fbmcp: install and use
Audience: the person who installs and runs fbmcp, and did not build it. No Go or Firebird internals knowledge is needed. State as of 2026-11-16, version 2027.2.24 (local and remote mode, HQbird install, fbagent bridge, HQbird cluster, result grid).
Short versions, not on this site: quickstart-en.md, quickstart-ru.md,
quickstart-pt.md. Details, in English: remote.md (remote
mode), agent-bridge.md (fbagent),
tool-reference.md (every tool, its tier and arguments),
claude-desktop.md (client wiring on the dev host),
ADRs in decisions/.
1. What fbmcp is
fbmcp is an MCP server for Firebird 2.5–5.0 on Windows and Linux. An MCP
client (Claude Desktop, Claude Code, Cursor, Codex, ZCode, others) uses it
to work as a DBA: monitoring, backups and verified restores, SQL/DDL/DCL,
indexes, users and grants, firebird.conf, a scheduler, migrations, the
HQbird fbagent and the HQbird cluster. It has 122 tools: 56 read-only (Tier 0), 44 Tier 1,
21 Tier 2, 1 Tier 3 (disabled). fb_demo_write is there only when the
server runs with FBMCP_DEMO_TOOLS=1.
A client sees only the tools it can use. The tool list hides: Tier 3,
tools above the identity’s max tier, fb_confirm and fb_cancel for a
read-only identity, the fbagent tools while agents: is empty
(fb_agent_discover and fb_agent_register stay), the migration tools
while no database has migrations_dir, and the cluster tools while
fbmcp.yaml has no rcm: section. When a config reload changes this
set, the client gets a new list. A direct call to a hidden tool gets a
clear refusal.
| Binary | Role |
|---|---|
fbmcp |
The server. Default: MCP over stdio (a client starts it). -serve: no stdio, runs as a service (scheduler, attach socket, HTTPS listener if configured). fbmcp bridge: a stdio client for a remote fbmcp |
fbmcpctl |
Operator CLI: doctor, approve, secret, apikey, tls, verify, status, ping, setup, repair, hqbird-install, gate. Same program as fbmcp ctl |
fbmcp-tray |
Windows tray: Approve/Deny pop-up for requests that need an operator |
One kernel per state directory. A second client with the same config does not start a second kernel: it attaches to the running one.
2. Safety model
Every call passes policy → human confirmation → audited execution.
| Tier | What | How it is confirmed |
|---|---|---|
| 0 | Reads (engine-enforced read-only transactions) | Not needed |
| 1 | Routine changes: DML/DDL, backup, users, grants | In-band: the client shows a request id and a token; the person agrees; the client calls fb_confirm |
| 2 | High impact: service control, restore-replace, firebird.conf, exclusive window, registration, agent restarts |
Out-of-band only: the tray or fbmcpctl approve. The client can never approve it |
| 3 | Critical (fb_db_drop) |
Disabled by design |
- Tier 2 usually needs an open maintenance window (
fb_window_open, on the samedb) and a verified backup younger than 24 hours. The refusal says what is missing. - A pending request expires after 15 minutes. A wrong-channel attempt
does not consume it;
fb_cancelremoves it. - Every gated tool accepts
mode: preview: it shows the impact and does nothing. - Every decision (allow, pending, approved, denied) goes to the
hash-chained
audit.jsonl. - There is no approval web page, and there must never be one: a page the model can reach would break the design.
3. Choose a setup
| Setup | Where fbmcp runs | Use it when | Section |
|---|---|---|---|
| A. Local stdio | Your PC; the client starts it | One admin, Firebird on this PC or in the LAN | §6 |
| B. Local service | Windows service on the DB host (-serve) |
Schedules must run when no client is open | §7 |
| C. Remote | Service on the DB server, HTTPS + API keys | Several admins, a distant server, autonomous agents | §8 |
In setup A the kernel lives only while a client runs, so schedules fire
only then. Use B or C for unattended schedules (nightly_verify).
fbmcp in HQbird. HQbird 2027 setups and fbagent install fbmcp by
default: one service per server for all Firebird lines, set up by
fbmcp ctl hqbird-install. On such a server do not follow §6–§8; read
hqbird.md (layout, remote mode with --allow-from, API keys).
