The truly free, open-source, local-first, cross-platform CapCut replacement. Scriptable through a JSON-RPC API & MCP.
host, server, API, CLI, MCP: read at ce64dae (2026-10-20)
Read-only. Nothing built or run. "Verified" means I read it at the line given. "Reasoned" means inferred from code I read. "Unverified" means I could not check it.
Since 10-15: 27 commits touch concat-host, 7 concat-api, 3 concat-server, 4 concat-mcp (new crate), 1 concat-cli. +8 586 / -499 lines across the five crates (git diff --stat ab21457 HEAD).
The server step that slipped in both earlier audits has now landed (363800f), and the new MCP crate is a thin adapter of good quality. What is still broken: how claims and the lock interact with agent sessions, a throttle that any local process can turn against the user, and MCP annotations that misstate what tools do.
Numbers (#[test] count, unsafe count, .unwrap() count in src/)
| crate | lines | tests | unsafe | unwrap |
|---|---|---|---|---|
| concat-host | 17 541 (+~5 600) | 125 | 46 | 34 |
| concat-server | 2 513 (+1 120) | 27 | 0 | 0 |
| concat-api | 2 971 (+619) | 26 | 0 | 0 |
| concat-mcp | 3 041 (new) | 11 (3 e2e) | 0 | 0 |
| concat-cli | 520 (+3) | 3 | 0 | 0 |
| Biggest files: concat-api lib.rs 1 782, playback.rs 1 347, tools.rs 1 133, media.rs 1 109, titles.rs 878, projects.rs 867, updates.rs 862. |
Re-verification of every 10-15 item
10-15 new findings
| # | Item | Status | Evidence |
|---|---|---|---|
| 1 (M6) | pre-auth fd exhaustion | FIXED | registry entry dropped with the connection: concat-server/src/json.rs:158-169, 292-301. A seat is taken only after auth (:275). Pending callers capped at 32 with a 5 s whole-line deadline (:241-249, 261, 342-370). Accept backs off on errors (:206-224). Tests :578, :598, :618 |
| 2 (M13) | crash handler takes locks / allocates | UNCHANGED | crash.rs has no commits since 10-15 |
| 3 (M14) | Android second Vulkan device | UNCHANGED | preview.rs off_thread untouched (diff is only forget_plan :375-384). The cards still build a compositor every round (cards.rs:410). d9f7e08 only skips software adapters |
| 4 | api lib.rs:7 "an MCP server" | MOOT | MCP now exists (concat-mcp) |
| 5 | connections() counts subscribers |
FIXED | counts authenticated seats: server lib.rs:339-342 |
| 6 (L11) | ConsolidateClips adopts any file |
UNCHANGED, wider reach | concat-project/src/commands/mod.rs:630, clips.rs:670. An agent can now reach it through MCP edit_apply (tools.rs:241-247) |
| 7 (L12) | weak tokens, no throttle, 0.0.0.0 | PARTLY | check_token (token.rs:32-46). The window refuses a weak token (settings.rs:637). Throttle added (throttle.rs). Off-loopback warning exists (settings.rs:876, is_loopback :1053). But the estimator passes guessable tokens (new L2), and the throttle is a new DoS (new M1) |
| 8 | two crash codes can never arrive | UNCHANGED | crash.rs untouched |
| 9 | merge bakes un-analysed masks | unverified (window code, not in scope) | — |
| 10/11 | mono output loses right channel; RT-thread frees | UNCHANGED | playback.rs no commits |
| 12 (M16) | README promises gRPC + CLI | FIXED | README.md:68 "A build from source adds gRPC and a command-line tool". Releases still build -p concat only (build-app.yml:550) |
| 13 (L15) | download size unbounded / Windows single rename | size UNCHANGED (models.rs:224-269), now reused by sounds (sounds.rs:432). Rename FIXED |
5 retries with backoff on Windows (projects.rs:357-372, 388-392) |
| 14 | merged renders never collected | unverified (window) | — |
Older items carried by 10-15
| Item | Status | Evidence |
|---|---|---|
| 12Oct#5 24 held stills FIFO | unchanged (pool not re-read) | preview.rs:167-170 hold_still as before |
| 12Oct#6 shared save scratch | FIXED | per-call create_new temporary with pid and nonce (projects.rs:304-345). Ticket ordering under a per-folder mutex (:181-298). Commits e75af9b, 9197340. The cross-process case is now covered by the lock. The doc at projects.rs:265-271 still says "a project folder has none [lock] today" |
| 12Oct#30 512 MiB live-title ceiling | UNCHANGED | titles.rs:98 |
