mcp server
bellcrank exposes a remote MCP server. Connect your agent, and it builds the axle, runs the rig and reads the solver's channels back — the same API your own scripts use, with the same rules.
Exhibit 002
The hand-clicked parameter sweep
Change a hardpoint, click Run, export to spreadsheet, paste, repeat. Survived every software generation on habit alone.
the loop
When a client connects, the server hands it the recommended order of calls, so your agent can start at the right tool.
list_templates indexes the catalog, and
your agent pulls the full structure of the option it picks.
create_axle builds the axle from that topology.
A saved motion defines the sweep to run.
run_kinematics queues the run and answers
at once. A wait tool does the polling for your agent.
get_result starts with a compact summary,
then pulls back just the channels your agent asks for.
You: On "FS26 front", raise both upper-wishbone inner pivots by 5 mm. Compare roll center height, camber gain and bump steer over ±50 mm of heave, before and after.
whoami() list_assets(kind="axle") # "FS26 front" list_assets(kind="motion") # "Heave ±50 mm", two_post_rig run_kinematics(runs=[{type:"two_post_rig", assembly_id:"3f2c…", motion_id:"9a41…"}]) wait_for_result(result_id="c07e…") # complete — baseline get_asset(kind="axle", id="3f2c…") # definitions + revision_checksum update_axle(id="3f2c…", patch={ definitions:{suspension:{…, values:{…, "double_wishbone.point.uca_inner_fore_l.position": [x, y, z+0.005], …}}}, parent_checksum:"e81b…"}) # all four uca_inner points, +5 mm in z run_kinematics(runs=[{type:"two_post_rig", assembly_id:"3f2c…", motion_id:"9a41…"}]) wait_for_result(result_id="d5a2…") # complete — revision 2 get_result(result_id="c07e…", channels=[ "axle.roll_center_z", "front.wheel.camber_gain_wheel_l_post_l", "front.wheel.toe_wheel_l", "front.wheel.wheel_travel_wheel_l"]) get_result(result_id="d5a2…", channels=[…same four…]) Agent: Roll center height, camber gain and bump steer (toe against wheel travel), for both revisions, side by side.
Ask it things like
"Sweep ±30 mm of roll on FS26 front. How far does the roll center move, sideways and vertically?"
"Drop both tie-rod inner points 3 mm and show toe against wheel travel before and after."
"On the FS26 car, what anti-dive and anti-squat do we have over ±40 mm of heave?"
resources and prompts
Attach the guide and the channel vocabulary once, and a single prompt starts the whole loop for you.
bellcrank://guide bellcrank://channels bellcrank://templates/{name} run-and-summarize compare-revisions connect
Mint a key under Account → API keys, then add the server:
claude mcp add --transport http bellcrank https://api.bellcrank.io/mcp \
--header "Authorization: Bearer bc_api_..."Cursor and any other client that can send a header use the same URL. More in the MCP guide
faq
Claude Code, Cursor and any other MCP client that can send a static header connect the same way: one URL and an API key, no browser sign-in.
bellcrank's solver computes every channel, and the tool reads it back. Your agent chooses what to run and which channels to read.
No. A queued run spends credits from your normal pool, the same as it would from the app.
Yes. Every plan can mint a key and connect an agent; your plan's usual limits still apply.
No. A result starts as a compact summary, and an answer too large to be useful is refused with a hint to narrow it, never dumped in full.
Yes, and every delete is marked destructive, so a client that honors annotations asks you first.
Create an account, mint a key, run one command. The free tier is permanent and needs no payment details.