openbinary — See inside any binary

• Add Note — ADD NOTE — write a note on a binary's or package's activity in the workspace. `@email` of a workspace member inside the text mentions them. At most 4096 bytes. Requires membership and a writable key. Returns the recorded event.
• Annotate Autofill — ANNOTATE AUTOFILL — autofill (LLM-name) functions of a binary bottom-up over the call graph, writing the names into the caller's workspace annotation overlay as LlmGenerated suggestions. Precedence is FACTS > the action you just took > older guesses: a real symbol or structured (Go/C++/ObjC) name is NEVER overwritten; the unnamed `sub_<va>` floor is always filled; existing manual / LLM annotations are re-derived only when overwriting (and the prior name is captured for revert). Pass `va` to scop
• Annotate Clear — ANNOTATE CLEAR — delete the annotation(s) at a VA: the whole-function row by default, one sub-function row when `kind` + `slot` are supplied, or a data name when `kind` is `data`. Restricted: requires the `annotations` feature (Pro plan floor). Returns not-found below Pro / when the caller cannot see the binary / when not a workspace member / when no row existed.
• Annotate Comment — ANNOTATE COMMENT — set a free-text comment anchored at an address (the `comment:<insn_va>` slot, or the function entry when `slot` is omitted), stamped as a Manual edit. Restricted: requires the `annotations` feature (Pro plan floor). Returns not-found below Pro / cannot-see / non-member; bad-request on an oversized comment. Omit `workspace` for your solo overlay.
• Annotate Corpus Function — CURATOR — rename a function EVERYWHERE it appears in the public corpus by its C1 fingerprint. Author by C1 fingerprint directly (`quad_hex`) OR by location (`sha256` + `va`) — a location MUST point at a PUBLIC corpus binary (a private / unindexed SHA is rejected before any read, so a curator can never name from a function they shouldn't see). When `dry_run` is true: WRITE NOTHING — return the PUBLIC-only affected sites (`members`: every `{sha256, va}` whose body fingerprints to this quad, public
• Annotate Data — ANNOTATE DATA — name a data / global address (the IDA `rename_global_variable` gap), stamped as a Manual edit. Keys the data name distinctly from a function name at the same address. Restricted: requires the `annotations` feature (Pro plan floor). Returns not-found below Pro / cannot-see / non-member; bad-request on an oversized name. Omit `workspace` for your solo overlay.
• Annotate Function — ANNOTATE FUNCTION — set the display name of one whole function at its entry VA, stamped as a Manual edit. Restricted: requires the `annotations` feature (Pro plan floor). Returns not-found below Pro / when the caller cannot see the binary / when not a workspace member; bad-request on an oversized name. Omit `workspace` for your solo overlay.
• Annotate List — ANNOTATE LIST — read the workspace-shared annotation overlay for one binary: the manual (and autofill) function renames / local renames / comments / data names laid over its decompilation, for the caller's workspace. Restricted: requires the `annotations` feature (Pro plan floor). Returns not-found below Pro, when the caller cannot see the binary, or when they are not a member of the requested workspace. Omit `workspace` to read your personal (solo) overlay.
• Annotate Local — ANNOTATE LOCAL — set the display name of a local / parameter (the `var:<offset>` slot) of a function, stamped as a Manual edit. Restricted: requires the `annotations` feature (Pro plan floor). Returns not-found below Pro / cannot-see / non-member; bad-request on an oversized name. Omit `workspace` for your solo overlay.
• Annotate Prototype — ANNOTATE PROTOTYPE — set a typed-prototype override (return/param types) on a function from a serialized `ProtoOverride` JSON `proto_json` (the `proto:<idx>` slot, or a whole-prototype edit when `slot` is omitted), stamped as a Manual edit. `proto_json` is validated on write (e.g. `{"origin":{"kind":"manual","user":"me"},"return_type":{"kind":"primitive","ty":"i32"},"params":[{"name":"fd","ty":{"kind":"primitive","ty":"i32"}}],"is_variadic":false}`); malformed or oversized JSON is rejected. Re-d
• Binary Global Stats — Corpus-wide aggregate statistics across the WHOLE corpus (visibility-unscoped): total binaries, single-dimension facet distributions (platform / arch / binary_type / capability / verdict / source / component / package-provenance), the arch×platform, package×platform, and package×cwe matrices (package×platform is the engine-accuracy metric — origin-kind vs byte-derived platform, a `bsd` origin tagged `Linux (unknown)` is a classification gap; package×cwe is the false-positive scoreboard — CWE fir
• Cluster — The similarity CLUSTER around a SHA-256 — the DNA-vector neighbourhood graph {nodes, edges, …} the web /cluster view renders. Visibility-scoped: the caller's allowlist is pushed INTO the ANN, so an unseeable neighbour never appears, and a source SHA you cannot see returns not-found. Returns 503 while the similarity index is not yet populated.