12Oct#32a gpu capability never appears |
UNCHANGED | Api::with_dirs uses Monitor::new() (api lib.rs:296), which sets gpu: None (preview.rs:92-100), so has_gpu() is always false (:156-165). lib.rs:506 never fires. The test pins its absence (lib.rs:1262) |
| 12Oct#32b stop waits on open gRPC Events | UNCHANGED (reasoned) | The Events sender lives in the hub subscriber list (grpc.rs:75-77). Graceful shutdown waits for the stream (:170-173). close_doors joins the listeners (lib.rs:404-406) before hub.close() (:420). Only gRPC embedders and tests are affected; the window has no gRPC |
| 12Oct#32c CLI render in place, no MAX_SIDE | UNCHANGED | concat-cli/src/main.rs:419-431 |
| 12Oct#33 dirs.rs "downloaded models" | WORSE | dirs.rs:32 still says "Large state: downloaded models", but data now also holds irreplaceable user content: the audio library and cloned voices (dirs.rs:89-100, sounds.rs, voices.rs:75-77). On Windows that is non-roaming %LOCALAPPDATA% |
| 11Oct#21 docs | PARTLY | api lib.rs:7 now true. concat-host/Cargo.toml:37 still claims a CPU compositor |
| 11Oct#6 stop hangs on wildcard bind (Windows) | FIXED | loopback() (server lib.rs:89-100, 393-395), test :515 |
| 11Oct#7 refused thread kills accept loop | FIXED | json.rs:252-256; errors back off :218-221 |
| 11Oct#10 API create unchecked | UNCHANGED | api lib.rs:516-533 calls projects::create with no checked_size/checked_rate. A rate_den of 0 is written, open then refuses it (projects.rs:453), and the folder is left behind |
| 11Oct#15 server edges | PARTLY | gRPC slow subscriber silently dropped: unchanged (grpc.rs:75-77). A poisoned subscriber list silences events: unchanged (hub.rs:54-57). CLI api dies on non-UTF-8: unchanged (cli main.rs:175-176). key_of doc on impl Drop: unchanged (api lib.rs:1064-1066) |
| 11Oct#16 stop blocks window; jobs die silently | stop FIXED (stop_detached server lib.rs:369-379, used at settings.rs:624). Jobs: PARTLY |
reap books a vanished job as failed (api lib.rs:897-907). Under panic = "abort" (src/Cargo.toml:127) a panic still ends the process. playback.rs:683, 693 unwraps unchanged |
| 11Oct#17 reads unconfined / caller strings to FFmpeg | UNCHANGED, now documented | api lib.rs:360-365. preset passes through unvalidated (:708-711) |
| 11Oct#23 host lows | UNCHANGED, one WORSE | The silent data-folder move (dirs.rs:124-136) now also moves audio/voices user content (:111-116). updates.rs has no commits. templates.rs:357 poster as preview.jpg |
| 11Oct#25 WAV header overflow | FIXED | record.rs:621-626 clamps to u32::MAX - 36; header :590-606 |
11Oct#29 --socket deletes whatever file |
UNCHANGED | server lib.rs:272 |
| S1 unsigned update chain, quarantine strip | UNCHANGED | updates.rs has no commits since 10-15. Still this area's worst standing risk |
| S2 API root = whole home | UNCHANGED, wider | server lib.rs:131-137, 206-210. MCP headless uses the same default (concat-mcp/src/backend.rs:185-188) |
| S3 | FIXED | see #1 |
| S4 Clear cache deletes reversed/enhanced media | unverified | reverse.rs untouched; enhance.rs changed and was not re-read for this |
| S6 embedded API has its own monitor/GPU | UNCHANGED (reasoned) | api lib.rs:296. It also shares the process-wide decode pool with the window (preview.rs:94), so an agent's previews evict the user's frames |
| S8 HDR proxy name | UNCHANGED | proxy.rs untouched |
| S9 composites on hub thread; slow subscribers | hub UNCHANGED (api lib.rs:840-868). JSON PARTLY |
A full outbox (256) now silently unsubscribes the caller while its connection stays (json.rs:316, 447-455, hub.rs:58-66). A writer stalled more than 10 s is hung up on (json.rs:63, 430-437) |
| S13 per-sample clip scan, S14 silent missing font | UNCHANGED | files untouched |
| S16 connections count | FIXED | — |
| S17 Unix socket perms from umask | UNCHANGED | server lib.rs:273 (no chmod) |
Tally, about 40 items: 11 fixed, 6 partly, 3 worse, about 17 unchanged, the rest unverified or moot. Every 11-Oct and 12-Oct server-step item that was in scope landed in 363800f, and each landed with a test.
