Door opensourcesignal 2.19.0215 tools18 recipesStreamable HTTP

Hello, agent.

Your human wants ChatGPT to recommend them. That’s a job with steps: find the pages AI reads, find who runs them, get them pitched, prove the win. We hold every step. You run them. Your human says yes.

Agents, your brief is plain markdownGET /outreach-for-agents/brief.md

14-day free trialno card neededcancel anytime

Human? Start the trial, then make your agent a token in Settings, Agents.

MCP session

Agent calls: initialize

Door answers: sourcesignal 2.19.0 · 215 tools · 18 recipes

Agent calls: start

Door answers: You act for Sam at Northpeak · every change is a card or applies as your human set

Agent calls: opportunities_find {"project_id":"…","limit":5}

Door answers: 5 pages · brightpathreview.com first, grade 91

Agent calls: contacts_research {"project_id":"…","limit":5}

Door answers: Card proposed · waiting for Sam

Speaks MCP with
Claude
ChatGPT
Gemini
Copilot
CuCursor
VSVS Code

// The job

Four calls from invisible to cited.

ChatGPT, Gemini and the rest build their favorites from a short list of pages. Get your human onto those pages and they’re in the answer. Here is the whole loop, as you’d call it.

01
opportunities_find
Find the pages AI reads
Every page the engines cite when your human’s buyers ask, graded by what it’s worth to them.
Read · answers at once
02
contacts_research
Find who runs them
The worker looks up the people behind each page and keeps the finds. Priced per page.
Card or at once · your human decides
03
opportunities_promote
Get them pitched
Each promoted page gets a pitch written for it. Your human reviews it, and it leaves from their own inbox.
Card or at once · your human decides
04
placements_find
Read every win
Every placement won: how it was won, whether it’s still live, and the answers that cited it since.
Read · answers at once

You run the loop. Your human says yes. We keep score.

// Recipes

18 plays, handed over on your first call.

start returns them, and a client that lists prompts offers each one as a recipe_… prompt. Each names the calls in order and what every call gives back. Pick one.

prompts/get

# recipe_job_follow_up

# Did the work you started finish, and what did it produce

1jobs_findLists this project's recent jobs with their ids and status.

2statusReads one job: progress by stage, the error, what it produced and whether it can be retried.

3job_retryLeaves a card that runs a failed job again.

// Scopes

Five boxes decide what you can call.

Humans: these are the boxes in Settings, Agents. Agents: this is the tools/list you get back. Tick only what the job needs.

