Quick Start

Pick your agent harness to use Om

Install the Om Discovery plugin with npx, or connect the hosted Om MCP in Claude, Codex, Cursor, VS Code, Copilot, Gemini, Grok, or another remote MCP client. After OAuth, you score proteins, start Discovery, and read Wallet Credits from the agent. You do not need the Python SDK for this path.

Pick an agentExample prompts

Om MCP

One URL. Sign in once.

Add the hosted Om MCP in your agent, complete the browser login, restart the client, then ask for om_status. Use the in-app login. You do not paste an API key into this setup.

FieldValue
MCP URLhttps://agents.omtx.ai/mcp
TransportStreamable HTTP
AuthOAuth 2.0 in the client
Pluginnpx skills add omtx-ai/omtx
Om accountSign in at omtx.ai with the same email you will authorize
Sourcegithub.com/omtx-ai/omtx

choose one harness

Pick your agent harness to use Om

Verified paths are Claude and Codex. Cursor, VS Code, Copilot, Gemini, Grok, and other remote MCP clients use the same hosted URL. If a client cannot complete browser OAuth, use the Om API with an API key instead.

one command

Install the Om Discovery plugin

Add the hosted Om MCP and the Om Discovery skill from GitHub, then complete OAuth in your agent. Or pick a harness below for the native client command.

npx skills add

npx skills add omtx-ai/omtx

npx plugins add

npx plugins add omtx-ai/omtx

The plugin points at https://agents.omtx.ai/mcp. Source: github.com/omtx-ai/omtx. Claude Code can also run /plugin marketplace add omtx-ai/omtx then /plugin install om-discovery@omtx.

Claude app

Claude.ai and Claude Desktop

Verified
  1. 1Open Customize → Connectors (claude.ai/settings/connectors).
  2. 2Add a custom connector named Om.
  3. 3Paste https://agents.omtx.ai/mcp, then Add and Connect.
  4. 4Finish Om OAuth in the browser and enable Om in the chat + menu.

Start a new chat, enable Om, then ask for om_status.

Read docs

copy these into the agent

Prompt templates and example replies

Start with Hello World, then the highest-value work in this order: score a protein, score a proteome, order molecules, fine-tune and rescore, submit Discovery, launch Hub jobs, then Diligence. Replies are example shape, not live results. Paid actions wait for an explicit confirm.

0. Hello World

om_status, credits_get, pricing_get

You say

Run om_status. If Om MCP is healthy, show my Wallet Credits
and list the highest-value things I can do next.

Example agent reply

Om MCP is connected. Wallet Credits are available on this account. Highest-value next steps: 1. Score a protein against Om Accessible Space with LULA-2 2. Score one molecule across a protein panel (proteome score) 3. Order selected molecules 4. Fine-tune LULA on your labels and rescore 5. Submit a Discovery program with AS-MS 6. Launch Hub jobs such as Boltz-2 7. Use Diligence to learn more about a target

1. Score a protein

lula2_score, jobs_wait, jobs_get_top_scored_molecules

You say

Score this protein with LULA-2 against Small Om Accessible Space,
n=1000, and return the top 25 hits. Wait until scoring finishes.
Also map hotspot residues for the top hit with lula2_residue_map.

[paste a methionine-start amino acid sequence]

Example agent reply

I will call lula2_score with source=om and the Small Accessible Space tier, wait on the job, then return the top 25 molecules. For the top hit I can run lula2_residue_map (protein sequence + that SMILES) to show ligand-conditioned hotspot residues. Nothing is ordered yet.

2. Score a proteome

artifacts_upload_bytes, lula2_proteome_score, jobs_wait

You say

I have one SMILES and a FASTA of proteins.
Upload the FASTA, then score this SMILES across that protein panel with LULA-2.
Return the top proteins this molecule is predicted to bind.

Example agent reply

I will upload the FASTA with artifacts_upload_bytes, then call lula2_proteome_score with that protein_artifact_id and your SMILES. When the job completes I will return the top protein hits. This is one molecule against many proteins, the inverse of Accessible Space scoring.

3. Order molecules

molecule_accessible_space, molecule_fulfillment_pricing, molecule_fulfillment_order

You say

Take the top 10 LULA-2 hits from the last score.
Quote Molecule Fulfillment, then wait for my confirm before placing a
Wallet Credits order to ship those molecules.

Example agent reply

I will look up fulfillment pricing and quote the 10 selected rows. I will not call molecule_fulfillment_order until you confirm the Wallet Credits total and a shipping address. Physical molecules are a separate order from Discovery validation data.

4. Fine-tune LULA and rescore

lula2_finetune, finetuned_derived_models_list, lula2_score

You say

I have a labeled CSV of binders and non-binders as an Om artifact.
Fine-tune LULA-2 on it, then rescore the same protein against Small
Accessible Space with that derived model.

Example agent reply

I will call lula2_finetune with your csv_artifact_id, wait until the derived checkpoint is ready, then call lula2_score with derived_model_id set to that checkpoint. That rescore uses your labels instead of the base LULA-2 weights.

5. Submit a Discovery request

discovery_launch_quote, discovery_launch_order, discovery_launch_get, credits_get

You say

Quote a Discovery program for this protein: 10 molecules, Small space,
AS-MS included. Show Molecules / Validation / Total Wallet Credits.
Do not place the order until I confirm.

[paste a 300–1500 aa sequence starting with M]

Example agent reply

discovery_launch_quote for 10 molecules, Small space, AS-MS on: - Molecules: 50,000 Wallet Credits - Validation: 30,000 Wallet Credits - Total: 80,000 Wallet Credits This is the same product as the website Discovery flow: LULA ranks Om Accessible Space, then Om runs AS-MS and returns binder and non-binder data. I will only call discovery_launch_order after you confirm. After it is placed, discovery_launch_get checks status.

6. Launch Hub jobs

hub_models_catalog, hub_boltz2, hub_submit, jobs_wait

You say

Show active Hub models. Then launch Boltz-2 on this protein-ligand pair
and wait until the job finishes. Return artifacts when ready.

Example agent reply

I will call hub_models_catalog, then hub_boltz2 (or hub_submit) with your inputs, then jobs_wait. Boltz-2, BoltzGen, Chai-1, OpenFold3, DiffDock, and other active Hub models are available from the same Om MCP session.

7. Learn more with Diligence

diligence_search, diligence_gather, diligence_deep_diligence

You say

Use Om MCP to search KRAS and summarize the therapeutic strategy,
current clinical context, and the highest-signal citations.

Example agent reply

I will run diligence_search, then gather or deep-diligence only if the first result warrants it. The brief will stay inside retrieved evidence. If a claim is not in the sources, I will say so.

If a tool does not appear

  • Restart the client after adding the server. MCP lists load on startup.
  • Finish OAuth in the client. Claude uses /mcp. Codex uses codex mcp login --scopes email omtx. Other clients use their own MCP connect or auth command.
  • Confirm the URL is exactly https://agents.omtx.ai/mcp with https and /mcp.
  • Use the same email on Om that you authorized in the browser.
  • Om asks for email only. If the client requests openid, profile, or phone, cancel instead of approving.
TroubleshootingTool listDiscovery