadata do not affect the identity; whitespace, paths, modes and blob bytes do. Read-only and deterministic.

• hunch_change_proof

  Bind an exact committed Git transition to its change identity, Project DNA, base/result semantic graphs, current decisions and constraints, blast radius, conformance, guard verdict, and explicit gaps. Read-only and deterministic; grants no execution, CI, deployment, merge, ranking, promotion, or policy authority. Needs two committed refs. Not for a verdict on staged work (hunch_merge_verdict) or a

• hunch_check_constraints

  Return constraints whose scope matches a glob/path, sorted by severity. Call this BEFORE editing code to avoid breaking intentional invariants. Returns each constraint's id, severity, enforcement, statement, and rationale. Not for who-depends-on-this (hunch_get_dependents) or invariants reachable only through dependents (hunch_blast_radius).

• hunch_compare

  Given several candidate branches/commits (e.g. N solutions to one task), replay each against engineering memory and RANK them best-fit first — the candidate that trips the fewest in-force invariants, reverses no decisions, and adds the least sprawl wins. Deterministic (the same merge-verdict per candidate, no LLM). Use to choose among multiple solutions before committing to one.

• hunch_conformance

  Intent-conformance (the inversion of a normal guard): for every in-force decision carrying a conformance predicate, deterministically verify the CODE still satisfies its intent over the dependency graph — e.g. 'pay still reaches verifySession'. Returns the violations: intent the code has silently drifted away from, with NO diff required. Run before a refactor or merge to catch intent erosion a dif

• hunch_constitution_g2_behavior_candidates

  Derive a bounded private review packet from human-grounded rejected structural proxies and newly added literal node:test cases in their exact fixing commits. Read-only: candidates remain unselected and create no policy, corpus, proof, authority, warning, or block.

• hunch_constitution_g2_behavior_materialization

  Bind the complete current private behavior review and exact selected attestations, then report whether their durable meanings are expressible by the supported Policy IR. Read-only and fail-closed: unsupported behavior creates no policy, corpus, plan, proof, authority, warning, or block.

• hunch_constitution_g2_behavior_policy_materialize

  Materialize every current exact selected behavior attestation into a separate private Policy IR v2 proposal, exact corpus and plan, and P3 executable proof. Writes private non-authoritative artifacts only; activation remains a separate explicit human action.

• hunch_constitution_g2_behavior_replaywrite action

  Execute one exact behavior candidate without a shell in disposable known-bad and known-good worktrees, transplanting the hash-bound known-good test file into both. Diagnostic only: writes no Constitution artifact and grants no policy or G2 authority.

• hunch_constitution_g2_candidates

  Return a bounded private review packet of exact structural candidates from fix-labeled git history, including the current append-only human selection/rejection when present. Read-only: proposed before/after corpus refs are not replayed evidence, and the tool creates no attestation, policy, proof, corpus, authority, warning, or block.

• hunch_constitution_g2_operational_drillwrite action

  Execute the selected private G2 runbook's exact safety regression and return a content-addressed hash-only receipt. Diagnostic only: it writes no rehearsal or shadow evidence, grants no authority, and never signs off G2.

• hunch_constitution_g2_readiness

  Return the exact private G2 dogfood evidence packet: human-selected policies, bound proof/corpus/shadow evidence, operational runbook rehearsals, and blockers. Read-only; it never creates evidence, signs off G2, activates policy, warns, or blocks.

• hunch_constitution_g2_shadow_queue

  Return a bounded private queue of exact-current-proof G2 shadow violations that still require human classification. Read-only; it never records an observation or disposition, changes lifecycle, grants authority, warns, or blocks.

• hunch_constitution_g3_readiness

  Return the exact private G3 advisory packet: human-selected policies and clients, immutable experiment preregistrations, proof-card comprehension/review measurements, executable adapter conformance, scorecard, and blockers. Read-only; it never records evidence, activates policy, or signs off G3.

• hunch_context

  Given a file, symbol, or task phrase you're about to work on, return the MINIMAL relevant memory — invariants to preserve, decisions explaining the design, bug history not to reintroduce, and the blast radius — as a compact brief. Call this FIRST when starting work on something. A task phrase that resolves to no file/symbol falls back to the closest graph matches. Returns a budgeted brief plus a d

• hunch_current_decision

  Decision-grounding: return the single CURRENT (accepted, non-superseded) decision anchored to a topic — the authoritative answer a doc or diff is checked against, plus what it rejected. A shared-store miss is confirmed against fresh team memory before Hunch says the topic has no current decision. If freshness is unavailable, or a topic has an unresolved collision (>1 live), it injects nothing (fai

• hunch_escalations
• hunch_evidence_map

  Compile supplied red-target/green-control, execution, and intervention observations into a bounded read-only evidence map. Useful for bugs, regressions, design invariants, and any other testable behavior. This tool executes no code, mutates nothing, and never converts behavioral influence into an exact correction-owner claim.