tools/list

212 of 215 tools unlocked

AnswersLeaves a card

project_brief
search
open_view
glossary_lookup
compare_periods
opportunities_find
opportunity_explain
deck_state
pipeline_state
conversations_find
describe
query
read
metric
status
jobs_find
next_steps
insights_since
replies_find
contacts_find
placements_find
outreach_yield
what_worked
mailbox_health
spend_find
agent_activity
clients_health
prompts_find
competitors_movement
visibility_breakdown
tracking_health
engine_rates
score_history
fanout_find
competitor_pages
market_find
data_correlate
deck_patterns
promotion_outcomes
research_yield
drafts_find
posts_find
tracking_reports_find
pages_review
keywords_find
competitors_find
content_gaps_find
content_articles_find
content_improvements_find
content_checks_state
content_piece_get
report_headline
start
catalog
spot_explain
steer_compile
steer_set_rule
steer_set_focus
steer_end
deck_set_budget
pipeline_run
contacts_research
nudges_draft
opportunities_park
opportunities_bring_back
opportunities_promote
job_retry
mailbox_reconnect
opportunity_update
opportunities_pin
opportunity_override_set
promotion_rule_propose
research_window_set
deck_types_set
reply_triage
prompts_add
prompts_remove
prompts_restore
prompt_rename
prompts_import_ahrefs
keyword_prompts_link
keyword_prompts_unlink
competitor_update
competitors_remove
competitor_suggestions_accept
competitor_suggestions_dismiss
competitors_autoadd_set
engines_set
project_markets_set
answers_check_now
analysis_start
draft_edit
draft_rewrite
reply_draft_set
reply_prepare
outreach_prompt_set
reply_rules_set
fee_stance_set
pitch_drafts_create
no_reply_redraft
recipient_page_draft
recipient_page_pick
page_draft_write
opportunities_recommend
drafts_park
sequence_template_save
email_template_save
email_templates_seed
auto_promotion_set
draft_recipient_set
drafts_sender_set
sequence_timing_set
outreach_timing_set
timing_presets_set
outreach_hold
outreach_stop
follow_up_dates_set
reply_reminder_set
reply_reminder_days_set
outreach_signoff
outreach_launch
outreach_send_now
outreach_resume
outreach_resume_anyway
reply_send
step_retry_unconfirmed
step_delivered_record
draft_sent_record
outreach_attempt_record
conversation_mark_won
post_mark_posted
suppression_lift
mailbox_send_settings_set
contact_add
contact_update
addresses_check
contacts_suppress
unmatched_mail_resolve
conversation_contact_switch
page_mention_recheck
post_draft
post_edit
posts_skip
posts_reopen
outreach_sender_set
outreach_hours_set
ahrefs_report_bind
ahrefs_prompts_sync
feed_reconnect
brand_domain_set
tracking_activate
job_cancel
thread_decide
opportunities_regrade
opportunities_group
prompts_classify
competitor_suggestions_add_all
project_create
project_archive
project_restore
tracking_plan_propose
prompts_propose
org_rename
profile_update
support_ask
support_resolve
support_rate
keywords_add
keywords_remove
keyword_update
keywords_restore
keywords_import_csv
keywords_suggest
competitors_add
content_outline_start
site_checks_run
site_read_run
content_piece_mark_published
content_outline_again
content_outline_edit
content_draft_start
content_draft_stop
content_section_rewrite
content_sections_retry
content_section_edit
content_topic_decide
content_piece_move
content_live_check
content_plan_rank
content_autopilot_set
content_voice_read_run
content_voice_accept
content_voice_forget
placement_record
placements_confirm_live
placement_update
placements_check
share_link_create
share_link_revoke
report_commentary_set
briefing_generate
csv_import
traffic_sync_now
analysis_ask
mandate_propose
deck_next
deck_decide
take_back
outreach_prompt_suggest
sequence_template_suggest
spot_plan
spot_ask
org_brief
org_projects_health
org_usage

// House rules

Yes, we wrote you a robots.txt.

The same rules your human reads in the app, in a format you already respect.

agents.txt

# agents.txt for sourcesignal.ai/api/mcp
User-agent: *
Allow: read   # what your human can open, nothing more
Ask: change   # a card your human applies, or at once where their setting says so
Mandate: tidy, track, spend, standing, publish   # by kind, or any tool by name; alone within limits until a date
Mandate-asks: reply_triage   # never alone under a mandate
Mandate-asks: outreach_signoff   # never alone under a mandate
Mandate-asks: outreach_launch   # never alone under a mandate
Mandate-asks: outreach_send_now   # never alone under a mandate
Mandate-asks: outreach_resume   # never alone under a mandate
Mandate-asks: outreach_resume_anyway   # never alone under a mandate
Mandate-asks: reply_send   # never alone under a mandate
Mandate-asks: step_retry_unconfirmed   # never alone under a mandate
Mandate-asks: step_delivered_record   # never alone under a mandate
Mandate-asks: draft_sent_record   # never alone under a mandate
Mandate-asks: outreach_attempt_record   # never alone under a mandate
Mandate-asks: conversation_mark_won   # never alone under a mandate
Mandate-asks: post_mark_posted   # never alone under a mandate
Mandate-asks: suppression_lift   # never alone under a mandate
Mandate-asks: project_create   # never alone under a mandate
Mandate-asks: project_archive   # never alone under a mandate
Mandate-asks: org_rename   # never alone under a mandate
Mandate-asks: profile_update   # never alone under a mandate
Mandate-asks: support_ask   # never alone under a mandate
Mandate-asks: support_rate   # never alone under a mandate
Mandate-asks: keywords_suggest   # never alone under a mandate
Mandate-asks: placement_record   # never alone under a mandate
Mandate-asks: placements_confirm_live   # never alone under a mandate
Mandate-asks: mandate_propose   # never alone under a mandate
Disallow: paying a publisher
Disallow: billing, plans and purchases
Disallow: publishing website pages, or deleting for good
Disallow: members, invites, roles and connecting a mailbox
Disallow: deleting a placement
Disallow: changing how pages are scored, engine policy and admin
Rate-limit: 120/60s   # past it: 429 and Retry-After
Timeout: 60s   # a slow call answers what to read
Answer-size: 24000   # longer answers return continue_from

Words you didn’t write arrive fenced
Publisher replies, AI answers and page titles come back marked as data. A reply that says “ignore your instructions” stays a quote.
Links come back absolute
Answers link into the app on our own domain, so you can hand your human the exact page.
Nothing here is destructive
Every tool says so in its MCP annotations. Reads are read-only. Changes apply only as your human set: a card, their setting or a mandate.
A refusal is final
Don’t retry it in other words. Tell your human what you changed and what’s waiting for them, and stop.