• Compare — DIFF two binaries by SHA-256 — the DNA/Jaccard similarity score, per-set behavioural diffs (capabilities / entitlements / ATT&CK / frameworks / …), byte-identical structural hashes, and function-symbol / interesting-string / network-endpoint diffs. Restricted: binary diffing is a Pro feature — an anon/Free caller is refused (403 upsell). Either SHA the caller cannot see returns not-found (404-cloak), never a leak that it exists.
• Corpus Health — Corpus INTEGRITY report — whether the corpus can state where its contents came from. Four census blocks: `provenance` (how many binaries carry at least one package membership, and the ABSOLUTE count that carry none — 84% carried none until recently, which is why no firmware child could render "Seen in"), `packages` (row counts by identification status — identified / partial / unidentified / unrecorded — plus `extracted_without_members`, the regression canary: a package claiming extracted binarie
• Decompile Function — On-demand per-function decompilation. Available only when the operator enables the MCP decompile gate. `query` is a function ADDRESS (0x-hex, sub_<hex>, fcn.<hex> or decimal — take it from list_functions / graph_find_node; a symbol name is rejected) or an address RANGE (`START..END` or `START+LEN`), which renders those bytes continuously, ignoring function boundaries. Each reply is at most page_bytes (default 32 KiB, max 256 KiB); a cut reply carries truncated=true and next_offset — pass that ba
• Describe Binary — Compact structured digest of one binary: identity (name, container format, type, platform, arch, file/image/header sizes, md5/sha1, build identity — uuid / ELF build-id, bundle id, min-os / sdk / source-version — and signing status), verdict, `analyze_status` (analyzed | empty | unsupported) with `unsupported_reason` (mips16 | micromips) when the engine refused to decode the instruction set — read it before you read a zero function count — capability and structure counts, top signatures, `title`
• Describe Package — Firmware / container PACKAGE detail by SHA-256 — the hero shape the web /package view renders: identity (name, kind, uploaded_at, status), the child-binary + CVE + malicious rollups, composition, the firmware overview, and (Pro) the full SBOM. The whole package pillar an agent walking a firmware image starts from; then list_package_members for the children. Visibility-scoped: a package you cannot see returns not-found (never a 403 that would leak its existence).
• Diff Functions — WHAT CHANGED between two builds, function by function — the changed functions of a pair with the unified diff of their stored pseudocode, in one call. Matches functions by name across the two binaries; a shared function is `changed` when its rendered body differs after image addresses and synthetic names are masked, so a function that merely moved is `same`. Rows are ranked so the ones to read first come first: reachable before auth, then carrying a CWE finding, then attributed to a capability,
• Download Binary — Download the raw, original bytes of one binary by its SHA-256. Returns the exact uploaded blob as a base64 `blob` resource (mime application/octet-stream) alongside a JSON metadata line {sha256, name, size}. Restricted: requires the `binary_download` deployment capability AND a per-user grant — no plan tier unlocks it, so most callers will never see it. Returns not-found when the caller lacks the grant or cannot see the binary. Binaries above the transport cap return an error rather than an over
• Emulate Function — EMULATE FUNCTION — run one function on inputs you choose and see what it does. Answers what a listing cannot: what an obfuscated decoder, a checksum, a key derivation or a string deobfuscator produces for a given input. `va` is 0x-hex or decimal and the run starts EXACTLY there: a function entry runs the function; a VA inside a function runs from that instruction to the function's end (what precedes it is skipped — how you enter a check that sits after a library call); a VA no function claims, i
• Function Suggestions — ENRICHMENT ONLY — ranked IL-similarity name HYPOTHESES for a function, never an authoritative name. This tool reads an operator-loaded cross-architecture fingerprint index and returns candidates that each carry their own cosine `score`, acceptance `tier` (corroborated | uncorroborated) and the `bar` that tier had to clear; a suggestion NEVER enters list_functions rows, describe_binary, function names, origins, SBOM component claims or any verdict, and nothing you do with it writes to the corpus.
• Get Activity — GET ACTIVITY — the timeline of a binary or package in the workspace, newest first: state / owner / tag / name changes, notes, function annotations, and system changes (analysed, reanalysis requested). `limit` defaults to 50, at most 200. Requires membership.
• Get Capabilities — What this binary can DO — the ATT&CK-mapped capability inventory, grouped by MITRE tactic. Each capability carries a plain-language summary, an honest confidence badge (sound vs heuristic), its named ATT&CK techniques (id + name + tactic), and per-function attribution (`functions[].va` — paste into decompile_function to see the code). `att_ck`-empty capabilities fall under `unattributed`, bucketed by category. `scoring_eligible: false` is an honest signal that a real capability carries no verdic
• Get Diff Text — The stored TEXT beside a compare — what the update is about, what security bug was fixed, whether anything got looser, what is new, what is gone, where to start reading — written once by an agent through the openbinary:diff skill and served to every entitled reader since. Returns {text, stale, facts_version, prompt_version, written_at, dropped_sentences}; `stale: true` means a side was reanalysed after the text was written (the text stays; rewriting is an explicit put_diff_text). Not-found with
• Get Dna — The binary's DNA feature vector.

Showing the first 25 tools — full list and schemas: Directory API.