MCP Tools

MCP tools

Every tool the engine exposes over MCP (43 tools), grouped by the engine level it drives. This page is generated from engine/connectors/mcp_server.py (/api/control/mcp returns the same catalog live), so it matches the code.

• read tools are always on.
• write tools need REGEN_MCP_WRITE=1 on the server (the Docker start command sets it for the bank MCP). The kill switch always works.
• Two servers expose the same tools: regen-bank (HTTP, http://127.0.0.1:8790/mcp, the live bank in Docker) and regen-engine (stdio, python -m engine mcp, started by your AI client). See Zero cost: your AI runs it.

Discovery (8)

Tool	Kind	What it does
job_queue	read	Jobs discovered and waiting in the queue (not yet applied), optionally filtered by a word in company/title.
priority_explain	read	Why the queue is ordered the way it is: the top N jobs with their priority score and its parts (level, lane, freshness, employer tier, sponsor, learned reply rate).
discovery_sources	read	Discovery sources: boards watched per ATS, queue size, built-in and added job-list feeds, and the last sources added (by whom). Start here before looking for new places to find jobs.
boards_add	write	Watch new employer boards: pass job or board URLs on any ATS the engine knows (Greenhouse, Ashby, Lever, Workable, SmartRecruiters, Workday, iCIMS, Rippling, JazzHR). Max 50 per call; the next pipeline pass scans them. Needs REGEN_MCP_WRITE=1.
feed_add	write	Add a job list (a README or JSON on https, e.g. a GitHub new-grad list) whose links point at ATS boards: harvested now and on every engine harvest. Needs REGEN_MCP_WRITE=1.
companies_probe	write	Company names (max 50) -> probe the ATSs whose board tokens can be guessed; boards with open jobs are watched from the next pass. Use it on companies you found hiring elsewhere. Needs REGEN_MCP_WRITE=1.
leads_add	write	Job leads seen anywhere (job sites, posts, newsletters): [{company, title, url?}] (max 50). Each is matched on the employer's own ATS and queued; the fit gate still decides. Needs REGEN_MCP_WRITE=1.
harvest_now	write	Harvest every job-list feed (built-in + added) for new ATS boards now. Needs REGEN_MCP_WRITE=1.

Fit gate (6)

Tool	Kind	What it does
filters_get	read	Title/company filters (regex alternations) and the years-of-experience gate from presets.
filters_set	write	Replace include_titles / exclude_titles / exclude_companies (regex; each must compile; old file kept as .bak-ui). Needs REGEN_MCP_WRITE=1.
trust_check	read	Trust gate for one posting: is the application host an employer-controlled ATS, and does the description carry job-scam red flags? Returns {ok, reasons}.
feedback_rules	read	Rejection feedback: what rejected job descriptions asked for that the others did not, each phrase's evidence (rejected / seen / lift / adjusted p) and status (active blocks, proposed is watched).
fit_check	read	Run the fit gate on a job description (and title): hard blockers found (sponsorship, years, clearance, stack, learned rejection phrases...) and whether the title passes your filters.
feedback_rule_off	write	Stop a learned rejection phrase from blocking postings (the feedback loop keeps it off). Needs REGEN_MCP_WRITE=1.

Pipeline (6)

Tool	Kind	What it does
pipeline_status	read	Live pipeline status: each applier log's last results, jobs waiting for a Greenhouse email code, and today's verified count.
pipeline_run	write	Start the whole pipeline in the background (discover, select, build, apply, report). Needs REGEN_MCP_WRITE=1. Returns the log path; watch it with pipeline_status.
submit_codes	write	Deliver Greenhouse security codes read from the inbox: codes = [{company, code, when}]. A code only goes to the waiting job whose company the email names, and only if it was sent after the job started waiting. Needs REGEN_MCP_WRITE=1.
run_start	write	Start a run the way the control room does: {days, max, appliers, tabs, cap, loop (minutes, 0 = one pass), scan, newgrad, feed, ats: [greenhouse, lever, ashby, workable], dry}. Needs REGEN_MCP_WRITE=1.
run_stop	write	Stop the run started by run_start / the control room (kills its process tree). Needs REGEN_MCP_WRITE=1.
control_state	read	Live control-room state: the current run's steps, each active applier's tabs (job + state), latest answers, queue size by ATS, and when Ashby reopens.