• hunch_findings

  List LIVE findings (observed gaps/debt with no fix yet — triage open/accepted-risk/scheduled) concerning a file, glob, or symbol; omit scope for the whole ledger. Call before planning work in an area to inherit past audits instead of re-discovering them. Advisory; resolved/stale findings are excluded unless all:true. Not for invariants (hunch_check_constraints) or bug history (hunch_bug_lineage):

• hunch_get_dependents

  Return everything that transitively depends on a symbol/component (callers + dependent components) so a change's blast radius is known before editing. Returns dependents nearest first with depth and edge kind. Not for the invariants those dependents carry (hunch_blast_radius) or constraints on the target itself (hunch_check_constraints).

• hunch_merge_verdict
• hunch_now

  What just happened and what's next, straight from the graph: the last N decisions, the ROADMAP, and any inline human question such as an imported ADR awaiting explicit approve/decline. Call at session start to orient, or before planning what to work on. Same data as the wiki's now.md. Public store only, EXCEPT a queued commit-repair's liveness is checked against the full store (so a private-overla

• hunch_path

  How does A reach B? Returns the shortest chain of call/import/dependency/contains edges connecting two symbols, files, or components — walked in either direction. Use to understand coupling before a refactor, to verify the actual route behind a must-reach invariant, or to explain why editing A shows up in B's blast radius. Deterministic, read-only.

• hunch_policy_candidates

  List compiled/proposed deterministic Policy IR candidates. Read-only; candidates carry no authority and cannot block. Uses the same Git-native policy store for every MCP client.

• hunch_policy_card

  Return the deterministic proof-card view for a policy: exact assertion/scope, raw evidence vector, uncertainty, blocking readiness, authority, limitations, and next actions. Read-only and grants no authority.

• hunch_policy_evaluate

  Evaluate one or all deterministic Policy IR records and return canonical neutral receipts (satisfied, violated, not_applicable, unknown, error). This is the same evaluator used by CLI and strict CI; models never decide the verdict.

• hunch_policy_plan

  Return the canonical Git-native ProofPlan for a policy candidate: immutable source/current commits, known-good/bad corpus, mutation operators, expectations, and budgets. Planning executes no replay, model, test, or authority transition.

• hunch_policy_proof

  Return the full content-addressed proof artifact for a policy. Read-only; exposes baseline, mutations, proof class, artifact hashes, and limitations without changing authority.

• hunch_policy_shadow

  Return the append-only shadow evaluation ledger, current human dispositions, raw precision counts, unknown/error rate, thresholds, and P4-review recommendation for one policy. Read-only: it never records a sample, changes lifecycle, activates, warns, or blocks.

• hunch_policy_upgrade_correction

  Upgrade the exact supported static ESM import-declaration package projection of one captured correction into a deterministic review packet when the baseline is clean. Writes proposal, plan, proof, and evidence artifacts only; never activates, warns, blocks, or grants authority. Unsupported corrections keep their immediate legacy guard and create no policy.

• hunch_pr_impact

  Given a change (staged, working tree, a branch vs base, or a single commit), return its IMPACT SURFACE: the files whose code transitively depends on the changed files, the invariants directly in scope and those reached via blast radius, and the recorded decisions concerning the touched files. Read-only and advisory — use hunch_merge_verdict for the gate. Call before review to know what a PR can br

• hunch_project_dna

  Derive a bounded, deterministic profile of how this repository communicates, reviews, and builds from one exact committed revision. Read-only: observations remain advisory and are never adopted into graph authority automatically.

• hunch_project_dna_delta

  Return a sealed, explainable delta between two immutable Project DNA profiles. Read-only: reports observation drift and never rewrites history or graph authority.

• hunch_project_match

  Explainably score a commit subject, PR, issue, or message using only deterministic checks supported by an exact-revision Project DNA profile. Advisory: never changes policy or enforcement authority.

• hunch_query

  Full-text + graph search across the engineering memory (decisions, bugs, constraints, components, symbols). Returns ranked records with provenance. Use this to ask 'why' questions about the codebase. Not for orienting on a known file or symbol (hunch_context / hunch_why give the curated slice with its blast radius) or for finding where code lives (hunch_structure).

• hunch_record_correction
• hunch_record_decision

  Persist a new Decision (ADR) into Hunch with provenance. Use after making a non-trivial design choice so future sessions are grounded in it. Set private:true to keep a SENSITIVE decision out of a (possibly public) repo — it is written to the HUNCH_PRIVATE_DIR overlay store and stays queryable locally, never committed here. Returns the stored id, home, and status. Not for a rule the agent must obey

• hunch_record_finding
• hunch_report

  Read task reports: exact delivered memory, agent-reported applications, observed command results and explicit unknowns, as a bounded summary (identities and verdicts, not envelope text; the full report is `hunch report <id> --json`). Supply lesson for exact revision history across tasks. With neither task_id nor lesson, lists recent tasks without guessing which is yours. html writes a local privat

• hunch_review_imported_adr

  Use ONLY after the human explicitly answers the currently surfaced imported-ADR question with approve or decline in this conversation. Never infer approval from silence, continued work, a generic earlier sign-off, or the ADR file's own status. The source and review hashes bind the answer to both the exact bytes and mapped meaning shown. Approve grants human-confirmed authority; decline records rev

• hunch_runbook

  Look up the proven 'how-to' (ordered steps + files) for a recurring task — runbook-SCOPED retrieval (searches within runbooks, not the whole graph). Use at the START of a task to reuse a known procedure instead of re-deriving it. Advisory. Not for design rationale (hunch_why) or free-text memory search (hunch_query).