New findings
HIGH: none new. S1 (the unsigned update chain, unchanged) remains the area's top risk.
MEDIUM
M1. The auth throttle turns loopback into a shared lockout: any local process, or a web page, can lock the user's own agent and scripts out of the window's API indefinitely. Verified.
- Every TCP caller is keyed by IP alone (
json.rs:93-97), so all local callers arePeer::Ip(127.0.0.1). Every Unix-socket caller is onePeer::Local(:114-116). - Any first line that is not a correct
authcounts as a failure (json.rs:267-271), and that includesGET / HTTP/1.1. - After 3 failures the address is refused before its token is read (
json.rs:228-240), with waits doubling to 60 s (throttle.rs:110-116). The test pins that even the right token is refused (json.rs:650-659). - Failure: a browser tab runs
fetch("http://127.0.0.1:7420")four times, or any unprivileged local process sends one garbage line a minute.concat --mcpis then refused permanently. Inautomode it silently falls back to headless (backend.rs:81-86, stderr only), which is the path the window mode exists to avoid. - Second vector: 32 silent sockets renewed every 5 s fill
MAX_PENDING, and everyone getsbusy(json.rs:241-249). Silent connections are not even throttled (:262-266). - The throttle buys nothing for a minted 128-bit token; the module says so itself (
throttle.rs:6-9). - Fix: skip the throttle when the token was minted. Count only well-formed auth attempts. Cap pending connections per peer.
M2. Projects an agent opens stay claimed and locked until the server stops or the process exits. The window then refuses them with a message that offers no way out, and the agent's unsaved edits are discarded with no warning. Verified.
- Claims are released only by
project.closeor when theApiis dropped (api lib.rs:579-592, 1066-1073). - A connection that ends does nothing to sessions (
json.rs:332-335). - The Window backend's
finishis the no-op default (backend.rs:47;Windowdoes not override it,:143-167). - A headless
concat --mcpkeeps itsconcat.lockfor the whole agent session, because clients keep MCP servers alive. - Failure: the agent builds "Promo" through the window's server and the session ends. The user clicks Promo in Recents and gets "This project is open through the Remote API" (
en.json:721,studio.rs:8169-8171). The only way to free it is to switch Automation off, and that drops the agent's unsaved edits without a word (theApidrop does not save). - Every keystroke in the host, port or token field restarts the server (
settings.rs:356-372→apply_server→stop_detached). Each restart discards every agent session and cancels its export. - Fix: track sessions per connection and close (or save-prompt) them when it ends. Have
Window::finishclose what it opened. Let the window take over an API-held project with a prompt. Apply settings changes on Enter, not on every keystroke.
M3. MCP annotations misstate what tools write, and writes reach anywhere under home. Verified.
preview_frameisreadOnlyHint: true,destructiveHint: false(tools.rs:445-453). Withoutputit callsFile::createon any path under the roots, with no extension check and no check for an existing file (api lib.rs:811-813, 1148-1163).export_startclaimsdestructiveHint: false(tools.rs:413-419) but overwrites an existingoutput(api lib.rs:665, muxer guessed from the extension).project_openis unconfined, andproject_savethen writesconcat.json, plus the lock and caches, into that folder wherever it is (api lib.rs:360-365, 600-602).- Failure: a client that auto-approves read-only tools (behaviour varies by client, unverified per client) runs an agent that is prompt-injected through title text or media names. Those strings flow into tool output via
summary.rs. It callspreview_frame{output:"~/.ssh/config"}and the file is replaced by a PNG with no prompt. docs/MCP.mdpresents "only under your home folder" as a safety property; home is where everything that matters lives.- Fix: correct the hints. Require
.pngand a new file, or an explicitoverwrite. Refuse to export over an existing file unless asked. Consider defaulting the roots to Videos/Movies plus the open project folders.
M4. A lock that fails with an I/O error lets the open go ahead with no guard. Path verified; Windows behaviour reasoned.
