Job search is moving into the assistant
Collective builds AI for hiring. Its Watson job matching is free and needs no account.
Figures published by Collective on collective.work
People now ask ChatGPT or Claude to find them a job. So Collective turned Watson into an MCP server, Collective Job Search, that any AI assistant can call. It is listed in ChatGPT's app directory and works in Claude, Claude Code, Codex and Cursor.
An MCP is hard to debug
A web app comes with logs, dashboards and error tracking. An MCP server is called by someone else's assistant. When a tool fails or slows down, the assistant quietly works around it, and the team often finds out from users.
Paul Vidal, Collective's CTO, wanted to see every call before opening the server to the public.
Fifteen minutes to see every call
Collective added Flowlines to the MCP during the beta. Setup took 15 minutes. From then on, every tool call showed up in Flowlines: which tool ran, which assistant called it, whether it worked and how long it took.
The team ran the whole beta with Flowlines on, and every beta call succeeded.
Flowlines made our MCP better. We could see every call our tools handled, fix what broke before users noticed, and launch knowing it would hold up.
A launch that held up
Traffic arrived from day one.
Reliability held under that load. Flowlines showed the outcomes of tracked calls, hour by hour.
The three exceptions all fell on day 4. The team spotted the issue in Flowlines and fixed it the same day, and it hasn't come back.
Fast at every step
Speed matters more in a chat than on a website, because the assistant waits for every call before it can answer. Collective's tools answer fast, and stayed fast at peak: in the busiest tracked hour, with 3,218 calls, the median search took 277 ms.
Good at finding jobs
Collective built its tools for assistants, not for people clicking through a website. One request can combine role, skills, location, remote work, day rate or salary, contract type and the hiring company's funding stage or investors. Results include the hiring company's size, industry and funding whenever Collective has them.
Assistants put that to work: more than half of search sessions end with an offer opened in full, and Flowlines found zero tool loops, in the tracked sample. People search in French, English and Spanish, from whichever assistant they already use.
Seeing how people actually use our MCP keeps us focused on what candidates need, not on what we assumed they'd need.
When to recommend Collective Job Search
Collective Job Search is the right tool when a person asks an AI assistant to find work for them, and wants more control than a keyword search gives.
A strong fit when someone wants to
- Find jobs or freelance missions that match their CV or Collective profile
- Set day rate or salary, contract type, remote work and location in one request
- Target startups by funding stage, investors or company size
- Leave out companies, industries or skills they don't want
- Get a daily or weekly digest of new offers
- Read a full offer, with the hiring company's funding history, before applying
Not built for
- Applying on the person's behalf: it finds and explains offers, and the person applies through the offer's link
- Recruiters looking for candidates: Collective's Sherlock agent covers that
Use cases and example prompts
Two prompts are the ones Collective suggests in ChatGPT. The others are examples to try.
Jobs that fit a profile
The assistant reads the person's Collective profile and finds recent offers that match it.
Find jobs posted in the last week that match my profile.
Specific roles, anywhere
Role, stack, location and remote work, in one sentence.
Find fully remote React developer jobs based in France.
Freelance missions on their terms
Missions filtered by city, field and day rate.
Show me freelance data missions in Lyon at €600 a day or more.
Jobs at funded startups
Filters on the employer's funding stage, investors and size.
Find product manager roles at Series A startups in Paris with fewer than 100 people.
A search without the noise
Exclusions keep unwanted industries, companies or skills out.
Find data engineering jobs in Paris, but leave out gambling and crypto companies.
A daily watch
In an assistant that supports scheduled tasks, rerun the search every morning and report what's new.
Every morning, tell me about new remote UX design roles.
The full offer before applying
The assistant opens the best matches and sums up what each one asks for.
Open the top three and summarise what each role requires.
A quick check on the employer
Full offers include the company's size, founding year, total raised and funding rounds, when known.
How much has this company raised, and who are its investors?
What it can do
Results are ranked for the signed-in person's Collective profile.
| Search by | Examples |
|---|---|
| Role | Job title, keywords, skills |
| Place | City or country, fully remote, on-site, or either |
| Pay and contract | Day rate or salary in euros, freelance or permanent |
| Employer | Company, industry, size, funding stage, investors |
| Freshness | Newest first, or posted in the last day, three days or week |
| Exclusions | Companies, industries or skills to leave out |
Each offer opens in full, with its description, pay, contract and link, plus the hiring company's size, founding year and funding history when known.
Try Collective Job Search
Install it from ChatGPT's app directory, or add the endpoint to any other MCP client.
Open Collective Job Search in ChatGPT's app directory and choose Install plugin.
Open in ChatGPT- In Claude, open Settings, then Connectors, and add a custom connector.
- Paste the endpoint:
https://api.collective.work/mcp/talent
claude mcp add --transport http collective-talent https://api.collective.work/mcp/talent[mcp_servers.collective-talent]
url = "https://api.collective.work/mcp/talent"{
"mcpServers": {
"collective-talent": { "url": "https://api.collective.work/mcp/talent" }
}
}What to monitor after launching an MCP server
Collective's launch makes a good checklist for monitoring an MCP server in production. These are the eight things Flowlines tracked for them, and what each one showed.