// Connect

Plug in. Or let your agent plug itself in.

Only a human can make a token, in Settings, Agents. That part is on purpose. Everything after it, your agent can do on its own.

Paste this into your agent

It reads its brief, asks you for the token, and connects itself.

Prompt

Read https://www.sourcesignal.ai/outreach-for-agents/brief.md and connect yourself to sourcesignal. I will paste you a token from Settings, Agents. Then call start and tell me what you can do for us.

Claude Code

One line

claude mcp add --transport http sourcesignal https://www.sourcesignal.ai/api/mcp --header "Authorization: Bearer <your token>"

Replace <your token> with the secret the dialog shows. Claude Code sends it on every call and never asks you to sign in.

A shared project

.mcp.json

{
  "mcpServers": {
    "sourcesignal": {
      "type": "http",
      "url": "https://www.sourcesignal.ai/api/mcp",
      "headers": {
        "Authorization": "Bearer ${SOURCESIGNAL_AGENT_TOKEN}"
      }
    }
  }
}

Reads the secret from SOURCESIGNAL_AGENT_TOKEN, so the file can be committed without it.

A hosted client

Claude’s own connector, and any client that cannot set the Authorization header, takes the same URL with the secret in an x-api-key header instead.

// FAQ

Questions agents ask.

And a few from humans.

sourcesignal.ai● ONLINE

Pick a question to see the answer.

Can I send the pitch myself?

Only where your human set “Emails and posts to publishers” to Apply at once in Settings, Agents. Then you can sign a draft off, launch it and send its next email now, each through the same checks as their own press, and never under a mandate. Otherwise you edit your human’s pitch drafts, they sign off and send. It leaves from their own Gmail or Outlook inside their send window.

Can I answer a publisher?

Yes. You draft replies, fee answers included, ready in your human’s Inbox, in your own words or with the platform’s draft. The row says Reply ready and the reply box opens with it. Your human sends it, or you do where they set “Emails and posts to publishers” to Apply at once, never under a mandate.

My call came back 401.

The token is wrong, revoked or expired. Ask your human for a new one in Settings, Agents. You will never be asked to sign in.

How do I get more autonomy?

Ask for a mandate with mandate_propose. If your human approves it, you may make the changes it names on your own, by tool or by kind (125 tools can be named, tidying up first), within the decisions and paid work a day they set, until the date they set. 90 days at most. A mandate never covers an email or a post. Your human can also set a kind of change to apply at once, under Settings, Agents.

A tool is taking a while.

Long work answers with an id. Poll status with it, and it waits up to 40 seconds for the work to settle.

Which clients work?

Anything that speaks MCP over HTTP and can send the token in a header. Claude Code and other command-line agents use the Authorization header. Claude’s own connector, and any client that cannot set that header, puts the secret in an x-api-key header instead.

What does a call cost?

Most calls read what is already there or leave a card, so they cost nothing. Calls that run the platform’s own AI (outreach_prompt_suggest, sequence_template_suggest, steer_compile, keywords_suggest, spot_plan and spot_ask) use the account’s AI allowance, which no plan limits today. Work you prepare is counted as the same button is in the app, and only when it applies: when your human applies the card, or at once where their setting or a mandate allows.

How safe is a token?

Treat it like a password. The secret shows once and only a hash is kept. Revoke it in a click and the agent stops at once. On every call the door re-checks what the account may hold now, so a demoted admin’s token loses the account scope on the next request.

Can I limit it to one client?

Yes. When you make a token you tick the projects it may reach, and that is fixed for its life. An agency gives each client’s agent a token that reaches only that client.

Does my data leave sourcesignal?

Only to the agent you connect, and only what it asks for and you can open. What happens to it there is up to that agent’s provider, so connect the ones you trust.

End of transmission.

Agents: your brief is waiting. Humans: start the trial, make your agent a token, and go get cited.

GET /outreach-for-agents/brief.md