• hunch_shortlist

  Experimental, deterministic, read-only repository-adaptive diagnostic for a schema/validation issue or reproduction. Preserves a flat top five, adds a transfer-tested hierarchical inspection view, and emits an efficiency-tested advisory progressive inspection queue capped at eleven declarations with deterministic receipts. Optional authenticated same-claim evidence is annotated but cannot reorder

• hunch_structure

  Orient WITHOUT grep/glob rounds: the graph already holds the repo's structure. No target → repo map (components + directories by symbol weight). A directory → its files with their symbols. A file → its outline (symbols, fan-in/out, callers). An exact symbol name → its definition site(s) with one-hop neighbors. Call this FIRST when exploring unfamiliar code — it tells you exactly which file to read

• hunch_task

  Start once per user task; pass the returned task_id to hunch_context. Finish before your final response and include the returned concise contribution card, without asking the user. Applications are explicitly agent-reported and must name an exact delivered occurrence and record hash. Completion never implies successful verification. Not for storing decisions or claiming tests passed; use the CLI t

• hunch_timeline

  Time-travel: the decisions touching a file/symbol over time — what was believed, its valid-time window, and what superseded it. Use to understand how (and why) the design changed, and to avoid re-introducing a deliberately-retired approach.

• hunch_why

  Return the decisions, bugs, and constraints that explain a file path or symbol — the 'why' and the 'what must not break', with evidence. Pass `as_of` (a commit/tag/branch) to time-travel: see what was believed at that point in history. Use when you need the full rationale for ONE target. Not for a budgeted task brief (hunch_context), keyword search (hunch_query), or where-is-it questions (hunch_st

• hunch_wiki_status

  Which generated wiki pages are fresh vs stale (graph moved, source doc changed, hand-edited), plus the specs ledger's doc grades. Call before trusting wiki pages or when deciding whether `hunch wiki --heal` is needed. Public home only — the private overlay wiki is a local concern.

• nuryel_capabilities

  Negotiate before depending on anything: returns the contract version, the capability list (verbs + record schemas), the repository partition this store serves, and which partition kinds it can hold. A capability you need that is missing here is a typed refusal on use, never a degraded answer.

• nuryel_capture
• nuryel_capture_batch
• nuryel_read
• nuryel_records

  Fetch state records by id (from a subscribe event, a read ref, or a write result). Every id is accounted for: found (with its facet), denied (its scope is outside your grants — named, never described) or missing.

• nuryel_subscribe

  Return the change events for a scope with seq > after_seq, strictly ordered. Unfiltered, the events are contiguous (a gap means resynchronize); with facets/subjects filters the response is a subsequence and head_seq is still your next cursor. Each event names the record, its hash, what changed, what it invalidates, and the cause.

• nuryel_writewrite action

• Code scan454 source files scanned20/25
• –Live reliabilityno gateway calls yet and no remote to proben/a
• –Tool poisoningtools not inspected (local package is not executed); not countedn/a
• Auth qualitylocal package, no credentials required12/15
• Maintenancelast push 4 days ago15/15
• Maintainer identityregistry namespace matches repository owner; GitHub account older than a year8/10

Safety score

85/100

higher than 70% of 48,255 servers

Tools

57

the typical server has 7

Tools that change things

5%

typical: 0%. Fewer write tools means less an agent can break

GitHub stars

11

more than 88% of 28,688 servers

Downloads a week

2,483

more than 95% of 12,191 servers with a package

Grade history

• 2026-09-19restoreC → Ascore 85: eval / new Function used

Hunch: common questions

Is Hunch MCP server safe?

Yes, by our scan: it is graded A (85/100). Read the Hunch safety report

How do I install Hunch?

It runs on your machine. Copy the Claude Code, Cursor, VS Code or Claude Desktop config from the install section.

Does Hunch need an API key?

Not as far as the registry entry and our scan can tell: no credentials are declared or required.

Is Hunch maintained?

The last commit was 24 days ago (2026-10-11). The latest release is v1.38.1.

What can I use instead of Hunch?

Servers from other publishers that do the same job: Concord MCP server and State Memory MCP server.

How does Hunch compare with other MCP servers?

safety score 85/100 (higher than 70% of 48,255 servers); tools 57 (the typical server has 7); tools that change things 5% (typical: 0%. Fewer write tools means less an agent can break). Full comparison.