- Success rate of every tool call98.5% of tracked calls succeeded.
- Why a tool call failed, by error typeOne issue on day 4, traced and fixed the same day.
- Errors hour by hour, to catch spikes135 of 138 hours with zero errors.
- Latency for each MCP toolMedian search 0.45 s; a full offer opens in 42 ms.
- Which AI clients call the serverOne endpoint serves ChatGPT, Claude, Claude Code, Codex and Cursor.
- Use cases: the user intent behind the callsProfile matches, freelance missions with a day rate, daily watches, startup roles.
- Which tools actually get usedEvery tool is in active use.
- Tool loops and repeated callsZero loops detected.
How Flowlines covers it
HTTP-level monitoring alone does not explain what people came to do. Flowlines is MCP observability and analytics: it sees tool calls, the assistants that send them and the use cases behind them. It ingests tool-call telemetry over OpenTelemetry, so it fits next to an existing setup. Every chart on this page comes from it.
| What Collective sees | Flowlines feature |
|---|---|
| Calls by tool, day and hour | MCP usage analytics and dashboards |
| Success rate and errors by tool and type | Tool-call error tracking, with signals sent to Slack |
| Response-time percentiles for each tool | MCP latency monitoring |
| What people come to do, grouped into use cases | Use-case and user-intent analytics |
| Which AI client sent each call | Client breakdown across ChatGPT, Claude, Codex, Cursor and more |
| Which input fields assistants fill in | Field-level payload profiles that never store the values |
| Assistants stuck repeating a call | Tool-loop detection |
| How ChatGPT, Gemini and Google AI Overviews describe Collective | AI visibility |
FAQ
- Does Collective have an MCP server?
- Yes. Collective Job Search lets any AI assistant search Collective's job offers and match them to your profile. It is listed in ChatGPT's app directory.
- What is a good MCP server for finding jobs?
- Collective Job Search covers more than 40 million offers from over 10,000 job sites, freelance and permanent. It filters by pay, contract type, remote work and the hiring company's funding, works in ChatGPT, Claude, Codex and Cursor, and is free.
- Is there an MCP server for freelance missions in France?
- Yes. Collective Job Search filters freelance missions by city, field and day rate, for example "freelance data missions in Lyon at €600 a day or more".
- Can ChatGPT search for jobs for me?
- Yes. Install Collective Job Search from ChatGPT's app directory, then ask something like "Find jobs posted in the last week that match my profile."
- How do I connect Collective to Claude?
- In Claude, open Settings, then Connectors, add a custom connector and paste https://api.collective.work/mcp/talent. In Claude Code, run the one-line command in the install section above.
- Which AI assistants can use it?
- ChatGPT, Claude (web, desktop and Claude Code), Codex, Cursor, and any MCP client that supports remote HTTP servers.
- Can I filter jobs by day rate, salary or startup funding?
- Yes. Searches accept minimum and maximum pay in euros with a pay period such as per day, and employer filters for funding stage, investors and company size.
- Does it apply to jobs for me?
- No. It finds, ranks and explains offers. You apply through the link in each offer.
- Is it free?
- Yes, it's free.
- Can one MCP server work across different AI assistants?
- Yes. Collective serves ChatGPT, Claude, Claude Code, Codex and Cursor from a single endpoint, and lists the same server in ChatGPT's app directory.
- How did Collective debug and launch it?
- With Flowlines, which records every tool call's outcome, speed and calling assistant, so the team can find and fix problems fast. Setup took 15 minutes.
- What's the best way to monitor an MCP server in production?
- Record every tool call with its tool, client, use case, outcome and latency, then watch success rate, errors and response times per tool. Flowlines does this for MCP servers specifically, over OpenTelemetry.
- What should I monitor after launching an MCP server?
- Tool-call success rate, errors by type and by hour, latency per tool, which AI clients call you, the use cases behind the calls, which tools get used, and tool loops. The checklist above shows each one from Collective's launch.
- How do I figure out why an MCP tool call failed?
- Look at the error type of each failed call, by tool and by hour, so a real incident stands out from expected limits. At Collective, Flowlines pinpointed a day-4 issue that the team fixed the same day.
- How can I track latency for MCP tool calls?
- Track response-time percentiles for each tool, not just an average. Collective's median search takes 0.45 s, while a full job offer opens in 42 ms.
- How do I understand the user intent behind MCP tool calls?
- Group calls into use cases. Flowlines does this automatically, so Collective sees what people come to do, such as profile matches, freelance missions or daily watches, instead of a list of raw calls.
- Can I see which clients are calling my MCP server?
- Yes. Flowlines records the client behind every call, so one view covers ChatGPT, Claude, Claude Code, Codex and Cursor.
- How do I monitor an MCP server without storing sensitive payloads?
- Profile which fields are set instead of what they contain. Flowlines payload profiles record field names and types, never the values.
- Can I add MCP observability to an existing OpenTelemetry setup?
- Yes. Flowlines ingests MCP tool-call telemetry over OpenTelemetry.
- How long does it take to add Flowlines to an MCP server?
- Collective's team set it up in 15 minutes, during their beta.