4. Get the binaries
Requirements: Go ≥ 1.26.11 (an older Go 1.21+ downloads go1.26.11 by itself on the first build; this needs internet), CGO not needed; Firebird 2.5–5.0; fbagent ≥ 2.67 with the Local API for the agent tools (optional); fbagent 3.1.0+ and RCM 2027.5.11+ for the cluster tools (optional).
git clone https://github.com/alexeykovyazin/hqbirdmcp.git
cd hqbirdmcp
$env:CGO_ENABLED = "0"
go build -trimpath -ldflags "-s -w" -o dist/fbmcp.exe ./cmd/fbmcp
go build -trimpath -ldflags "-s -w" -o dist/fbmcpctl.exe ./cmd/fbmcpctl
go build -trimpath -ldflags "-s -w -H=windowsgui" -o dist/fbmcp-tray.exe ./cmd/fbmcp-trayLinux binaries (cross-build from Windows, or build on Linux without
GOOS):
$env:GOOS = "linux"; $env:GOARCH = "amd64"
go build -trimpath -ldflags "-s -w" -o dist/fbmcp ./cmd/fbmcp
go build -trimpath -ldflags "-s -w" -o dist/fbmcpctl ./cmd/fbmcpctlChecksums: Get-FileHash dist/* or sha256sum dist/*. More targets:
packaging/README.md. Put the binaries in a folder
of their own (examples use C:\fbmcp\): mkdir C:\fbmcp; copy dist\*.exe C:\fbmcp\
(Linux: install -m 0755 dist/fbmcp dist/fbmcpctl /usr/local/bin/).
fbmcp ctl and fbmcpctl are the same program. fbmcpctl help lists the
commands. Commands that read the config take --config <file>; without it
they use FBMCP_CONFIG, then fbmcp.yaml in the current folder, then next
to the program, then the HQbird default (C:\HQbirdData\fbmcp\fbmcp.yaml,
/etc/fbmcp/fbmcp.yaml), and print using config <path>.
5. Configure fbmcp.yaml
Full example: packaging/fbmcp.yaml.example.
Minimal Windows config:
state:
dir: C:\ProgramData\fbmcp\state
# secrets:
# dir: C:\ProgramData\fbmcp\secrets # setups B and C
instances:
- id: fb5
addr: localhost:3050
bin_dir: C:\Program Files\Firebird\Firebird_5_0 # folder with gbak, gfix, gstat
version: "5.0"
service: FirebirdServerDefaultInstance # Linux: firebird.opt_fb50
service_user: SYSDBA # instance tools (fb_firebird_log, ...)
service_secret_env: FBMCP_FB5_SERVICE_PW
databases:
- id: main # tools and chats use the id, never the path
instance: fb5
path: D:\Data\main.fdb
backup_dir: D:\fbmcp\backups
work_dir: D:\fbmcp\work # restore tests; Linux: not under /tmp
ro_user: FBMCP_RO # a read-only Firebird user
ro_secret_env: FBMCP_MAIN_RO_PW
admin_user: SYSDBA
admin_secret_env: FBMCP_MAIN_ADMIN_PW
notify:
webhook_url: "" # optional: a chat webhook for pending requests
retention:
keep_days: 0 # 0 = never delete backupsSecrets. The YAML holds only variable names. fbmcp looks up each
value in this order: environment variable → file <secrets.dir>/<NAME> →
OS keyring of the account that runs fbmcp.
C:\fbmcp\fbmcpctl.exe secret set FBMCP_MAIN_RO_PW # keyring of the current user (setup A)
C:\fbmcp\fbmcpctl.exe secret set --dir C:\ProgramData\fbmcp\secrets FBMCP_MAIN_RO_PW # secrets dir (setups B, C)The value is read from stdin (type it, then Ctrl+Z, Enter; Linux Ctrl+D).
A service account (LocalSystem, NT SERVICE\...) cannot see your keyring:
use secrets.dir for services. A changed secret file applies in a few
seconds without a restart.
Read-only user. Create a separate Firebird user for ro_user (SQL:
quickstart-en.md:
CREATE USER FBMCP_RO, GRANT SELECT on every table, and on Firebird 4.0+
a role with MONITOR_ANY_ATTACHMENT so fb_sessions sees all connections).
SYSDBA as the RO user is for dev only; doctor warns about it.
Live reload. fbmcp applies changes to fbmcp.yaml by itself (file
watcher), or on fb_config_reload. No client or process restart. A reload
waits for jobs on the changed databases (30 s), then applies. Refused:
a state.dir change, duplicate database paths, an insecure remote config.