OpenProjects::claimmapsLockError::Ioto "no lock, open anyway" (api lib.rs:183-189).acquireturnsPermissionDeniedintoIoafter 4×50 ms (lock.rs:97-106).- On Windows a lock file held without sharing by an AV scanner, OneDrive or Search, or left delete-pending on volumes without POSIX delete semantics (FAT/exFAT, SMB, older builds; unverified which builds), gives exactly that error.
- Windows
same_fileis onlypath.exists()(lock.rs:257-263), so a lock taken on a delete-pending file passes the check. Every later opener then getsPermissionDenied→Io→ an unguarded open. This is two processes saving one folder, the thing 68ceffe set out to prevent, on the priority platform. - Network file systems with no locks degrade silently too.
- Fix: refuse, or open read-only, when the folder is writable but cannot be locked. Retry
PermissionDeniedfor longer.
M5. concat --mcp as shipped is never run by CI, and its panic safety net is dead code. Verified.
- The shipped profile is
panic = "abort"(src/Cargo.toml:127). That makescatch_unwindinprotocol.rs:426-429("a panic in a tool is that call's failure") andhub.rs:91-104(inside the window's server) no-ops. - Any engine panic kills the agent session and its running export. Inside the window it kills the window and the user's unsaved work.
- The release binary is GUI-subsystem on Windows (
concat/src/main.rs:8); its use as a stdio server is untested. - The e2e test drives the dev
concat-mcpbinary (tests/e2e.rs:29). The workflows contain no--mcp(grep of.github/workflows). - The discovery →
Window::findpath has no end-to-end test either: e2e setsCONCAT_API_ADDRESS/TOKEN(e2e.rs:385-391). - Fix: a CI smoke test that runs
concat --mcp headlessfrom the app-profile build on Windows and macOS, sendinginitialize,tools/listand one edit.
LOW
- L1. The Window backend re-sends any request once after a connection error (
backend.rs:144-162). If the server applied it and the reply was lost (server restart mid-call, or the 10 s write-timeout hang-upjson.rs:430-437), aclip_addorcaptions_addlands twice. The hang-up is reachable: the client reads events only during a call (client.rs:95-115), so an idle agent during an export backs up the 256-line outbox. Retry only idempotent tools. - L2. The token estimator (
token.rs:53-72, floor 64) passes12345678901234567890(about 66 bits) andabcdefghijklmnopqrst(about 86). On a 0.0.0.0 bind, IPv6 address rotation evicts strikes once 4 096 peers are booked (throttle.rs:76-89). Any success from a shared IP wipes its strikes (json.rs:273,grpc.rs:204). - L3. Inline previews are unbounded.
width/heightup to 8192 (tools.rs:878-904defaults only when both are absent) gives a ~256 MB RGBA frame, PNG-encoded and base64'd into one JSON line headed for the agent's context. Above the client's 256 MiBMAX_REPLY(client.rs:21) the line is cut, read as "not JSON-RPC", and retried through L1, so the frame renders again. - L4. Recents are written non-atomically (
projects.rs:545-553) by the window, the window's API and every headless MCP process (api lib.rs:570-572). A read during a write returns empty, and the nextrememberwipes the list. Agent projects also fill the user's Recents and home page. - L5.
Api::cancelcancels the shared export slot (api lib.rs:786-789), not its job's flag, contrary tojob.rs:13-16's own rationale. A small race can cancel the window's export. - L6. Job names restart at
j1for eachApi(api lib.rs:892-895). After a server restart, an agent pollingj1gets "no job" or someone else's job. - L7. The Tripwire decides by pid liveness (
jobs.rs:157-165), whilelock.rs:18-23explains why pids answer wrongly. On Windows and Android a reused pid lets a crash-loop job restart itself. - L8. No read timeout after auth (
client.rs:55-59). With one request at a time (protocol.rs:18-22), a stuck hub or a 600 sexport_waitblocks every other tool,export_cancelincluded. - L9. Two windows: the second overwrites
automation.json, and when its server stops it removes the listing (discovery.rs:134-141), so the first window's running server is no longer listed. - L10. Sound downloads trust a digest from the same manifest as the URL, with no size cap (
sounds.rs:383-395, 432-447;models.rs:224-269). Same class as S1. - L11. Stale or misplaced docs:
prefs.rs:230-231says an empty token means "none"; it means a minted one.projects.rs:265-271says there is no lock.settings.rs:964-969:new_token's doc sits onLOG_LEVELS.settings.rs:1044-1052:split_listen's doc is merged intois_loopback's.api lib.rs:1064-1066: key_of's doc sits onimpl Drop.protocol.rs:426is false under abort.concat-host/Cargo.toml:37still claims a CPU compositor.dirs.rs:32"downloaded models".is_exposed(server lib.rs:77-82) is unused; the window duplicates it asis_loopback.