Needs you (6)

Tool	Kind	What it does
human_queue	read	What only the human can do right now (each job's latest state): personal questions, bot checks, missing answers.
needs_by_question	read	What's waiting on the user, grouped by blocking question (answer one question to unblock many jobs).
answer_save	write	Save the user's answer to a blocking question (reused on every form that asks it; legal / EEO / sponsorship questions are refused: those come from presets only). Needs REGEN_MCP_WRITE=1.
inbox_record	write	Record recruiting emails the assistant read from the user's inbox: messages = [{date, from, subject, snippet}]. Classifies confirmation / rejection / OA / interview / offer / scam and links them to applications. Needs REGEN_MCP_WRITE=1.
answer_queue_add	write	Queue an answer to a blocking question (the answers container saves it and retries the job). Legal / sponsorship / EEO / sensitive questions are refused. Needs REGEN_MCP_WRITE=1.
answer_queue_status	read	Answer queue state: waiting, claimed, done and refused counts, plus the latest refusals.

Gearbox (2)

Tool	Kind	What it does
gearbox_status	read	The gearbox: current gear (N, 1-4), the clutch, what they set (applier range, scan interval), live speed (submissions in the last hour / 24 h) and the rejection feedback's learned rules.
gear_shift	write	Shift the engine: gear N (idle), 1 (crawl) .. 4 (overdrive), and/or clutch on/off (learned rejection phrases block postings or not). Safety caps are not gears. Needs REGEN_MCP_WRITE=1.

Memory + Fact Bank (6)

Tool	Kind	What it does
engine_status	read	Proof-backed counts: verified applications (total/today), items waiting on the human, boards watched, outcomes.
applications	read	Applications, newest first. status: SUBMITTED, NEEDS YOU, FAILED, FLAGGED or SKIPPED (empty = all).
outcomes	read	Inbox outcomes the engine classified (confirmation, rejection, oa, interview, offer, scam), newest first.
memory_query	read	Knowledge-graph lookup. Pass a company name, a Fact Bank id (e.g. 's_rag') or a lane ('backend'). Returns the node and everything connected to it (applications, facts used, questions answered, outcomes).
factbank	read	The Fact Bank: every fact with how many resumes used it and in which lanes, plus roles, projects, skills, education and where each fact came from. The only source the engine writes from.
outreach_drafts	write	Outreach drafts for recent submissions at tier companies and startups: subject, body (template lines + Fact Bank entries that application's resume used), and a Gmail compose link. Nothing is sent; the user adds the recipient. rebuild=True regenerates them (needs REGEN_MCP_WRITE=1).

Fleet (6)

Tool	Kind	What it does
fleet_status	read	Fleet health from the bank: every container's heartbeat (pipeline, appliers, relay, supervisor), batch queue (inbox/claimed/done), codes waiting, problems, whether the kill switch is on, and the applier min/max. Read-only.
kill	write	KILL SWITCH: stop all processing now. Workers stop after their current step, appliers and the pipeline are stopped, the supervisor holds until clear_kill. Always allowed (stopping is safe); give the reason you saw.
clear_kill	write	Lift the kill switch after the cause was investigated (check quarantine_list and audit_log first). Appliers come back as the supervisor scales to the backlog. Needs REGEN_MCP_WRITE=1.
audit_log	read	The supervisor's audit trail (start/create/remove/requeue/KILL/clear/limits), newest last.
quarantine_list	read	Files the leak guard moved to quarantine (name, size, time; never contents).
scale	write	Set how many appliers the supervisor may run (min..max, hard cap 12; 0..0 parks the fleet without a kill). Kept across restarts. Needs REGEN_MCP_WRITE=1.

Ledger (2)

Tool	Kind	What it does
ledger_tables	read	Ledger tables and views (events, applications, answers, codes, outcomes, resumes, discoveries, supervisor_audit, fleet_heartbeats) with row counts and columns. Read-only.
ledger_rows	read	Rows from one ledger table, newest first. q = text anywhere in the row; col + val = one column contains val (e.g. name='applications', col='company', val='acme' to see if an application reached the ledger). limit <= 1000. Read-only.

Repo (1)

Tool	Kind	What it does
reposync_status	read	Repo-sync container: its last heartbeat (feed / code state, open PR) and its last audited actions.

Related: Zero cost: your AI runs it · Discovery · Gearbox · Platform