Check: C:\fbmcp\fbmcpctl.exe doctor must end with doctor: green
(remote mode: doctor --remote, no FAIL lines). It makes a missing state
folder, logs in to every database with both users, and each FAIL line ends
with Fix: and the command to run.
6. Setup A — local stdio
Use absolute paths. Do not set listen. Restart the client after a config
change: clients load servers only at start.
| Client | Where | Entry |
|---|---|---|
| Claude Desktop | Settings → Developer → Edit Config (%APPDATA%\Claude\claude_desktop_config.json) |
"mcpServers": {"fbmcp": {"command": "C:\\fbmcp\\fbmcp.exe", "args": ["-config", "C:\\fbmcp\\fbmcp.yaml"]}} |
| Claude Code | command line | claude mcp add --transport stdio fbmcp -- "C:\fbmcp\fbmcp.exe" -config "C:\fbmcp\fbmcp.yaml" |
| Cursor | %USERPROFILE%\.cursor\mcp.json or .cursor\mcp.json in the project |
same JSON as Claude Desktop |
| Codex | codex mcp add fbmcp -- C:\fbmcp\fbmcp.exe -config C:\fbmcp\fbmcp.yaml or %USERPROFILE%\.codex\config.toml |
[mcp_servers.fbmcp] command = 'C:\fbmcp\fbmcp.exe' args = ["-config", 'C:\fbmcp\fbmcp.yaml'] |
| ZCode | .zcode\config.json in the project |
"mcp": {"servers": {"fbmcp": { ...same command and args... }}} |
Passwords are in the keyring (§5), so no env block is needed. If you put
a password into a client env block anyway, keep that file out of git.
Then start the tray for Tier 2 approvals:
C:\fbmcp\fbmcp-tray.exe C:\fbmcp\fbmcp.yaml (at every logon:
.\packaging\windows\install-tray.ps1 -Bin C:\fbmcp\fbmcp-tray.exe -Config C:\fbmcp\fbmcp.yaml).
Smoke test in the client: “call fb_ping” (shows the server name and your
identity with its max tier), then “fb_db_list”, then “fb_db_health for main”.
7. Setup B — local Windows service
Elevated PowerShell:
.\packaging\windows\install-service.ps1 -Bin C:\fbmcp\fbmcp.exe -Config C:\fbmcp\fbmcp.yaml
sc.exe start fbmcp- The service runs
fbmcp -serveas LocalSystem, restarts on failure, and stops cleanly on a service stop. - Use
secrets.dirfor passwords (§5). - Lock the folder: only SYSTEM and Administrators may write to
C:\fbmcp(the script prints theicaclscommand). A folder made underC:\lets every user change the files, and the service runs them as SYSTEM. - The tray user and local stdio clients need read/write on the state dir:
icacls C:\ProgramData\fbmcp\state /grant "<user>:(OI)(CI)M". A stdio client with the same config then attaches to the service kernel. - Readiness:
C:\fbmcp\fbmcpctl.exe ping --config C:\fbmcp\fbmcp.yaml. - Log:
<state_dir>\fbmcp.log(5 × 10 MiB; crashes infbmcp-crash.log).
For a service that admins reach over the network, use setup C instead.
8. Setup C — remote (HTTPS + API keys)
Full procedure: remote.md. Short form:
Windows server (elevated):
- Binaries and
fbmcp.yamlinC:\fbmcp\(template: remote.md §3:listen,tls,identities,allowed_origins,secrets.dir). C:\fbmcp\fbmcpctl.exe tls selfsign --out C:\fbmcp\tls --dns srv1 --ip 10.0.0.5 --grant "NT SERVICE\fbmcp"— note the pinsha256:....- One key per person or agent:
C:\fbmcp\fbmcpctl.exe apikey new alice --max-tier 2 --secrets-dir C:\ProgramData\fbmcp\secrets(printed once). With--config C:\fbmcp\fbmcp.yamlit also adds theidentities:entry and usessecrets.dir. - Passwords and agent tokens:
fbmcpctl secret set --dir C:\ProgramData\fbmcp\secrets NAME. .\packaging\windows\install-remote.ps1 -Port 8443 -AllowFrom 10.0.0.0/24 -FirebirdService FirebirdServerDefaultInstance -WhatIf, then without-WhatIf. It creates the service, the firewall rule, the folder ACLs and the Firebird service rights.C:\fbmcp\fbmcpctl.exe doctor --remote --config C:\fbmcp\fbmcp.yaml→ no FAIL;curl.exe --cacert C:\fbmcp\tls\tls.crt https://srv1:8443/healthz→ok.- Tray for the approver:
.\packaging\windows\install-tray.ps1.