- L12. The MCP "modern" protocol
2026-07-28withserver/discover(protocol.rs:36-41) is past my knowledge: unverified against the spec.
Commits since 10-15, area view
- 363800f server hardening. Correct in shape: RAII seats and places, a registry that cleans itself up, a whole-line auth deadline, a bounded outbox with a write timeout, and stop cancelling before joining. Each guard has a test (
json.rs:545-778). What it gets wrong is the throttle model (M1). - 56ff055 discovery + client. Owner-only 0600 write via staging and rename (
discovery.rs:76-110); removal checks the pid. Fine. Windows relies on the folder's ACL (unverified). - db529b2 / f70332c / 30584f7 MCP. A thin adapter, with schemas generated from the API types and round-trip tested (
tools.rs:1087-1116). Cancellation and progress are handled properly. Annotations wrong (M3); shipped path untested (M5). - 68ceffe / 46e04b6 project lock. An OS lock rather than a pid file, takeover after a crash, an inode recheck against the release race (
lock.rs:113-118), a real child-process test (:300-345). Holes: M4, and M2's lifetime problem. - e75af9b save ordering. Fixes 12Oct#6 properly.
- Host features (sounds, voices, stabilize, enhance preview, tripwire, voiceover recorder): skimmed for path and security handling only. File names are sanitised (
sounds.rs:162-168, 224-252;voices.rs:89-99) and the voice decode is bounded (voices.rs:117-120). Not audited for behaviour.
Good
- The fd-leak fix is complete and tested.
- Stop is detached from the UI thread.
- The lock is designed properly.
- Saves are ordered, with their own temporaries and Windows retries.
- MCP schemas come from the types, so they cannot drift.
- Stdout carries protocol only: no
println!in the engine crates outside tests (grep). - The discovery file is owner-only.
Scores (area; 10-15 is the mean of the host, server, API and CLI rows)
| Aspect | 10-20 | 10-15 | Why |
|---|---|---|---|
| Code quality | 8 | ~7.3 ↑ | The server rewrite is careful RAII with no unwraps. Dead catch_unwind and misplaced docs remain |
| Design architecture | 7 | ~7.3 → | One dispatcher with thin transports, MCP as an adapter: right. Claims are not tied to connections (M2), the lock fails open (M4), the embedded API has its own monitor and the shared pool |
| Design philosophy | 8 | ~6 ↑ | "Refuse rather than lose", an OS lock rather than pids, schemas from types. The Tripwire contradicts it (L7) |
| Maintainability | 7 | ~7.3 → | Clear modules, but api lib.rs is 1 782 lines and tools.rs 1 133; docs drift (L11) |
| User likeability | 6 | (Remote API 4) ↑ | MCP is real value with a one-line setup. Dead-end refusals (M2), Recents pollution (L4), settings edits that kill agent work |
| Performance | 6 | ~5.7 → | One hub thread composites synchronously; previews unbounded (L3); an agent shares the window's decode pool |
| Scalability | 5 | 5 → | One API thread, one export slot, 64 seats: fine for a local tool, nothing more |
| Security | 5 | ~4.5 ↑ | Pre-auth fd leak fixed, token floor, owner-only discovery. A new lockout DoS (M1), annotations that understate writes plus home-wide roots (M3), and S1 untouched |
| Testing | 7 | ~5.5 ↑ | 27 server tests, a cross-process lock test, MCP e2e headless and via server, schema round trips. The shipped --mcp path, the discovery path and Windows lock races are untested (M5, M4) |
| Documentation | 6 | 6 → | docs/MCP.md is clear and accurate on modes, but overstates the safety of "home". About ten stale or misplaced comments (L11) |
Top fixes, in order: M1 (no throttle on a minted token), M2 (claims tied to sessions), M3 (correct hints and refuse overwrites), M5 (CI smoke of concat --mcp on Windows), M4 (fail closed), then S1.