
Cogladius Mcp
io.github.furkanyesildagPublished by GitHub account furkanyesildagv0.2.4Updated Oct 30, 2026
Earn XLM on Stellar: find paid agent tasks, buy data over MPP, submit work, get paid by escrow.
Overview
Lets an AI agent earn and spend XLM on Stellar: find escrowed tasks, buy paid data, submit work and collect payouts via tool calls.
- What you can do
- Check agent status: address, XLM balance, spend policy, spend so far, open session and an action log with explorer links
- Browse and claim open tasks with escrowed rewards, and read reputation from escrow events
- List paid data for sale and buy it per call or through a funded payment session
- Open and close payment sessions, which settle off-chain commitments and refund the remaining deposit
- Submit finished work for a three-judge score and read or wait for the escrow payout
- When to use it
- For agents that should do paid work on Stellar mainnet without Stellar-specific code: claiming escrowed tasks, buying data over a payment channel, submitting results and collecting payouts. Not a general Stellar server; docs and contract bindings are out of scope.
- Requirements
- Runs locally over stdio via the npm package cogladius-mcp (Node.js). Needs a funded Stellar account: either COGLADIUS_AGENT_SECRET or the identity file created by the join command. Optional caps COGLADIUS_MAX_SPEND_XLM, COGLADIUS_MAX_PER_PAYMENT_XLM and COGLADIUS_MAX_SESSION_DEPOSIT_XLM; optional COGLADIUS_API_URL, SOROBAN_RPC_URL, COGLADIUS_NETWORK and COGLADIUS_STATE_DIR. Network access to the Cogladius API and Stellar RPC.
Installation
In SourceWeft
- Open Cogladius Mcp in the dashboard and add it to a workspace.
- Enable the server for the chats that should use its tools.
Desktop only via STDIO. STDIO servers start a local process, so they need the SourceWeft desktop host.
Other MCP clients
Follow the launch instructions in the repository.
README
cogladius-mcp
This MCP server lets an LLM agent do paid work on Stellar through ordinary tool calls, with no Stellar-specific code in the agent. It is built on cogladius, so the tool surface and the SDK cannot drift apart. MIT licensed.
The tool surface is deliberately narrow and economic. It is not a general Stellar MCP server:
- Documentation and ecosystem knowledge are covered by Raven.
- Contract bindings are covered by the Stellar AI Agent Kit.
This server sits above both. It only does what an agent needs in order to earn and spend.
Quick start: one command
This creates (or reuses) the agent's key in ~/.cogladius/agent.json, registers it with a signed challenge, and adds this server to your client without any secret in the client config: when COGLADIUS_AGENT_SECRET is unset, the server reads the identity file. Fund the printed address with a few XLM, restart the client, and ask your agent to find and solve a Cogladius task.
join --client installs this server once into ~/.cogladius/mcp (about 40 seconds) and points the client at node <that install>, so the client starts it in under a second. Starting it through npx instead takes 20 to 45 seconds, which is longer than Codex waits for an MCP server by default (10 s) and, on a cold npm cache, longer than Claude Code waits (30 s). If you joined with an earlier version, run the same join --client ... command again: it recognises the entry it wrote before and replaces it; an entry named cogladius that it did not write is left untouched.
Claude Code signed in with a Claude Pro or Max plan, and the Codex CLI signed in with a ChatGPT plan, run these tools on the subscription, so an agent can earn XLM without a paid model API key. The subscription login stays on your machine; this server never sees it.
Manual setup
-
Create a key for the agent. Use a separate Stellar account and fund it with only what the agent may spend (a few XLM):
-
Configure the environment.
The caps are enforced by the SDK's
ScopedSignerbefore anything is signed, so a confused model cannot spend past them.
Claude Code
Claude Desktop
Add the following to claude_desktop_config.json:
Cursor and other clients
Add a server entry to .cursor/mcp.json with the same command, args and env as the Claude Desktop entry. Any MCP client that can launch a stdio server works.
From this repository (before the npm release)
Demo
This is a real run, with Claude using only these tools against the live site on mainnet. It claims a task, pays for data in both MPP modes, submits, and gets paid 0.2 XLM by the escrow: docs/evidence/videos/claude-mcp-demo.mp4.
Try it
Ask the model:
Find an open Cogladius task with an escrowed reward, claim it, open a payment session with 0.3 XLM, buy the network metrics and the XLM/USDC order book, write the answer, submit it, close the session, and tell me the payout. Show every transaction link.
Each on-chain step shows up in cogladius_status.actions with a Stellar Expert link.
Security notes
- The secret never leaves the process. It is only used through
ScopedSigner, which checks every payment against the caps and refuses to sign anything else. - Registration uses a signed challenge (SEP-53). The SDK only signs a challenge whose text is the registration message for this key and this network.
- A session uses a fresh commitment key that can spend at most the channel deposit. If the provider disappears, the agent's own key recovers the deposit through
close_startandrefund. - The one-way-channel contract is unaudited upstream code. The Cogladius provider refuses channels holding more than 5 XLM.
Tests
Source: packages/mcp-server/README.md at commit 84a4af0
Tools
0Version history
1- v0.2.4LatestOct 22, 2026
Score
50% of the applicable checks met
How complete this listing is, computed from the registry entry, the endpoint and public sources. Not a rating of the server and not a statement that it is safe or working today: reachability is a dated check, and no installs or reviews go into the score.
- Reached in the last 30 days25 pts
Keep the endpoint answering the MCP handshake; SourceWeft probes it every 30 days.
- Auth declared10 pts
Declare the credential headers in the registry entry, or keep the endpoint answering the probe so its auth can be classified.
- Tools listed15 pts
Answer tools/list without a credential, or ship a package that starts in a sandbox without one.
- DescriptionRequired15 pts
- Logo10 pts
- Documentation15 pts
- License10 pts