Linux server (root): copy fbmcp, fbmcpctl and packaging/linux/, then
install -m 0755 fbmcp fbmcpctl /usr/local/bin/
sh packaging/linux/install-remote.sh --user firebird --port 8443 --allow-from 10.0.0.0/24 --dry-run
sh packaging/linux/install-remote.sh --user firebird --port 8443 --allow-from 10.0.0.0/24The script makes the folders, the certificate, a first config with an
admin key (printed once), the systemd unit, the sudo rule for the
Firebird unit and the ufw rule, then runs doctor --remote. Add
databases: and agents: to /etc/fbmcp/fbmcp.yaml. Approvals on Linux
(no tray): sudo -u firebird fbmcpctl approve [id]. It finds
/etc/fbmcp/fbmcp.yaml by itself, asks y/N (--yes does not ask), and
--deny <id> --reason "..." refuses.
Admin PC:
C:\fbmcp\fbmcpctl.exe secret set FBMCP_KEY_SRV1(the key you got).- Any stdio client (Claude Desktop, Cursor, Codex, ZCode): the §6 entry
with these
args:["bridge", "--url", "https://srv1:8443/mcp", "--key-env", "FBMCP_KEY_SRV1", "--pin", "sha256:<hex>"]. The pin checks the server key, so a self-signed certificate needs no CA setup. - Claude Code can also connect directly over HTTP (
.mcp.json"type": "http", headerAuthorization: Bearer ${FBMCP_KEY_SRV1},NODE_EXTRA_CA_CERTSfor a self-signed cert). Do not pass the key withclaude mcp add --header: it is stored in plain text. - Test:
fb_pingshows the server name and your identity.
Autonomous agents (OpenClaw and similar): a separate key with
apikey new openclaw-1 --profile autonomous --allow-from 10.0.0.21/32.
Read-only by default; every Tier 1 request waits for an operator, and
fb_confirm from the agent is refused. Limits: remote.md §6–7.
9. fbagent bridge and HQbird cluster (optional)
The agent tools (fb_agent_status, fb_agent_job_run, …) use the HQbird
fbagent Local API on the same host.
They are off until agents: has an entry.
- In the client: “
fb_agent_discover”. It lists the local agents withbase_url,instance_idand the suggested token name. It never reads tokens. - The operator stores the agent’s
local_api.token(notagent.auth_token):fbmcpctl secret set [--dir <secrets.dir>] FBMCP_AGENT_TOKEN_P<port>. Never type a token into the chat. - In the client: “
fb_agent_registerfor that agent” → approve in the tray. fbmcp adds theagents:entry, reloads and checks the identity. - Test:
fb_agent_status,fb_agent_instances.
fb_agent_config_set (change job settings: backup, sweep, …) also needs
config_path on the entry and read access to the live agent_config.json
plus create rights on its folder. That file holds the token and Firebird
credentials, so this grant is the operator’s decision
(agent-bridge.md). When the agent check fails, the tool
writes nothing.
HQbird cluster (RCM). The cluster tools call the HQbird cluster
manager (hqbirdrcm) on the same host. They are shown only when
fbmcp.yaml has an rcm: section (url, ca_file, user,
password_env). On a Linux HQbird server fbmcp ctl hqbird-install writes
it from /opt/hqbirdrcm/rcm.json and stores the first RCM password in the
secret FBMCP_RCM_PASSWORD; elsewhere add the section by hand
(tool-reference.md). Then:
fb_cluster_rcm_password: replace the first RCM password (Tier 1).- A new cluster: open the pairing window on each host
(
fb_agent_cluster_pairing), start the RCM on the master (fb_agent_cluster_rcm), link the hosts (fb_cluster_link), make the replicas (fb_cluster_reinit, Tier 2) and compare the data (fb_cluster_verify). Steps:prompts/cluster-setup.md. - Cluster services (switchover, failover, rejoin):
fb_cluster_serviceshows them and the checks; the moves are Tier 2. Steps:prompts/cluster-services.md.
