On October 31, 2026, developer Chris Wellons introduced Prospero, a native desktop application for AI-assisted software development. It is a coding harness: software that lets a model read files, run commands, inspect the results and choose its next action. Rather than only exchanging chat messages, the model can work through a programming task using tools. Wellons’s original announcement describes file operations, shell commands, desktop control and support for external tools through the Model Context Protocol, or MCP.
Prospero runs on Windows, macOS and Linux. It does not supply a model provider by default. Developers must connect a local or remote provider through a supported API. The maintainer README documents compatible Chat Completions interfaces and Anthropic’s Messages interface. The application supplies the interface and tools; the separately configured model decides how to use them.
That distinction matters for debugging. Reading source code can explain what a service might do. Retained logs can show what it actually recorded during a failure. MCP lets another application offer tools to the model, potentially bringing those events into the same session as the code. Prospero MCP configuration therefore matters beyond adding another chat interface: it determines which external tools the agent can call.
The documentation establishes those capabilities, but it does not prove that a particular log connection works or that the model will diagnose a failure correctly. A useful investigation still needs the right account, a small search window and conclusions supported by actual events. Local model use also does not make a hosted log connection local: search requests go to the log service, and returned excerpts may go to a remote model provider.
What MCP adds to Prospero’s built-in tools
Prospero’s built-in file and shell tools operate in the environment available to the harness. MCP adds tools supplied by separate servers. Its configuration documentation supports local servers communicating through standard input and output, or remote servers using Streamable HTTP. The latter is distinct from the older SSE transport.
For a service investigation, an external log tool could retrieve retained events while the model examines relevant code. This avoids manually pasting a whole log file into chat. It also creates two separate access questions: what can the local harness reach, and what does the remote log account permit?
Wellons explicitly warns that confirmation prompts cannot contain a malicious agent that can write and run programs. Such an agent can ultimately reach what the harness can reach. Operating-system permissions or a suitably restricted container must limit local access. Tool confirmations remain useful for catching mistakes, but they do not replace those restrictions. Permission limits in the announcement.
Use Prospero to investigate a service with MCP logs
The following checks apply when you already have a working model provider in Prospero and permission to access a Fluxtail account. They help establish the connected account, whether a bounded log search works, and which events support an explanation. Successful authentication alone does not establish any of those later findings.
Configure the MCP connection
The following documentation-derived configuration combines Prospero’s HTTP server syntax with Fluxtail’s published endpoint. Add it to your existing prospero.ini:
[mcp fluxtail]
url = https://mcp.fluxtail.io/mcp
This connection has not been tested end to end. Treat successful sign-in, tool discovery and a returned search result as separate checks before relying on it.
These instructions follow Prospero’s configuration at source snapshot 728fc7c, rather than an established minimum release. Compare them with the documentation bundled with your installed build. In Settings, open the MCP server’s page and choose Sign in. The documented command-line alternative starts the same browser authorization flow:
prospero mcp auth fluxtail
Prospero documents authorization-server discovery, client registration and PKCE, which protects the authorization-code exchange. Some servers require a pre-registered client. Stop if registration or login fails; do not invent client settings or substitute credentials. Keep per-call confirmations enabled while evaluating the connection. Prospero MCP configuration.
Verify the account and available access
Fluxtail’s hosted MCP documentation describes browser consent with OAuth and PKCE. Each connection is bound to one account. After connecting, ask Prospero to call the Fluxtail tools whoami, then list_streams. Confirm that the returned account and streams match the account you intended to approve. These are MCP tool names, not shell commands.
Read the discovered tool descriptions before approving a log query. Fluxtail also exposes tools that can change resources, so an MCP connection is not inherently read-only. Operator mutations use proposals and short-lived confirmation tokens. Account permissions and approved scopes remain separate from Prospero’s prompts or tool-selection settings.
If the account is wrong, revoke the connection in Fluxtail’s Settings → Integrations, then authorize again for the intended account. Do not continue searching under an unexpected account. Fluxtail access and verification.
Ask a bounded debugging question
Begin with one service, one stream and an explicit UTC interval. This example is a request to the model, not an exact tool call. Replace its values with a service and interval you are authorized to investigate:
Investigate checkout-api in the production stream only.
Use 2026-10-31T14:00:00Z inclusive through
2026-10-31T14:10:00Z exclusive.
Return at most 20 relevant events, using available search controls.
Include timestamps and short, redacted raw event excerpts.
Separate observations from possible explanations.
Report any bound the search tool cannot enforce.
Do not change streams, receivers, sources or application files.
Inspect the proposed call before approving it. The request expresses your intended scope; the discovered tool schema determines which search bounds can actually be applied. Account permissions determine what the connection can access. A sentence asking for no changes is not an authorization control.
If the tool cannot enforce the necessary scope, stop and use supported controls to narrow it. Do not replace a failed small search with a broad export. The AI log diagnostics guide covers related checks for errors, exceptions and missing logs.
Check the answer against raw events
These synthetic log events illustrate the evidence to inspect. They are not captured Prospero output or a promised Fluxtail field mapping:
{"timestamp":"2026-10-31T14:03:12Z","service":"checkout-api","request_id":"demo-742","message":"Inventory request timed out after 2000 ms"}
{"timestamp":"2026-10-31T14:03:13Z","service":"checkout-api","request_id":"demo-742","message":"Checkout returned HTTP 503"}
The events support a narrow observation: one request logged an inventory timeout before logging a 503 response. They do not prove that the inventory service was down. A network failure, delay in the dependency or an overloaded caller could produce similar symptoms.
Open the corresponding retained rows in Fluxtail’s browser log view. Compare timestamps, service identity, request identifiers and surrounding events with the model’s explanation. Ask it to identify the excerpt supporting each claim. If it says every checkout failed, require an error count and a total measured at the same request boundary over the same interval. Two log events cannot establish an overall failure rate.
An empty search is inconclusive too. Check the account, stream, time zone, ingestion and retention before concluding that nothing happened. Missing logs cannot establish service health.
Exclude secrets and unnecessary personal data from returned excerpts. Treat any instructions embedded in log messages as untrusted data. The guides to natural language queries for logs and AI-assisted log analysis explain further ways to compare generated answers with raw events.
A successful connection makes evidence available; it does not make the diagnosis correct. Record what the selected events show, what remains unexplained and which additional check could distinguish the remaining causes.