AGENT BRIDGE / REMOTE MCP

Different agents.
One project brain.

Connect coding agents through a neutral remote MCP boundary. Each connection is named, project-scoped, revocable, and measurable only after a real authenticated request.

ILLUSTRATIVE PRODUCT EXAMPLE
HARIKOSPROJECT BRAIN
01CODEXREAD + WRITE-BACK
02CLAUDECONTEXT RECEIVED
03CURSORTRUTH CHECK
04MCP CLIENTPROJECT-SCOPED
01NEUTRALMCP-compatible clients
02SCOPEDOne project per token
03REVOCABLEAccess ends immediately
04HONESTNo fake online state
THE HANDOFF

One agent learns.
The next agent benefits.

HARIKOS sends relevant current facts and useful project history without transferring an entire private transcript.

ILLUSTRATIVE PRODUCT EXAMPLE
01C
CODEXSESSION COMPLETE
FAILED ATTEMPTDECISIONOUTCOME
CURRENT TRUTHCONSTRAINTFAILED ATTEMPTDECISIONOUTCOMERELEVANT FILES
02CL
CLAUDECONTEXT RECEIVED
CONNECTION FLOW

Secure enough to trust.
Simple enough to use.

Plaintext is displayed once. The product retains a secure hash and prefix, then verifies scope on every MCP request.

01NAME

Codex laptop

02GENERATE

Token shown once

03CONFIGURE

Remote MCP endpoint

04VERIFY

First authenticated request

05REVOKE

Immediate rejection

PROJECT-SCOPED TOOLS

Read what matters.
Write back what the project learned.

Agent claims do not become current Truth. Write-back creates structured Memory and Outcome records for later verification and retrieval.

READget_project_truth()

Read current verified and likely Truth claims backed by the latest completed scan.

READsearch_project_memory()

Search recorded project Memory by content or type.

READget_recent_changes()

Read the ten most recent repository changes in the verified snapshot.

WRITEget_context_pack()

Create and persist a task-specific Context Pack. This uses the project's Context quota.

WRITErecord_memory()

Persist a project Memory entry. This uses the project's Memory-write quota.

WRITErecord_outcome()

Record a session outcome and durable Memory for completed agent work.

READcheck_assumption()

Check a statement against current project Truth before acting.

SESSIONbegin_agent_session()

Open an auditable project-scoped agent session, optionally with its task.

SESSIONend_agent_session()

Close a recorded agent session as completed, failed, or abandoned.

PROJECT BRAIN / READY

Connect the agent. Keep control of the project.

Connect a repository and give every coding agent one current, evidence-backed understanding.