10. Daily use
Ask the client in plain words; it picks the tools. Examples:
| Task | Tools |
|---|---|
| Is everything up? | fb_ping, fb_db_list, fb_db_health, fb_service_status |
| Read data | fb_query (read-only, rows + plan + per-table stats; a grid in MCP-Apps clients); check a draft with fb_query_check |
| Sessions, long transactions | fb_sessions, fb_transactions, fb_session_kill (Tier 1) |
| Slow query | fb_analyze_query, fb_index_advice → apply with fb_write; fb_agent_perf_report |
| Backup | fb_backup_start → job id → fb_job_status |
| Verify a backup | fb_restore_test (restores into work_dir, marks the backup verified) |
| Nightly backup + verify | fb_schedule_create with target=nightly_verify (confirmed once; setup B or C) |
| Schema / DDL | fb_schema_list, fb_describe, fb_write (each statement is classified and tiered), fb_diff_schema |
| Migrations | fb_migration_status, fb_migration_plan, fb_migration_apply |
| Server config | fb_config_get, fb_config_diff, fb_config_set (Tier 2) |
| Logs | fb_firebird_log |
| Capacity | fb_trends, fb_gstat, fb_index_stats |
| Cluster state | fb_cluster_status, fb_cluster_service |
| Move the master | fb_cluster_switchover (no data loss) or fb_cluster_failover (master lost), then fb_cluster_rejoin (Tier 2) |
Jobs are async: submit → fb_job_status. Jobs on one database run one at
a time. A job survives an fbmcp restart only as interrupted: submit it
again if it is still needed.
fb_config_set changes only its own line in firebird.conf (comments,
indentation and line ends stay), keeps the old file as firebird.conf.prev,
and never restarts Firebird. Restart with fb_service_restart (Tier 2): it
waits until the service is stopped and running again.
11. Operator tasks
| Task | Command |
|---|---|
| List / approve / deny requests | fbmcpctl approve; fbmcpctl approve <id>; fbmcpctl approve --deny <id> --reason "..." (or the tray) |
| Rotate a key | fbmcpctl apikey rotate alice --config <fbmcp.yaml> (old key stops in seconds) |
| List keys | fbmcpctl apikey list (names, sources, tiers; never values) |
| Remove a person | delete the identities: entry, then fbmcpctl secret drop --dir <dir> NAME |
| Renew the certificate | fbmcpctl tls selfsign --renew --out <dir> --dns ... --ip ... (pin unchanged, no restart) |
| Check the audit chain | fbmcpctl verify |
| Is the kernel alive? | fbmcpctl ping; fbmcpctl status (offline read of the state) |
| Allow service control | packaging/posture/verify.ps1 or verify.sh, then fbmcpctl setup --write-posture |
| Upgrade | stop the service (or the clients), replace the binaries, start, doctor |
| Remove | Windows: sc.exe stop fbmcp; sc.exe delete fbmcp; Remove-NetFirewallRule -DisplayName fbmcp-remote. Linux: remote.md §10. Keep the state dir (audit) unless the owner says otherwise |
Dev host only: .\packaging\rebuild-install.ps1 -Force rebuilds dist\,
stops and replaces the running binaries (old ones kept as *.exe~),
merges client configs and runs doctor.
12. Crash and restart recovery
Verified by the chaos harness (internal/killharness).
| After a crash or kill | On restart | Operator action |
|---|---|---|
| Job was running or queued | Marked interrupted; never resumed |
Submit again if needed |
| Pending request, not confirmed | Restored from state.json (TTL still applies) |
Confirm or let it expire |
fb_restore_replace in progress |
Finishes: restore, validate, database online; <db>.pre-restore kept |
Check fb_db_health; delete .pre-restore later |
fb_shutdown_window in progress |
Finishes the gfix -online step |
Check online |
Stale instance.lock |
Taken over automatically | None |
state.json corrupt |
Kernel refuses to start; copy kept as state.json.corrupt-<ts> |
Inspect; repair or remove state.json on purpose (fbmcpctl repair rebuilds schedule grants from the audit) |
audit.jsonl corrupt or cut |
Kernel refuses to start | §13.2: treat it as a tamper signal |
| Clock moved back | Scheduler skips; no double run | None |
| Windows service process killed | SCM restarts it in about 5 s | None |
The kernel fails closed on anything it cannot prove intact. A restart loop is a finding: read the log.
13. Incident playbooks
13.1 Something is wrong with a database
fbmcpctl doctor --config <file>: config, secrets, Firebird logins, folders, posture, schedules. Each FAIL line has aFix:.fb_db_list→fb_db_health <db>;fb_service_statusfor the instance.fb_sessions,fb_transactionsfor attachment or OAT-gap problems;fb_firebird_logfor engine errors.
13.2 Suspected tamper or corruption: audit first
fbmcpctl verify --config <file>→audit chain OK: N entriesorBROKENwith the reason.- If it fails: keep the whole state dir (it is the evidence), then
investigate. Do not delete
audit.jsonlto make the kernel start.
13.3 Restore drill (every quarter)
fb_backup_start→fb_job_status= succeeded.fb_restore_test→ the backup is verified.fb_window_openfor the database.fb_restore_replace→ approve out-of-band. Downtime until online.- Check:
fb_db_health, then a real read (fb_schema_list,fb_info). - The old file stays as
<db>.pre-restoreuntil you delete it.
13.4 Client cannot connect
fbmcpctl ping --config <file>,fbmcpctl status --config <file>.- Read the client’s MCP log (stdio) or
<state_dir>\fbmcp.log/journalctl -u fbmcp(service). - “another fbmcp instance is active”: one kernel per state dir. Piped clients attach by themselves; a second kernel fails fast by design.
In the commands above you can omit --config when the lookup in §4 finds
the file.
14. Troubleshooting
| Symptom | Meaning / fix |
|---|---|
| Tools do not appear | JSON syntax, absolute paths, full client restart; the client’s MCP log |
| One tool is missing from the list | Hidden on purpose (§1): above your max tier, no agents: entry, no migrations_dir, or no rcm: section |
| Server exits at once | fbmcpctl doctor; a missing secret; a service cannot read your keyring (use secrets.dir) |
| Tier 2 “confirmation rejected: channel policy” | Expected in the chat. Approve in the tray or with fbmcpctl approve |
| Tier 2 “no open maintenance window” | fb_window_open for the same db you pass to the tool |
fb_restore_replace denied: preconditions |
Needs a verified backup < 24 h: fb_backup_start + fb_restore_test, and a window |
| Service control denied | posture.verified missing (fbmcpctl setup --write-posture) or the account has no rights on the Firebird service |
fb_config_set: cannot write |
The account may not change that file (Linux unit ProtectSystem=strict; Windows ACL). Edit by hand, or see remote.md §8 |
| Tray shows nothing | It shows only requests that need an operator; check the tray user’s rights on the state dir and tray.log there |
| 401 (remote) | Wrong key, or the server did not load it: apikey list, doctor --remote |
| 403 (remote) | Source IP not in allow_from / firewall, or an Origin header not in allowed_origins |
| 429 (remote) | Limits; wait for Retry-After |
| Pin mismatch | The server key changed; get the new pin from the server admin |
agent-unreachable |
Agent service down, local_api.enabled: false, or wrong base_url (default 127.0.0.1:(10000 + firebird port)) |
agent-auth |
The stored token is not the agent’s local_api.token; set it again |
agent-timeout on restarts |
agents[].timeout_seconds (130) must be larger than the agent’s restart_timeout_seconds (120) |
Error loading plugin MySQLEngine … on attach |
On HQbird this usually means the database file does not exist; check the path |
Bare DROP INDEX in fb_write needs a window |
It is Tier 2 there; use fb_index_drop (Tier 1; refuses constraint indexes) |
| New index not used | Run statistics (fb_index_rebuild action=statistics); ORDER BY … DESC needs a DESCENDING index |
| Tools fail for one database only | It is offline (fb_db_health); tools degrade per database |
15. State directory reference
| Path | Meaning |
|---|---|
state.json (+ .tmp, .corrupt-*) |
Pending requests, jobs, backup catalog, workflows, schedules. Corrupt ⇒ refuses to start |
audit.jsonl (+ .head) |
Hash-chained audit log; .head detects truncation |
events.jsonl |
Job events (job.succeeded, job.failed with the message) |
approvals\, denials\ |
Out-of-band decisions (tray / fbmcpctl approve), used within ~2 s |
instance.lock |
One-kernel lock |
attach.addr, attach.token |
Attach endpoint for extra stdio clients |
fbmcp.log, fbmcp-crash.log |
Service log (rotated) and crash output |
query-log.jsonl |
Every fb_query call |
trends\ |
Sampler history for fb_trends |
config-journal\ |
fb_config_set history (old and new value) |
posture.verified |
Marker that allows service control |
tray.log |
Tray errors |
16. Walkthrough status
Pending (M1): a walkthrough on a fresh Windows host using only this guide.
Findings go back into this file. Lab results for remote mode (Linux host,
fresh VM, Windows service): findings/phase12-lab-2026-11-03.md.